TWI823501B - Photoplethysmography based non-invasive blood glucose prediction by neural network - Google Patents
Photoplethysmography based non-invasive blood glucose prediction by neural network Download PDFInfo
- Publication number
- TWI823501B TWI823501B TW111128479A TW111128479A TWI823501B TW I823501 B TWI823501 B TW I823501B TW 111128479 A TW111128479 A TW 111128479A TW 111128479 A TW111128479 A TW 111128479A TW I823501 B TWI823501 B TW I823501B
- Authority
- TW
- Taiwan
- Prior art keywords
- ppg
- cnn
- hba1c
- signal
- medical treatment
- Prior art date
Links
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 94
- 238000013186 photoplethysmography Methods 0.000 title claims description 194
- WQZGKKKJIJFFOK-GASJEMHNSA-N Glucose Natural products OC[C@H]1OC(O)[C@H](O)[C@@H](O)[C@@H]1O WQZGKKKJIJFFOK-GASJEMHNSA-N 0.000 title claims description 64
- 239000008103 glucose Substances 0.000 title claims description 64
- 210000004369 blood Anatomy 0.000 title claims description 58
- 239000008280 blood Substances 0.000 title claims description 58
- 238000011282 treatment Methods 0.000 claims abstract description 161
- 238000012549 training Methods 0.000 claims abstract description 134
- 238000000034 method Methods 0.000 claims abstract description 63
- 238000013527 convolutional neural network Methods 0.000 claims description 49
- 238000012545 processing Methods 0.000 claims description 24
- 230000008569 process Effects 0.000 claims description 10
- 230000008859 change Effects 0.000 claims description 6
- 230000035487 diastolic blood pressure Effects 0.000 claims description 4
- 230000035488 systolic blood pressure Effects 0.000 claims description 4
- 238000004458 analytical method Methods 0.000 description 46
- 229940079593 drug Drugs 0.000 description 35
- 239000003814 drug Substances 0.000 description 35
- 239000013598 vector Substances 0.000 description 33
- 238000012360 testing method Methods 0.000 description 25
- 206010012601 diabetes mellitus Diseases 0.000 description 18
- 210000000748 cardiovascular system Anatomy 0.000 description 17
- 238000005259 measurement Methods 0.000 description 11
- 238000012544 monitoring process Methods 0.000 description 10
- 230000011664 signaling Effects 0.000 description 9
- 239000011159 matrix material Substances 0.000 description 8
- 238000005516 engineering process Methods 0.000 description 7
- 238000011176 pooling Methods 0.000 description 7
- NOESYZHRGYRDHS-UHFFFAOYSA-N insulin Chemical compound N1C(=O)C(NC(=O)C(CCC(N)=O)NC(=O)C(CCC(O)=O)NC(=O)C(C(C)C)NC(=O)C(NC(=O)CN)C(C)CC)CSSCC(C(NC(CO)C(=O)NC(CC(C)C)C(=O)NC(CC=2C=CC(O)=CC=2)C(=O)NC(CCC(N)=O)C(=O)NC(CC(C)C)C(=O)NC(CCC(O)=O)C(=O)NC(CC(N)=O)C(=O)NC(CC=2C=CC(O)=CC=2)C(=O)NC(CSSCC(NC(=O)C(C(C)C)NC(=O)C(CC(C)C)NC(=O)C(CC=2C=CC(O)=CC=2)NC(=O)C(CC(C)C)NC(=O)C(C)NC(=O)C(CCC(O)=O)NC(=O)C(C(C)C)NC(=O)C(CC(C)C)NC(=O)C(CC=2NC=NC=2)NC(=O)C(CO)NC(=O)CNC2=O)C(=O)NCC(=O)NC(CCC(O)=O)C(=O)NC(CCCNC(N)=N)C(=O)NCC(=O)NC(CC=3C=CC=CC=3)C(=O)NC(CC=3C=CC=CC=3)C(=O)NC(CC=3C=CC(O)=CC=3)C(=O)NC(C(C)O)C(=O)N3C(CCC3)C(=O)NC(CCCCN)C(=O)NC(C)C(O)=O)C(=O)NC(CC(N)=O)C(O)=O)=O)NC(=O)C(C(C)CC)NC(=O)C(CO)NC(=O)C(C(C)O)NC(=O)C1CSSCC2NC(=O)C(CC(C)C)NC(=O)C(NC(=O)C(CCC(N)=O)NC(=O)C(CC(N)=O)NC(=O)C(NC(=O)C(N)CC=1C=CC=CC=1)C(C)C)CC1=CN=CN1 NOESYZHRGYRDHS-UHFFFAOYSA-N 0.000 description 6
- 230000006872 improvement Effects 0.000 description 5
- 238000003062 neural network model Methods 0.000 description 5
- 238000010606 normalization Methods 0.000 description 5
- 230000002265 prevention Effects 0.000 description 5
- 230000001105 regulatory effect Effects 0.000 description 5
- 206010020772 Hypertension Diseases 0.000 description 4
- 230000036772 blood pressure Effects 0.000 description 4
- 238000013461 design Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 4
- 230000006870 function Effects 0.000 description 4
- 238000010801 machine learning Methods 0.000 description 4
- 208000010125 myocardial infarction Diseases 0.000 description 4
- 238000011160 research Methods 0.000 description 4
- 102000004877 Insulin Human genes 0.000 description 3
- 108090001061 Insulin Proteins 0.000 description 3
- 230000004913 activation Effects 0.000 description 3
- 238000013459 approach Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 3
- 238000004422 calculation algorithm Methods 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 3
- 239000000284 extract Substances 0.000 description 3
- 238000002347 injection Methods 0.000 description 3
- 239000007924 injection Substances 0.000 description 3
- 229940125396 insulin Drugs 0.000 description 3
- 230000031700 light absorption Effects 0.000 description 3
- 239000000203 mixture Substances 0.000 description 3
- 238000012552 review Methods 0.000 description 3
- 239000000523 sample Substances 0.000 description 3
- 208000024172 Cardiovascular disease Diseases 0.000 description 2
- 238000004497 NIR spectroscopy Methods 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 2
- 210000004027 cell Anatomy 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- HVYWMOMLDIMFJA-DPAQBDIFSA-N cholesterol Chemical compound C1C=C2C[C@@H](O)CC[C@]2(C)[C@@H]2[C@@H]1[C@@H]1CC[C@H]([C@H](C)CCCC(C)C)[C@@]1(C)CC2 HVYWMOMLDIMFJA-DPAQBDIFSA-N 0.000 description 2
- 230000001149 cognitive effect Effects 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 230000007774 longterm Effects 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 238000007637 random forest analysis Methods 0.000 description 2
- 210000003491 skin Anatomy 0.000 description 2
- 239000000243 solution Substances 0.000 description 2
- 238000004611 spectroscopical analysis Methods 0.000 description 2
- 230000009861 stroke prevention Effects 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- 210000001519 tissue Anatomy 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 206010061666 Autonomic neuropathy Diseases 0.000 description 1
- KHMVXSQLPUNRCF-UHFFFAOYSA-N DL-Adalin Natural products C1CCC2CC(=O)CC1(CCCCC)N2 KHMVXSQLPUNRCF-UHFFFAOYSA-N 0.000 description 1
- 206010018429 Glucose tolerance impaired Diseases 0.000 description 1
- 102000017011 Glycated Hemoglobin A Human genes 0.000 description 1
- 108010014663 Glycated Hemoglobin A Proteins 0.000 description 1
- 241000282412 Homo Species 0.000 description 1
- 241000124008 Mammalia Species 0.000 description 1
- 206010033372 Pain and discomfort Diseases 0.000 description 1
- 208000001280 Prediabetic State Diseases 0.000 description 1
- 206010067584 Type 1 diabetes mellitus Diseases 0.000 description 1
- 241000251539 Vertebrata <Metazoa> Species 0.000 description 1
- 238000009825 accumulation Methods 0.000 description 1
- 230000002411 adverse Effects 0.000 description 1
- 125000000129 anionic group Chemical group 0.000 description 1
- -1 anionic lipid Chemical class 0.000 description 1
- 230000003466 anti-cipated effect Effects 0.000 description 1
- QVGXLLKOCUKJST-UHFFFAOYSA-N atomic oxygen Chemical compound [O] QVGXLLKOCUKJST-UHFFFAOYSA-N 0.000 description 1
- 238000005452 bending Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- WQZGKKKJIJFFOK-VFUOTHLCSA-N beta-D-glucose Chemical compound OC[C@H]1O[C@@H](O)[C@H](O)[C@@H](O)[C@@H]1O WQZGKKKJIJFFOK-VFUOTHLCSA-N 0.000 description 1
- 230000004071 biological effect Effects 0.000 description 1
- 239000013060 biological fluid Substances 0.000 description 1
- 230000008081 blood perfusion Effects 0.000 description 1
- 230000037237 body shape Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 235000012000 cholesterol Nutrition 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 230000002508 compound effect Effects 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000013434 data augmentation Methods 0.000 description 1
- 210000004207 dermis Anatomy 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 229940029980 drug used in diabetes Drugs 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 230000002526 effect on cardiovascular system Effects 0.000 description 1
- 210000003743 erythrocyte Anatomy 0.000 description 1
- 230000007717 exclusion Effects 0.000 description 1
- 125000002791 glucosyl group Chemical group C1([C@H](O)[C@@H](O)[C@H](O)[C@H](O1)CO)* 0.000 description 1
- 150000003278 haem Chemical class 0.000 description 1
- 231100001261 hazardous Toxicity 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 230000003862 health status Effects 0.000 description 1
- 230000000004 hemodynamic effect Effects 0.000 description 1
- 238000010921 in-depth analysis Methods 0.000 description 1
- 208000015181 infectious disease Diseases 0.000 description 1
- 239000004615 ingredient Substances 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- YECIFGHRMFEPJK-UHFFFAOYSA-N lidocaine hydrochloride monohydrate Chemical compound O.[Cl-].CC[NH+](CC)CC(=O)NC1=C(C)C=CC=C1C YECIFGHRMFEPJK-UHFFFAOYSA-N 0.000 description 1
- 238000012886 linear function Methods 0.000 description 1
- 150000002632 lipids Chemical class 0.000 description 1
- 235000012054 meals Nutrition 0.000 description 1
- 238000013160 medical therapy Methods 0.000 description 1
- 238000002483 medication Methods 0.000 description 1
- 238000010197 meta-analysis Methods 0.000 description 1
- 230000002503 metabolic effect Effects 0.000 description 1
- 230000000116 mitigating effect Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012806 monitoring device Methods 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 238000012014 optical coherence tomography Methods 0.000 description 1
- 230000008520 organization Effects 0.000 description 1
- 229910052760 oxygen Inorganic materials 0.000 description 1
- 239000001301 oxygen Substances 0.000 description 1
- 206010033675 panniculitis Diseases 0.000 description 1
- 230000007310 pathophysiology Effects 0.000 description 1
- 230000037361 pathway Effects 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 230000002093 peripheral effect Effects 0.000 description 1
- 230000002688 persistence Effects 0.000 description 1
- 230000035790 physiological processes and functions Effects 0.000 description 1
- 229920001451 polypropylene glycol Polymers 0.000 description 1
- 201000009104 prediabetes syndrome Diseases 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 230000010349 pulsation Effects 0.000 description 1
- 238000011002 quantification Methods 0.000 description 1
- 239000012925 reference material Substances 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 238000007920 subcutaneous administration Methods 0.000 description 1
- 210000004304 subcutaneous tissue Anatomy 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 230000009897 systematic effect Effects 0.000 description 1
- 238000002834 transmittance Methods 0.000 description 1
- 208000001072 type 2 diabetes mellitus Diseases 0.000 description 1
- 208000019553 vascular disease Diseases 0.000 description 1
- 230000004218 vascular function Effects 0.000 description 1
- 230000002747 voluntary effect Effects 0.000 description 1
Landscapes
- Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
- Measuring Pulse, Heart Rate, Blood Pressure Or Blood Flow (AREA)
Abstract
Description
本發明一般而言係關於一種用於使用神經網絡預測一受試者之血糖值的非侵入式系統及方法。The present invention generally relates to a non-invasive system and method for predicting blood glucose levels of a subject using neural networks.
長期以來，人們需要並廣泛地研究精確且可靠之非侵入式血糖（NIBG）測量技術。手指扎刺血糖（BG）測量因侵入性質造成疼痛及不適，且引起感染的風險[1,2]，NIBG技術可藉由減輕定期侵入式測量的痛苦而大幅改善糖尿病病患的生活品質[3-8]。在所研討之NIBG測量中，光體積描記法（PPG）一直以來備受期待，這是由於此技術簡單、成本低、已廣泛用在各種穿戴式裝置中[9-15]，且已成功地應用在臨床測量氧飽和度（SpO2）及脈搏率。PPG裝置在血液通過周邊微血管時測量近紅外光之透射率或反射率的變化。已知特定波長之光吸收率及反射率對人體的血液動力學性質敏感，其與長期受血糖值影響之心血管系統的健康狀態高度相關，且可直接量測為脈搏型態曲線[9,11,13,16]。因此，找出PPG脈搏形態與血糖值之相關性可係達成NIBG預測之一可行方法。Accurate and reliable non-invasive blood glucose (NIBG) measurement technology has long been needed and extensively studied. Finger-prick blood glucose (BG) measurement causes pain and discomfort due to its invasive nature, as well as the risk of infection [1,2]. NIBG technology can greatly improve the quality of life of diabetic patients by reducing the pain of regular invasive measurements [3 -8]. Among the NIBG measurements studied, photoplethysmography (PPG) has been highly anticipated because this technology is simple, low-cost, has been widely used in various wearable devices [9-15], and has been successfully Used in clinical measurement of oxygen saturation (SpO2) and pulse rate. The PPG device measures changes in the transmittance or reflectance of near-infrared light as blood passes through peripheral capillaries. It is known that the light absorption and reflectance of specific wavelengths are sensitive to the hemodynamic properties of the human body, are highly related to the health status of the cardiovascular system that is affected by blood sugar levels for a long time, and can be directly measured as a pulse pattern curve [9, 11,13,16]. Therefore, finding the correlation between PPG pulse shape and blood glucose level can be a feasible method to predict NIBG.
先前對於NIBG測量之後分析的研究涵蓋各種不同的機器學習模型，諸如支援向量機（SVM）[17]、隨機森林[18]、K最近鄰（KNN）[19]、高斯過程回歸（GPR）[20]及人工神經網絡（ANN）[21]。提取自PPG信號之許多型態分佈及心率變化（HRV）特徵與人體之血管功能[22]及自主神經病變[23]相關。亦利用不同信號處理方法（諸如快速傅立葉轉換（FFT）、Kaiser-Teager能量（KTE）及譜熵）來提取不同域之特徵[9]。已觀察到，即使此等結果已宣稱對其有限數量之測試對象有效，但其中並無一者已成功地應用於臨床用途的商業產品上。此表明其方法對廣大人群的可行性可能相當有限。Previous research on post-measurement analysis of NIBG covers various machine learning models, such as support vector machine (SVM) [17], random forest [18], K-nearest neighbor (KNN) [19], Gaussian process regression (GPR) [ 20] and artificial neural network (ANN) [21]. Many pattern distributions and heart rate variation (HRV) features extracted from PPG signals are related to the human body's vascular function [22] and autonomic neuropathy [23]. Different signal processing methods (such as Fast Fourier Transform (FFT), Kaiser-Teager Energy (KTE) and spectral entropy) are also used to extract features in different domains [9]. It has been observed that even though these results have been claimed to be valid on a limited number of test subjects, none of them have been successfully applied to commercial products for clinical use. This suggests that the feasibility of its approach to the broad population may be rather limited.
基於此觀察，有可能大型資料集仍包含未有文件證明的變數，因此單一通用模型不適於涵蓋所有案例且可造成精確性降低。因此，需要一種神經網絡模型，其可對於預定的人口子集使用PPG信號進行精確的NIBG預測。Based on this observation, it is possible that large data sets still contain undocumented variables, so that a single general model is not suitable to cover all cases and can result in reduced accuracy. Therefore, there is a need for a neural network model that can perform accurate NIBG prediction using PPG signals for predetermined population subsets.
本發明提供一種基於光體積描記法（PPG）之非侵入式血糖（NIBG）預測系統，其包含一信號讀取器，其經配置以讀取來自一受試者之一或多個PPG信號；及一處理器，其依次包含一神經網絡，其中該處理器經配置以對藉由該信號讀取器讀取之該一或多個PPG信號進行信號處理；其中該處理器經配置以使用該神經網絡預測一受試者之血糖值；其中至該神經網絡之輸入包含藉由該處理器處理之該一或多個PPG信號之信號處理結果；其中該神經網絡係使用自未接受任何醫療之定群獲得的訓練資料訓練；及其中該受試者未接受任何醫療。The present invention provides a non-invasive blood glucose (NIBG) prediction system based on photoplethysmography (PPG), which includes a signal reader configured to read one or more PPG signals from a subject; and a processor, in turn comprising a neural network, wherein the processor is configured to perform signal processing on the one or more PPG signals read by the signal reader; wherein the processor is configured to use the A neural network predicts the blood glucose level of a subject; wherein the input to the neural network includes signal processing results of the one or more PPG signals processed by the processor; wherein the neural network is used from patients who have not received any medical treatment. The subject was trained using training materials obtained from the group; and the subject did not receive any medical treatment.
本發明亦提供一種基於PPG之非侵入式血糖（NIBG）預測的方法，其包含以下步驟：自一受試者讀取一或多個PPG信號；使用習知之手指扎刺法自該受試者獲得HbA1c；處理該一或多個PPG信號；訓練一神經網絡；及使用該經訓練之神經網絡預測該受試者之血糖值，其中至該神經網絡之輸入包含HbA1c及該一或多個PPG信號之處理結果；其中該訓練步驟係使用獲自未接受任何醫療之定群的HbA1c及PPG信號資料進行；及其中該受試者未接受任何醫療。The present invention also provides a method for non-invasive blood glucose (NIBG) prediction based on PPG, which includes the following steps: reading one or more PPG signals from a subject; Obtaining HbA1c; processing the one or more PPG signals; training a neural network; and using the trained neural network to predict the blood glucose level of the subject, wherein the input to the neural network includes HbA1c and the one or more PPGs The result of signal processing; wherein the training step is performed using HbA1c and PPG signal data obtained from a cohort that did not receive any medical treatment; and wherein the subject did not receive any medical treatment.
本發明之組成物可包含本文所述之本發明之基本要素及限制，以及任何本文所述之額外或視情況選用之成份、組份或限制，由其等組成，或基本上由其等組成。The compositions of the present invention may include, consist of, or consist essentially of the essential elements and limitations of the invention described herein, as well as any additional or optional ingredients, components, or limitations described herein. .
如本說明書及申請專利範圍所用，單數形式「一」、「一個」及「該」包括複數參照物，除非文中另外明確指明。例如，該術語「一」個細胞包括複數個細胞，包括其混合物。As used in this specification and claims, the singular forms "a", "an" and "the" include plural references unless the context clearly dictates otherwise. For example, the term "a" cell includes a plurality of cells, including mixtures thereof.
在數量值內容中之「大約」係指基於指示值最大±20%、±10%或±5%之平均偏差。例如，相對於該總脂質/兩親性體積莫耳濃度，大約30 mol %陰離子脂質之量係指30 mol % ±6 mol %、30 mol % ±3 mol %或30 mol % ±1.5 mol %之陰離子脂質。"Approximately" in the context of quantitative values refers to the average deviation based on the maximum ±20%, ±10% or ±5% of the indicated value. For example, an amount of approximately 30 mol % anionic lipid relative to the total lipid/amphiphilic volume molar concentration means 30 mol % ±6 mol %, 30 mol % ±3 mol %, or 30 mol % ±1.5 mol % Anionic lipids.
一「受試者」、「個體」或「病患」在本文中可交換使用，其係指一脊椎動物，較佳地一哺乳動物，更佳地一人類。A "subject," "individual," or "patient" are used interchangeably herein and refer to a vertebrate, preferably a mammal, and more preferably a human.
「神經網絡」（人工智能（AI）之一子集）包含節點之網絡，各節點可與一或多個其它節點通信，其中該網絡之各節點作用為一神經元，其能夠學習及模仿人類思考過程。該神經網絡通常包含一輸入層、一或多個隱藏層及一輸出層。各種神經網絡框架可包含（但不限於）Caffee、Keras、Microsoft Cognitive Toolkit、MXNET、DeepLearning4j、 Chainer及TensorFlow。深度學習神經網絡係包含三或多個隱藏層之神經網絡的一子集。"Neural network" (a subset of artificial intelligence (AI)) consists of a network of nodes, each node can communicate with one or more other nodes, where each node of the network acts as a neuron and is capable of learning and imitating humans thought process. The neural network usually includes an input layer, one or more hidden layers and an output layer. Various neural network frameworks may include (but are not limited to) Caffee, Keras, Microsoft Cognitive Toolkit, MXNET, DeepLearning4j, Chainer, and TensorFlow. Deep learning neural networks are a subset of neural networks that contain three or more hidden layers.
該術語「光體積描記法（PPG）」係一光學獲得之體積描記法，其可用於偵測在組織之微血管床中的血容量變化。[1][2] PPG通常藉由使用脈衝式血氧計獲得，其照明皮膚並測量光吸收之變化。[3]習知之脈衝式血氧計監測至皮膚之真皮及皮下組織的血液灌注。The term "photoplethysmography (PPG)" refers to an optically acquired plethysmography method that can be used to detect changes in blood volume in the microvascular beds of tissue. [1][2] PPG is usually obtained by using a pulse oximeter, which illuminates the skin and measures changes in light absorption. [3] The conventional pulse oximeter monitors the blood perfusion of the dermis and subcutaneous tissue of the skin.
「HbA1c」係血流中之血紅素，其隨著時間變得與葡萄糖化學結合。其已被美國糖尿病協會公認為診斷糖尿病及識別糖尿病前期的強國指數（powerhouse index）之一。測量HbA1c值可幫助個人測定其平均血糖值，及可在任何時間進行而不需事先飲食準備（例如：禁食）。不像血糖測試，HbA1c不會受到每日變異的影響。HbA1c值可用糖尿病控制與併發症試驗單位（DCCT，%），或國際臨床化學聯合會單位（IFCC，mmol/mol）來表示。對於本揭示案，所有HbA1c值使用DCCT單位以百分比（%）報告。由於紅血球細胞壽命有限，所測量之HbA1c反映大約過去三個月的平均血糖值。用於診斷糖尿病之經建議的HbA1c截止值係6.5%。但少於6.5%之HbA1c值未排除糖尿病之可能性[24]。儘管在HbA1c及BG值之間存在相關性，其等不可直接互換，如圖5B及5C所示。由於HbA1c僅需每季測量，其應不會造成太多不適及不便，因此其可接受來用在我們的模型建構。"HbA1c" is heme in the bloodstream, which becomes chemically bound to glucose over time. It has been recognized by the American Diabetes Association as one of the powerhouse indexes for diagnosing diabetes and identifying prediabetes. Measuring HbA1c helps an individual determine their average blood sugar level and can be done at any time without prior meal preparation (e.g. fasting). Unlike blood glucose tests, HbA1c is not affected by daily variation. HbA1c values can be expressed in Diabetes Control and Complications Trial units (DCCT, %), or International Federation of Clinical Chemistry units (IFCC, mmol/mol). For this disclosure, all HbA1c values are reported as a percentage (%) using DCCT units. Because red blood cells have a limited lifespan, the measured HbA1c reflects approximately the average blood glucose level over the past three months. The recommended HbA1c cutoff for diagnosing diabetes is 6.5%. However, an HbA1c value less than 6.5% does not rule out the possibility of diabetes [24]. Although there is a correlation between HbA1c and BG values, they are not directly interchangeable, as shown in Figures 5B and 5C. Since HbA1c only needs to be measured quarterly, it should not cause much discomfort and inconvenience, so it is acceptable for use in our model construction.
詞語「醫療」包括由美國食品藥物管理局（United States Food and Drug Administration）及/或歐洲藥品管理局（European Medicines Agency）規範為醫療程序、藥物或用藥方案的任何醫療程序、藥物或用藥方案。The term "medical treatment" includes any medical procedure, drug or medication regimen regulated as a medical procedure, drug or medication regimen by the United States Food and Drug Administration and/or the European Medicines Agency.
本發明提供一種基於PPG之非侵入式血糖（NIBG）神經網絡預測系統，其基於以下發現建立：一種用於基於PPG之NIBG預測的單一通用模型無法適當地涵蓋所有可能受試者之所有可能的變化，降低了整體NIBG預測的精確性。一項此種主要變化係源自一受試者之醫療，在與圖8A-F、10-13及14-15連結之實例中，其顯示對NIBG預測精確性有明顯的負面效應。特定而言，由於基於PPG之NIBG預測仰賴於血糖值與獲自該受試者之PPG信號（表示該受試者之心血管系統）的相關性，因此對心血管疾病之醫療可影響心血管系統及進而影響PPG信號，使得實質上改變或甚至消除了其對血糖（BG）值之相關性。The present invention provides a PPG-based non-invasive blood glucose (NIBG) neural network prediction system based on the discovery that a single general model for PPG-based NIBG prediction cannot appropriately cover all possible conditions for all possible subjects. changes, reducing the overall NIBG prediction accuracy. One such major change resulted from a subject's medical treatment, which in the examples linked to Figures 8A-F, 10-13, and 14-15 was shown to have a significant negative effect on NIBG prediction accuracy. In particular, because PPG-based NIBG prediction relies on the correlation of blood glucose values with PPG signals obtained from the subject (indicative of the subject's cardiovascular system), treatments for cardiovascular disease can affect cardiovascular disease. The system then affects the PPG signal, substantially changing or even eliminating its correlation with blood glucose (BG) values.
因此，在一具體實施例中，本發明之基於PPG之NIBG神經網絡預測系統包含一神經網絡，其配置以基於獲自一受試者之PPG信號預測該受試者之血糖值，其中該受試者未接受可影響該受試者之心血管系統之醫療，且該神經網絡係使用來自未接受可影響心血管系統之醫療之訓練定群的訓練資料來訓練，其中該訓練資料係基於獲自該訓練定群之PPG信號及參考BG。在另一具體實施例中，本發明之基於PPG之NIBG神經網絡預測系統係基於使用習知手指扎刺法測量之一受試者之HbA1c以及獲自該受試者之PPG信號預測該受試者之BG值，其中該受試者未接受可影響該受試者之心血管系統之醫療，且該神經網絡係使用來自未接受可影響心血管系統之醫療之訓練定群的訓練資料來訓練，其中該訓練資料係基於獲自該訓練定群之PPG信號、HbA1c及參考BG。在一具體實施例中，可影響心血管系統之醫療包含由美國FDA及/或歐洲藥品局規範之高血壓與高血壓預防相關之藥物及治療、中風與中風預防藥物及治療，以及心臟病發作與心臟病發作預防藥物及治療。在一具體實施例中，可影響心血管系統之醫療由美國FDA及/或歐洲藥品局規範之高血壓與高血壓預防相關之藥物及治療、中風與中風預防藥物及治療，以及心臟病發作與心臟病發作預防藥物及治療組成。Therefore, in a specific embodiment, the PPG-based NIBG neural network prediction system of the present invention includes a neural network configured to predict the blood glucose level of a subject based on a PPG signal obtained from the subject, wherein the subject The subject did not receive medical treatment that could affect the cardiovascular system of the subject, and the neural network was trained using training data from a training cohort that did not receive medical treatment that could affect the cardiovascular system, where the training data was based on PPG signal and reference BG from this training group. In another specific embodiment, the PPG-based NIBG neural network prediction system of the present invention predicts a subject based on the HbA1c measured using the conventional finger stick method and the PPG signal obtained from the subject. The BG value of a subject who did not receive medical treatment that could affect the subject's cardiovascular system, and the neural network was trained using training data from a training cohort that did not receive medical treatment that could affect the cardiovascular system. , where the training data is based on the PPG signal, HbA1c and reference BG obtained from the training cohort. In one embodiment, medical treatments that may affect the cardiovascular system include drugs and treatments related to hypertension and hypertension prevention, stroke and stroke prevention drugs and treatments, and heart attacks regulated by the U.S. FDA and/or the European Medicines Agency. and heart attack prevention medications and treatments. In one embodiment, medical treatments that may affect the cardiovascular system include hypertension and hypertension prevention-related drugs and treatments, stroke and stroke prevention drugs and treatments, and heart attack and prevention drugs and treatments regulated by the U.S. FDA and/or the European Medicines Agency. Heart attack prevention drugs and treatment composition.
此外，未直接影響心血管系統之任何醫療（諸如糖尿病藥物或諸如胰島素注射之治療）亦可用作排除一受試者及訓練定群之標準，以改善BG預測之精確性。因此，在一具體實施例中，用於排除本發明之受試者及訓練定群的醫療包含以下醫療：可影響如上文定義之心血管系統以及由美國FDA及/或歐洲藥品管理局規範之與糖尿病相關之醫療，諸如胰島素注射。In addition, any medical treatment that does not directly affect the cardiovascular system (such as diabetes medications or treatments such as insulin injections) can also be used as a criterion for excluding a subject and training cohorts to improve the accuracy of BG prediction. Therefore, in a specific embodiment, medical treatments used to exclude subjects and training cohorts of the present invention include medical treatments that affect the cardiovascular system as defined above and are regulated by the US FDA and/or the European Medicines Agency Diabetes-related medical treatments, such as insulin injections.
此外，在一具體實施例中，為更精確地限定本發明預測精確性之方法及系統，排除接受任何醫療之受試者或訓練定群，無論該醫療是否影響心血管系統或用於治療糖尿病。此種醫療包括由美國食品藥物管理局及/或歐洲藥品管理局認可為醫療、藥物或用藥方案的任何醫療、藥物或用藥方案，無論該醫療是否影響心血管系統或用於治療糖尿病。因此，在一具體實施例中，本發明之基於PPG之NIBG神經網絡預測系統包含一神經網絡，其經配置以基於獲自該受試者之PPG信號預測一受試者之血糖值，其中該受試者未接受任何醫療，且該神經網絡係使用來自未接受任何醫療之定群的訓練資料來訓練，其中該訓練資料係基於獲自該訓練定群之PPG信號。在另一具體實施例，本發明之基於PPG之NIBG神經網絡預測系統基於使用習知之手指扎刺法測量之一受試者之HbA1c以及獲自該受試者之PPG信號預測該受試者之血糖值，其中該受試者未接受任何醫療，且該神經網絡使用來自未接受任何醫療之定群的訓練資料訓練，其中該訓練資料係基於獲自該訓練定群之PPG信號、HbA1c及參考BG。Additionally, in a specific embodiment, to more precisely define the predictive accuracy of the methods and systems of the present invention, subjects or training cohorts who receive any medical treatment, regardless of whether the medical treatment affects the cardiovascular system or is used to treat diabetes, are excluded . Such medical treatment includes any medical treatment, drug or medication regimen recognized as a medical treatment, drug or medication regimen by the U.S. Food and Drug Administration and/or the European Medicines Agency, whether or not the medical treatment affects the cardiovascular system or is used to treat diabetes. Therefore, in a specific embodiment, the PPG-based NIBG neural network prediction system of the present invention includes a neural network configured to predict the blood glucose level of a subject based on the PPG signal obtained from the subject, wherein the The subjects did not receive any medical treatment, and the neural network was trained using training data from a cohort that did not receive any medical treatment, where the training data was based on PPG signals obtained from the training cohort. In another specific embodiment, the PPG-based NIBG neural network prediction system of the present invention predicts the subject's HbA1c based on measuring the HbA1c of a subject using the conventional finger stick method and the PPG signal obtained from the subject. Blood glucose values, where the subject did not receive any medical treatment, and the neural network was trained using training data from a cohort that did not receive any medical treatment, where the training data was based on the PPG signal, HbA1c, and reference obtained from the training cohort B.G.
圖1說明本發明之基於PPG之NIBG神經網絡預測系統之具體實施例。如圖1所示，本發明之系統包含受試者100、PPG信號裝置110、連接器120及處理器200。在一具體實施例中，該受試者100包含一個人，其血糖值係使用本發明之系統及方法預測，且未接受可影響他或她的心血管系統之任何醫療及/或未接受糖尿病醫療。在一具體實施例中，該受試者100包含一個人，其血糖值係使用本發明之系統及方法預測且未接受任何醫療。Figure 1 illustrates a specific embodiment of the NIBG neural network prediction system based on PPG of the present invention. As shown in Figure 1, the system of the present invention includes a subject 100, a PPG signal device 110, a connector 120 and a processor 200. In one embodiment, the subject 100 includes a person whose blood glucose levels were predicted using the systems and methods of the present invention and who is not receiving any medical treatment that may affect his or her cardiovascular system and/or is not receiving diabetes treatment. . In a specific embodiment, the subject 100 includes a person whose blood glucose level is predicted using the system and method of the present invention and who does not receive any medical treatment.
在一具體實施例中，該PPG信號裝置110包含一PPG信號讀取器112、信號發射器114及一信號模組118。該PPG信號讀取器112經配置以讀取自該受試者100發射之信號。該信號裝置110可進一步包含一信號發射器114，其用於輸出通過該受試者100的身體，然後從該受試者100發出以由該信號讀取器112讀取的信號。In a specific embodiment, the PPG signal device 110 includes a PPG signal reader 112, a signal transmitter 114 and a signal module 118. The PPG signal reader 112 is configured to read signals emitted from the subject 100 . The signaling device 110 may further include a signal transmitter 114 for outputting a signal that passes through the body of the subject 100 and is then emitted from the subject 100 for reading by the signal reader 112 .
該信號裝置110進一步包含一信號模組118，其經配置以與該信號讀取器112及信號發射器114通信並控制其等。該信號模組118亦可包含一或多個控制板，其允許一使用者控制傳入與傳出信號，諸如觸發信號及/或擷取信號。The signaling device 110 further includes a signaling module 118 configured to communicate with and control the signal reader 112 and signal transmitter 114 . The signal module 118 may also include one or more control panels that allow a user to control incoming and outgoing signals, such as trigger signals and/or capture signals.
連接器120經配置以允許在該信號裝置110及該處理器200之間的通信。在一具體實施例中，該連接器120可將藉由該信號裝置110讀取之信號傳輸至該處理器200，以及將來自該處理器200之指令傳輸至該信號裝置110以命令該信號模組118觸發及/或讀取信號。在一具體實施例中，該連接器120可係一實體線路。在另一具體實施例，該連接器120可係一無線連接，諸如使用Wi-Fi或藍芽技術者。Connector 120 is configured to allow communication between the signaling device 110 and the processor 200 . In a specific embodiment, the connector 120 can transmit signals read by the signaling device 110 to the processor 200, and transmit instructions from the processor 200 to the signaling device 110 to command the signal mode. Group 118 triggers and/or reads signals. In a specific embodiment, the connector 120 may be a physical circuit. In another embodiment, the connector 120 may be a wireless connection, such as using Wi-Fi or Bluetooth technology.
圖2進一步詳細說明處理器200。如圖2所示，處理器200包含一類比至數位轉換器（A/D轉換器）220、信號處理器222、特徵提取器224、神經網絡230、顯示器240及記憶體250。Figure 2 illustrates processor 200 in further detail. As shown in FIG. 2 , the processor 200 includes an analog-to-digital converter (A/D converter) 220 , a signal processor 222 , a feature extractor 224 , a neural network 230 , a display 240 and a memory 250 .
應注意，所述之該處理器200之組件可位在單一裝置上，如圖2中所說明，或位在分開的裝置或在雲端上。例如，該類比至數位轉換器220、該信號處理器222、該特徵提取器224、該神經網絡230及/或記憶體250可各自或以各種組合為一獨立裝置或雲端的部份。此外，顯示器240可係一獨立裝置，其經配置以允許使用者輸入至處理器200中，諸如經由螢幕上輸入，以及顯示諸如生理測量及/或HbA1C之資訊。It should be noted that the components of the processor 200 may be located on a single device, as illustrated in Figure 2, or on separate devices or in the cloud. For example, the analog-to-digital converter 220, the signal processor 222, the feature extractor 224, the neural network 230, and/or the memory 250 may be individually or in various combinations as a stand-alone device or part of a cloud. Additionally, display 240 may be a stand-alone device configured to allow user input into processor 200, such as via on-screen input, and to display information such as physiological measurements and/or HbA1C.
該A/D轉換器220經配置以將傳輸至該處理器200之類比信號數位化成數位信號252，其可儲存在記憶體250中。該信號處理器222經配置以處理該數位化信號252以便於提取來自該PPG信號的特徵。例如，在一具體實施例中，該信號處理器222可經配置以將該信號分解成AC及DC組份，進行傅立葉轉換等等，以促進PPG信號之分析及進一步處理，諸如將該PPG信號數位化成信號窗，及自該數位化信號252提取特徵256，如以下結合圖4進一步詳細描述。該信號處理結果254可儲存在記憶體250中。The A/D converter 220 is configured to digitize the analog signal transmitted to the processor 200 into a digital signal 252 , which can be stored in the memory 250 . The signal processor 222 is configured to process the digitized signal 252 in order to extract features from the PPG signal. For example, in one embodiment, the signal processor 222 may be configured to decompose the signal into AC and DC components, perform Fourier transforms, etc., to facilitate analysis and further processing of the PPG signal, such as converting the PPG signal into Digitize into signal windows, and extract features 256 from the digitized signal 252, as described in further detail below in conjunction with FIG. 4 . The signal processing results 254 can be stored in the memory 250 .
圖3A-D說明本發明之神經網絡230之各種具體實施例。如圖3A所示，本發明之該神經網絡230的一具體實施例包含一輸入層、一或多個隱藏層及一輸出層。在一具體實施例中，該輸入層經配置以接收一或多個輸入，諸如一或多個輸入向量或矩陣。在一具體實施例中，各隱藏層包含一或多個節點，其中各節點具有一相關權重及偏權，其可在該模型之訓練階段期間調整以最佳化預測精確性。在一具體實施例中，各節點連接至不同隱藏層之一或多個節點。在一具體實施例中，若任何個別節點之輸出高於特定臨限值，則啟用該節點，傳送資料至該網絡的下一層；否則，無資料傳遞至該網絡的下一層。在一具體實施例中，本發明之神經網絡包含任何數目的隱藏層。在一具體實施例中，本發明之神經網絡包含1至10個隱藏層，諸如1層、2層、3層、4層、5層、6層、7層、8層、9層或10層，包括落在此等值內的所有範圍及數目。Figures 3A-D illustrate various embodiments of the neural network 230 of the present invention. As shown in Figure 3A, a specific embodiment of the neural network 230 of the present invention includes an input layer, one or more hidden layers and an output layer. In a specific embodiment, the input layer is configured to receive one or more inputs, such as one or more input vectors or matrices. In one embodiment, each hidden layer includes one or more nodes, where each node has an associated weight and bias weight that can be adjusted during the training phase of the model to optimize prediction accuracy. In a specific embodiment, each node is connected to one or more nodes in different hidden layers. In a specific embodiment, if the output of any individual node is higher than a certain threshold, the node is enabled and data is transmitted to the next layer of the network; otherwise, no data is transmitted to the next layer of the network. In a specific embodiment, the neural network of the present invention includes any number of hidden layers. In a specific embodiment, the neural network of the present invention includes 1 to 10 hidden layers, such as 1 layer, 2 layers, 3 layers, 4 layers, 5 layers, 6 layers, 7 layers, 8 layers, 9 layers or 10 layers. , including all ranges and numbers falling within this value.
在一具體實施例中，本發明之NN 230的輸入可包含一或多個向量或矩陣，其依次包含該PPG信號之信號窗、使用習知之手指扎刺法測量之HbA1c、各種特徵或其組合。在一具體實施例中，本發明之NN 230的輸入包含一或多個一維向量，其依次包含該PPG信號之信號窗、使用習知之手指扎刺法測量之HbA1c、各種特徵或其組合。該等特徵在下文中進一步詳細討論。In a specific embodiment, the input of the NN 230 of the present invention may include one or more vectors or matrices, which in turn include the signal window of the PPG signal, HbA1c measured using the conventional finger stick method, various features, or a combination thereof . In a specific embodiment, the input of the NN 230 of the present invention includes one or more one-dimensional vectors, which in turn include the signal window of the PPG signal, HbA1c measured using the conventional finger stick method, various features, or a combination thereof. These features are discussed in further detail below.
在一具體實施例中，本發明之NN 230係使用訓練資料訓練，該訓練資料係諸如PPG信號及Hba1c及/或使用習知手指扎刺法獲得之參考血糖值，其來自未接受可影響他或她的心血管系統之經美國食品藥物管理局及/或歐洲藥品管理局所認定對心血管系統具有此一效應的任何醫療，及/或未接受用於治療糖尿病之任何醫療的訓練定群。在一具體實施例中，本發明之NN 230係使用訓練資料訓練，諸如來自未接受任何醫療之訓練定群之PPG信號及HbA1c及/或使用習知手指扎刺法獲得之參考血糖值。在一具體實施例中，訓練本發明之NN 230包含最小化如藉由在藉由本發明之系統預測之BG及該對應參考BG之間的總差異計算之損失。用於建構本發明之NN 230之各種神經網絡框架包含Caffee、Keras、Microsoft Cognitive Toolkit、MXNET、DeepLearning4j、Chainer及TensorFlow。In a specific embodiment, the NN 230 of the present invention is trained using training data, such as PPG signals and Hba1c and/or reference blood glucose values obtained using conventional finger stick methods, which are derived from unaccepted data that may affect other or any medical treatment on her cardiovascular system that has been determined by the US Food and Drug Administration and/or the European Medicines Agency to have such an effect on the cardiovascular system, and/or has not received training in any medical treatment for the treatment of diabetes. In a specific embodiment, the NN 230 of the present invention is trained using training data, such as PPG signals and HbA1c from a training cohort that did not receive any medical treatment and/or reference blood glucose values obtained using conventional finger stick methods. In a specific embodiment, training the NN 230 of the present invention includes minimizing the loss as calculated by the total difference between the BG predicted by the system of the present invention and the corresponding reference BG. Various neural network frameworks used to construct the NN 230 of the present invention include Caffee, Keras, Microsoft Cognitive Toolkit, MXNET, DeepLearning4j, Chainer, and TensorFlow.
在一具體實施例中，本發明之神經網絡230可包含一全連結神經網絡（FCNN）301。本發明之全連結神經網絡（FCNN）的一具體實施例在圖3D中說明。如圖3D所示，本發明之FCNN的一具體實施例包含一輸入層、一隱藏層及一輸出層。在一具體實施例中，該輸入層經配置以接收一或多個輸入，諸如一或多個向量或矩陣。在一具體實施例中，該輸入層經配置以接收個人生理特徵、脈搏型態特徵、心率變化特徵、HbA1c或其組合作為輸入。在一具體實施例中，如圖3D所示，該輸入層經配置以接收特徵向量F 340作為輸入，其包含16個特徵及HbA1c，如以下結合該等實例進一步詳細描述。如圖3D所示，本發明之FCNN包含一或多個隱藏層，其中各隱藏層包含一密集層306。在一具體實施例中，該密集層306包含一或多個節點，其中各節點連接至上一層之各節點。在一具體實施例中，該密集層之節點進行操作，其減少該輸入的大小以幫助識別用於預測BG之重要特徵。在一具體實施例中，該密集層之節點進行矩陣-向量乘法，其中可訓練該矩陣值，且可藉由反向傳播更新。如圖3D所示，本發明之FCNN進一步包含一輸出層，其經配置以輸出BG預測。In a specific embodiment, the neural network 230 of the present invention may include a fully connected neural network (FCNN) 301. A specific embodiment of the fully connected neural network (FCNN) of the present invention is illustrated in Figure 3D. As shown in Figure 3D, a specific embodiment of the FCNN of the present invention includes an input layer, a hidden layer and an output layer. In a specific embodiment, the input layer is configured to receive one or more inputs, such as one or more vectors or matrices. In a specific embodiment, the input layer is configured to receive as input a personal physiological characteristic, a pulse pattern characteristic, a heart rate variation characteristic, HbA1c, or a combination thereof. In a specific embodiment, as shown in Figure 3D, the input layer is configured to receive a feature vector F 340 as input, which includes 16 features and HbA1c, as described in further detail below in connection with the examples. As shown in Figure 3D, the FCNN of the present invention includes one or more hidden layers, where each hidden layer includes a dense layer 306. In a specific embodiment, the dense layer 306 includes one or more nodes, where each node is connected to each node in the previous layer. In a specific embodiment, the nodes of the dense layer perform operations that reduce the size of the input to help identify important features for predicting BG. In a specific embodiment, the nodes of the dense layer perform matrix-vector multiplications, where the matrix values can be trained and updated by backpropagation. As shown in Figure 3D, the FCNN of the present invention further includes an output layer configured to output BG predictions.
在一具體實施例中，本發明之神經網絡230可包含一卷積神經網絡（CNN）。本發明之CNN 302的一具體實施例在圖3B中說明。如圖3B所示，在一具體實施例中，本發明之CNN 302的隱藏層包含卷積模組310。該卷積模組310包含一或多個卷積層312，諸如312a、312b及312c，如圖3B所示。在一具體實施例中，該卷積模組310可包含任何數目之卷積層。在一具體實施例中，該卷積模組310可包含1至10層卷積層。In a specific embodiment, the neural network 230 of the present invention may include a convolutional neural network (CNN). A specific embodiment of the CNN 302 of the present invention is illustrated in Figure 3B. As shown in Figure 3B, in a specific embodiment, the hidden layer of the CNN 302 of the present invention includes a convolution module 310. The convolution module 310 includes one or more convolution layers 312, such as 312a, 312b, and 312c, as shown in Figure 3B. In a specific embodiment, the convolution module 310 may include any number of convolution layers. In a specific embodiment, the convolution module 310 may include 1 to 10 convolution layers.
在一具體實施例中，各卷積層312包含一卷積子模組334、批次正規化模組336、啟動功能模組338、池化模組339或其組合。在一具體實施例中，該啟動模組338經配置以進行ReLU操作。在一具體實施例中，該池化模組339經配置以進行最大池化或平均池化操作。在一具體實施例中，該卷積子模組334經配置以使用一或多個過濾器對來自另一卷積層之該一或多個輸入或輸出進行卷積操作。在一具體實施例中，該過濾器可具有任何長度。在一具體實施例中，該過濾器可具有1至500之長度，諸如1、5、10、15、20、30、40、60、80、100、200、250、300、350、400、450或500，包括落在此等值之所有範圍及數目。In a specific embodiment, each convolution layer 312 includes a convolution sub-module 334, a batch normalization module 336, a startup function module 338, a pooling module 339, or a combination thereof. In a specific embodiment, the startup module 338 is configured to perform ReLU operations. In a specific embodiment, the pooling module 339 is configured to perform a max pooling or average pooling operation. In a specific embodiment, the convolution sub-module 334 is configured to convolve the one or more inputs or outputs from another convolutional layer using one or more filters. In a specific embodiment, the filter can be of any length. In a specific embodiment, the filter may have a length of 1 to 500, such as 1, 5, 10, 15, 20, 30, 40, 60, 80, 100, 200, 250, 300, 350, 400, 450 or 500, including all ranges and numbers falling within this value.
在一具體實施例中，至本發明之CNN 302之輸入可包含一或多個向量或矩陣，其依次包含源自獲自該受試者之PPG信號之信號窗、使用習知之手指扎刺法測量之HbA1c、各種特徵或其組合，如上所述。在一具體實施例中，至本發明之CNN 302之輸入包含一維向量，其依次包含源自PPG信號的信號窗、使用習知之手指扎刺法測量之HbA1c、各種特徵或其組合，如上文所揭示。該卷積子模組334之輸出可藉由該批次正規化模組336、啟動功能338及/或池化模組339進一步處理。In a specific embodiment, the input to the CNN 302 of the present invention may include one or more vectors or matrices, which in turn include signal windows derived from the PPG signal obtained from the subject, using the conventional finger prick method. Measured HbA1c, various characteristics, or combinations thereof, as described above. In a specific embodiment, the input to the CNN 302 of the present invention includes a one-dimensional vector, which in turn includes the signal window derived from the PPG signal, HbA1c measured using the conventional finger stick method, various features, or combinations thereof, as above revealed. The output of the convolution sub-module 334 may be further processed by the batch normalization module 336, activation function 338 and/or pooling module 339.
該卷積模組310可進一步包含一攤平模組314及一或多個密集層316。在一具體實施例中，該攤平模組314經配置以減少該卷積層312之輸出的大小。各卷積模組310進一步包含一或多個密集層316，其經配置以進一步識別自攤平層314之輸出之重要部份。在一具體實施例中，各密集層316包含一密集啟動模組344、一漏失（dropout）模組346及/或一批次正規化模組348。在一具體實施例中，第二輸入包含使用習知之手指扎刺法測量之HbA1c，並且各種特徵可分開地輸入於該合併層352中，其經配置以接收及處理來自密集層316a及特徵向量340之輸出。The convolution module 310 may further include a flattening module 314 and one or more dense layers 316. In a specific embodiment, the flattening module 314 is configured to reduce the size of the output of the convolutional layer 312 . Each convolution module 310 further includes one or more dense layers 316 configured to further identify important portions of the output from the flattening layer 314 . In a specific embodiment, each dense layer 316 includes a dense activation module 344, a dropout module 346, and/or a batch normalization module 348. In one embodiment, the second input includes HbA1c measured using the conventional finger stick method, and the various features can be separately input into the merging layer 352, which is configured to receive and process the feature vectors from the dense layer 316a 340 output.
圖3C說明本發明之CNN 303的另一具體實施例，其具有兩個卷積模組。如圖3C所示，在一具體實施例中，本發明之CNN 303的隱藏層包含兩個平行的卷積模組320及330，其中各卷積模組具有不同長度之過濾器。較小過濾器長度經配置以擷取該輸入向量更詳細等級的態樣，而較大過濾器長度經配置以擷取該輸入向量更高等級的態樣。因此，在一具體實施例中，具有較短長度過濾器之卷積模組可稱為微卷積模組320，而具有較長長度過濾器之卷積模組可稱為巨卷積模組330。Figure 3C illustrates another specific embodiment of the CNN 303 of the present invention, which has two convolution modules. As shown in Figure 3C, in a specific embodiment, the hidden layer of the CNN 303 of the present invention includes two parallel convolution modules 320 and 330, wherein each convolution module has filters of different lengths. Smaller filter lengths are configured to capture more detailed level aspects of the input vector, while larger filter lengths are configured to capture higher level aspects of the input vector. Therefore, in a specific embodiment, the convolution module with a shorter length filter may be called a micro-convolution module 320, and the convolution module with a longer length filter may be called a macro-convolution module. 330.
在一具體實施例中，各卷積模組320、330之結構與卷積模組310相同，但用於該微卷積模組之過濾器或該微過濾器321之長度係較該巨過濾器331之長度短。在一具體實施例中，該微過濾器321之長度係該巨過濾器331的大約1/4至3/4，諸如1/4、1/2或3/4，包括落在此等值之任何範圍及數目。在一具體實施例中，該巨過濾器可具有任何長度。在一具體實施例中，該巨過濾器可具有1至500之長度，諸如1、5、10、15、20、30、40、60、80、100、200、250、300、350、400、450或500，包括落在此等值之所有範圍及數目。In a specific embodiment, the structure of each convolution module 320 and 330 is the same as that of the convolution module 310, but the length of the filter used in the micro-convolution module or the micro-filter 321 is longer than that of the macro filter. The length of device 331 is short. In a specific embodiment, the length of the micro filter 321 is approximately 1/4 to 3/4 of the macro filter 331, such as 1/4, 1/2 or 3/4, including those within these values. Any range and number. In a specific embodiment, the giant filter can be of any length. In a specific embodiment, the giant filter may have a length of 1 to 500, such as 1, 5, 10, 15, 20, 30, 40, 60, 80, 100, 200, 250, 300, 350, 400, 450 or 500, including all ranges and numbers falling within this range.
在一具體實施例中，該微卷積模組320包含一或多個微卷積層322，諸如322a、322b及322c，如圖3C所示。該巨卷積模組330包含一或多個巨卷積層332，諸如332a、332b及332c如圖3C所示。在一具體實施例中，各卷積模組320、330可包含1至10個卷積層。In a specific embodiment, the micro-convolution module 320 includes one or more micro-convolution layers 322, such as 322a, 322b and 322c, as shown in Figure 3C. The giant convolution module 330 includes one or more giant convolution layers 332, such as 332a, 332b and 332c as shown in Figure 3C. In a specific embodiment, each convolution module 320, 330 may include 1 to 10 convolution layers.
各卷積模組320、330可進一步包含一攤平模組324、334，一或多個密集層326、336或其組合。在一具體實施例中，該攤平模組324、334經配置以減少該卷積層322、332之輸出的大小。在一具體實施例中，各卷積模組320、330可進一步包含一或多個密集層326、336，其經配置以進一步識別自攤平層324、334輸出的重要部份。在一具體實施例中，各密集層326、336包含一密集啟動模組344、一漏失模組346、一批次正規化模組348或其組合。Each convolution module 320, 330 may further include a flattening module 324, 334, one or more dense layers 326, 336, or a combination thereof. In a specific embodiment, the flattening module 324, 334 is configured to reduce the size of the output of the convolutional layer 322, 332. In a specific embodiment, each convolution module 320, 330 may further include one or more dense layers 326, 336 configured to further identify important portions of the output from the flattening layers 324, 334. In a specific embodiment, each dense layer 326, 336 includes a dense activation module 344, a dropout module 346, a batch normalization module 348, or a combination thereof.
在一具體實施例中，該卷積子模組334經配置以使用一或多個過濾器對來自另一卷積層之該一或多個輸入或輸出進行卷積操作。In a specific embodiment, the convolution sub-module 334 is configured to convolve the one or more inputs or outputs from another convolutional layer using one or more filters.
在一具體實施例中，本發明之CNN 303進一步包含一合併模塊模組350，其經配置以接收特徵向量或矩陣340，以及該微卷積模組及巨卷積模組320、330之結果，以基於此等輸入分析及輸出該受試者之一預測之血糖值。在一具體實施例中，該合併模塊模組350包含一合併層352，其合併來自該兩個CNN模組及特徵向量340的輸出以用於進一步處理。在一具體實施例中，該合併模塊模組350進一步包含一或多個密集層354，其經配置以處理該合併之輸入以針對該受試者輸出一預測血糖值。在一具體實施例中，該特徵向量或矩陣340可包含本文所揭示之各種特徵，包括個人生理特徵、脈搏型態特徵、心率變化特徵或其組合。In a specific embodiment, the CNN 303 of the present invention further includes a merge module module 350 configured to receive the feature vector or matrix 340 and the results of the micro-convolution module and the macro-convolution module 320, 330 , to analyze and output a predicted blood glucose value of the subject based on these inputs. In a specific embodiment, the merging module module 350 includes a merging layer 352 that combines the outputs from the two CNN modules and feature vectors 340 for further processing. In a specific embodiment, the merge module module 350 further includes one or more dense layers 354 configured to process the merged input to output a predicted blood glucose value for the subject. In a specific embodiment, the feature vector or matrix 340 may include various features disclosed herein, including personal physiological features, pulse pattern features, heart rate variation features, or combinations thereof.
該特徵提取器224經配置以自該數位化信號252及/或該信號處理254之結果提取特徵。在一具體實施例中，所提取之特徵256可包含本文所揭示之任何特徵，及可用於產生該特徵向量340作為對本發明之NN 230之各種具體實施例的輸入。The feature extractor 224 is configured to extract features from the digitized signal 252 and/or the results of the signal processing 254 . In one embodiment, the extracted features 256 may include any of the features disclosed herein, and may be used to generate the feature vector 340 as input to various embodiments of the NN 230 of the present invention.
在一具體實施例中，對本發明之NN 230之任何具體實施例的輸入可包含數個輸入類型，諸如獲自該受試者之PPG信號的數位化區段，使用習知之手指扎刺法測量之HbA1c、各種特徵或其組合。在一具體實施例中，本發明之NN 230的各種輸入類型可包含一或多個輸入向量或矩陣。例如，該PPG信號之數位化區段可包含一輸入向量或矩陣，而該等特徵包含一不同之輸入向量或矩陣，諸如該特徵向量340。替代地，一或多個輸入類型可包含一單一輸入向量或矩陣，使得該PPG信號之數位化區段及該等特徵經序連以形成一輸入向量或矩陣。在一具體實施例中，各輸入向量或矩陣包含一維（1d）輸入向量。在另一具體實施例，該PPG信號包含一輸入向量，而包括HbA1c之該等特徵包含一分開之向量，諸如該特徵向量340，其中該PPG信號經直接輸入至該卷積模組，且該等特徵經分開地輸入至諸如該合併層352中。In one embodiment, the input to any embodiment of the NN 230 of the present invention may include several input types, such as digitized segments of the PPG signal obtained from the subject, measured using the conventional finger stick method. HbA1c, various characteristics, or combinations thereof. In a specific embodiment, various input types of the NN 230 of the present invention may include one or more input vectors or matrices. For example, the digitized portion of the PPG signal may include one input vector or matrix, and the features include a different input vector or matrix, such as the feature vector 340 . Alternatively, one or more input types may include a single input vector or matrix such that the digitized segments of the PPG signal and the features are concatenated to form an input vector or matrix. In a specific embodiment, each input vector or matrix includes a one-dimensional (1d) input vector. In another embodiment, the PPG signal includes an input vector and the features including HbA1c include a separate vector, such as the feature vector 340, where the PPG signal is directly input to the convolution module, and the The features are input separately into, for example, the merging layer 352.
圖4A說明一示例性PPG信號。在一具體實施例中，該PPG信號之數位化區段包含信號窗，如圖4D所示。在一具體實施例中，各信號窗包含大約1秒至大約20秒（諸如大約1秒、大約1.6秒、大約2秒、大約5秒、大約10秒、大約15秒或大約20秒）之一固定時間長度的PPG信號，或由其等組成。在一具體實施例中，各信號窗自該PPG信號的各波谷開始，使得若一受試者之心搏為每秒60下時，1分鐘長的PPG信號將造成60個信號窗。在一具體實施例中，該信號窗可使用一頻率過濾器較佳地定義，如圖4B及4C中所說明，以較佳地識別該PPG信號之峰及谷，如圖4D所示。在一具體實施例中，使用該Bigger-Fall-Side演算法識別該PPG波形之波谷[26]。然後從各波谷提取含有該脈搏之向後一秒長的區段（總共250個資料點）。將該等脈搏平均化，其用來表示整個分鐘之PPG信號以用於深度學習神經網絡。本發明之信號窗的一實例在圖4D中說明。Figure 4A illustrates an exemplary PPG signal. In a specific embodiment, the digitized section of the PPG signal includes a signal window, as shown in Figure 4D. In a specific embodiment, each signal window includes one of about 1 second to about 20 seconds (such as about 1 second, about 1.6 seconds, about 2 seconds, about 5 seconds, about 10 seconds, about 15 seconds, or about 20 seconds) A PPG signal of fixed length of time, or consisting of it. In one embodiment, each signal window starts from each trough of the PPG signal, such that if a subject's heartbeat is 60 beats per second, a 1 minute long PPG signal will result in 60 signal windows. In a specific embodiment, the signal window may be better defined using a frequency filter, as illustrated in Figures 4B and 4C, to better identify the peaks and valleys of the PPG signal, as shown in Figure 4D. In a specific embodiment, the Bigger-Fall-Side algorithm is used to identify the trough of the PPG waveform [26]. Then a segment containing one second of the pulse (a total of 250 data points) is extracted from each wave trough. These pulses are averaged and used to represent the entire minute of PPG signal for use in deep learning neural networks. An example of a signal window of the present invention is illustrated in Figure 4D.
在一具體實施例中，特徵可包含個人生理特徵、脈搏型態特徵、心率變化特徵或其組合。在一具體實施例中，該脈搏型態特徵及心率變化特徵亦可稱作經提取特徵256，因為此等包含源自該等PPG信號之特徵。在一具體實施例中，該個人生理特徵251可包含年紀、腰圍、身體質量指數、收縮壓、舒張壓或其組合。在一具體實施例中，源自該平均PPG脈搏之脈搏型態特徵可包含50%高度之脈搏寬度、該分鐘之總脈搏面積、平均脈搏面積、該脈搏面積之中位數，或從脈搏波谷至波峰之時間差或其組合。在一具體實施例中，源自該PPG窗之心率變化特徵可包含來自快速傅立葉轉換（FFT）之低頻功率、來自FFT之高頻功率、來自FFT之總功率、脈搏連續間隔變化超過20ms之百分比、連續間隔變化之標準差或其組合。在一具體實施例中，以上所列之所有17個特徵（對於包括HbA1c之模型再加上HbA1c）聚結成一特徵向量F 340作為至本發明之神經網絡的輸入。在另一具體實施例，亦可包括其它特徵作為特徵，諸如峰及重搏波位置、峰值振幅、該信號波形之第一及第二之衍生物上之波峰及波谷位置。In a specific embodiment, the characteristics may include personal physiological characteristics, pulse pattern characteristics, heart rate variation characteristics, or a combination thereof. In a specific embodiment, the pulse pattern features and heart rate variation features may also be referred to as extracted features 256 because they include features derived from the PPG signals. In a specific embodiment, the personal physiological characteristics 251 may include age, waist circumference, body mass index, systolic blood pressure, diastolic blood pressure, or a combination thereof. In a specific embodiment, the pulse pattern characteristics derived from the average PPG pulse may include the pulse width at 50% height, the total pulse area for the minute, the average pulse area, the median pulse area, or from the pulse trough. The time difference to the peak of the wave or its combination. In a specific embodiment, the heart rate variation characteristics derived from the PPG window may include low-frequency power from fast Fourier transform (FFT), high-frequency power from FFT, total power from FFT, and the percentage of continuous pulse interval changes exceeding 20 ms. , the standard deviation of changes in consecutive intervals, or a combination thereof. In a specific embodiment, all 17 features listed above (plus HbA1c for models including HbA1c) are coalesced into a feature vector F 340 as input to the neural network of the present invention. In another embodiment, other features may also be included as features, such as peak and dicrotic wave positions, peak amplitudes, and peak and trough positions on the first and second derivatives of the signal waveform.
本發明亦提供一種用於基於PPG之NIBG神經網絡預測的方法。圖6及7說明本發明之方法。如圖6及7所示，本發明之方法分別包含訓練階段及預測階段。The present invention also provides a method for NIBG neural network prediction based on PPG. Figures 6 and 7 illustrate the method of the invention. As shown in Figures 6 and 7, the method of the present invention includes a training stage and a prediction stage respectively.
圖6說明本發明之方法的訓練階段。本發明之方法的訓練階段開始於自訓練定群獲得一組訓練資料組以用於訓練本發明之神經網絡230。此包含步驟1000，其使用信號裝置110自一訓練定群獲取PPG信號，以及步驟1005，其輸入該訓練定群每個人的生理特徵，諸如BMI、年紀、腰圍、身高、體重等等。此外，步驟1000涉及獲得參考資料。在一具體實施例中，該參考資料組包含使用習知之手指扎刺或抽血法自該訓練定群每個人獲得之HbA1c及血糖值，以用於輸入以及用於訓練本發明之神經網絡230。Figure 6 illustrates the training phase of the method of the invention. The training phase of the method of the present invention begins with obtaining a set of training data from the training cohort for training the neural network 230 of the present invention. This includes step 1000, which uses the signaling device 110 to obtain PPG signals from a training group, and step 1005, which inputs the physiological characteristics of each person in the training group, such as BMI, age, waist circumference, height, weight, etc. Additionally, step 1000 involves obtaining reference materials. In a specific embodiment, the reference set includes HbA1c and blood glucose values obtained from each individual in the training cohort using conventional finger stick or blood draw methods for input and for training the neural network 230 of the present invention. .
在一具體實施例中，在該訓練定群存在超過100、500或1000個人。較佳地，該訓練定群包含具有各種性別、年紀及生理條件之人的多樣性。在一具體實施例中，該訓練定群每個人未接受任何醫療。在另一具體實施例，該訓練定群每個人未接受可影響他或她的心血管系統之任何醫療及/或未接受任何針對糖尿病之醫療。In a specific embodiment, there are more than 100, 500 or 1000 people in the training cohort. Preferably, the training group includes a diversity of people of various genders, ages, and physical conditions. In one embodiment, no one in the training cohort receives any medical treatment. In another embodiment, the training cohort includes each individual who does not receive any medical treatment that affects his or her cardiovascular system and/or does not receive any medical treatment for diabetes.
若該PPG信號以類比格式收集，則在步驟1010中藉由A/D轉換器220將該信號數位化。該數位化信號252可儲存在資料庫250中作為步驟1010之部份。接下來，在步驟1020中藉由該信號處理器222將該數位化信號252進行信號處理。在一具體實施例中，如上文所提及，該信號處理步驟1020可包含將該信號分解成高頻部份及低頻部份，諸如自該DC組份分離其AC組份。在另一具體實施例，該信號處理步驟1020可包含使用以下轉換方法轉換該信號：諸如傅立葉轉換、小波轉換、希伯特－黃（Hilbert-Huang）轉換或任何其它與任何時間-頻率分析相關的轉換。在一具體實施例中，將該數位化信號處理成信號窗，其係如上所述之該PPG信號的數位化區段。該信號處理1020之結果可儲存在記憶體250中作為步驟1020之部份。If the PPG signal is collected in analog format, the signal is digitized by the A/D converter 220 in step 1010 . The digitized signal 252 may be stored in the database 250 as part of step 1010 . Next, in step 1020, the signal processor 222 performs signal processing on the digitized signal 252. In a specific embodiment, as mentioned above, the signal processing step 1020 may include decomposing the signal into a high frequency component and a low frequency component, such as separating its AC component from the DC component. In another embodiment, the signal processing step 1020 may include converting the signal using a conversion method such as Fourier transform, wavelet transform, Hilbert-Huang transform, or any other method related to any time-frequency analysis. conversion. In a specific embodiment, the digitized signal is processed into a signal window, which is a digitized section of the PPG signal as described above. The results of signal processing 1020 may be stored in memory 250 as part of step 1020 .
接下來，在步驟1030中，該特徵提取器224從該數位化信號252及/或該經處理信號254提取經提取之特徵256，諸如脈搏型態特徵，包含在50%高度之脈搏寬度、該分鐘之總脈搏面積、平均脈搏面積、該脈搏面積之中位數、自脈搏波谷至波峰之時間差異或其組合，以及心率變化特徵，包含自快速傅立葉轉換（FFT）之低及高頻功率兩者、來自FFT之總功率、脈搏連續間隔變化超過20ms之百分比、連續間隔變化之標準差或其組合。所提取特徵256可儲存在記憶體250作為步驟1030之部份。Next, in step 1030, the feature extractor 224 extracts extracted features 256 from the digitized signal 252 and/or the processed signal 254, such as pulse pattern features, including pulse width at 50% height, the The total pulse area in minutes, the average pulse area, the median of the pulse area, the time difference from pulse trough to peak, or a combination thereof, and heart rate change characteristics, including both low and high frequency power from fast Fourier transform (FFT) Or, the total power from FFT, the percentage of continuous pulse interval changes exceeding 20 ms, the standard deviation of continuous interval changes, or a combination thereof. The extracted features 256 may be stored in memory 250 as part of step 1030 .
在提取所提取特徵256之後，可使用在先前步驟中獲得及衍生之訓練資料在步驟1040中訓練所揭示之該基於PPG之NIBG預測系統的任何具體實施例以估計血糖值。訓練方法為本領域所熟知，諸如在註腳1中所引用之參考文獻中討論者，其全文併入本文中。在一具體實施例中，訓練應導致對在該訓練定群中的每人在參考BG資料及對應BG預測之間的損失總和最小化，其係藉由修改在本發明之基於PPG之NIBG預測系統之該神經網絡230之任何具體實施例內的權重及偏權，諸如該等過濾器、密集層等等。After extracting the extracted features 256, any specific embodiment of the disclosed PPG-based NIBG prediction system may be trained in step 1040 to estimate blood glucose values using the training data obtained and derived in the previous steps. Training methods are well known in the art, such as those discussed in the references cited in footnote 1, the entire contents of which are incorporated herein. In a specific embodiment, training should result in minimizing the sum of losses between reference BG data and corresponding BG predictions for each person in the training cohort by modifying the PPG-based NIBG prediction of the present invention. The weights and biases within any embodiment of the neural network 230 of the system, such as the filters, dense layers, etc.
在該訓練階段後，本發明之基於PPG之NIBG預測系統準備好預測未接受任何醫療之受試者的血糖值，如圖7中所說明。在該訓練階段之後，本發明之基於PPG之NIBG預測系統亦可用於預測未接受可影響該受試者之心血管系統之任何醫療之受試者及/或未接受任何糖尿病醫療之受試者的血糖值，如圖7中所說明。如圖7所示，用於估計血糖值之方法包含步驟2000資料獲得、步驟2005使用者輸入、步驟2010數位化信號、步驟2020信號處理及步驟2030特徵提取。此等步驟與訓練程序步驟相同：步驟1000資料獲得、步驟1005使用者輸入、步驟1010數位化信號、步驟1020信號處理及步驟1030特徵提取，除了其等係針對一特定使用者（受試者100）進行，該受試者對基於PPG之NIBG預測感興趣。在步驟2040中，該經訓練之模型係用於進行NIBG預測。 實例 材料及方法 After this training phase, the PPG-based NIBG prediction system of the present invention is ready to predict the blood glucose values of subjects who have not received any medical treatment, as illustrated in Figure 7. After this training phase, the PPG-based NIBG prediction system of the present invention can also be used to predict subjects who have not received any medical treatment that can affect the cardiovascular system of the subject and/or subjects who have not received any medical treatment for diabetes. blood glucose values, as illustrated in Figure 7. As shown in Figure 7, the method for estimating blood glucose level includes step 2000 data acquisition, step 2005 user input, step 2010 digital signal, step 2020 signal processing and step 2030 feature extraction. These steps are the same as those of the training program: step 1000 data acquisition, step 1005 user input, step 1010 digital signal, step 1020 signal processing and step 1030 feature extraction, except that they are for a specific user (subject 100 ), the subject is interested in NIBG prediction based on PPG. In step 2040, the trained model is used for NIBG prediction. Example materials and methods
在本研究中，使用所有2538名受試者建立圖3C的模型作為基準線以代表對大樣本數使用單一通用模型的習知方法。接著我們將受試者基於他們是否接受任何種類之醫療（包括胰島素注射、心血管治療等等）而分成兩個定群。醫療包括由美國食品藥物管理局及/或歐洲藥品管理局規範為醫療程序、藥物或用藥方案的任何醫療程序、藥物或用藥方案。有接受醫療及未接受醫療之受試者的定群分別由1682及856名受試者組成。各定群之特性總結在表1中。第一組模型針對每一定群建構以與通用模型比較。接著第二組模型針對每一定群加上其HbA1c值作為新特徵來建構以評估性能改善。所有五個模型使用圖3C中所說明之完全相同的CNN架構，以用於涉及醫療及HbA1c的公平比較。本研究之想法的一視覺表示呈現在圖5A中。
此外，具有一卷積模組之圖3B的模型而非圖3C之模型的兩個平行卷積模組係用於訓練以及測試BG預測，其使用藉由有接受醫療及未接受醫療之標準決定之兩組受試者及訓練定群獲得之資料。使用四種輸入類型，包括：1. 僅PPG信號，2. PPG信號加上如在該實驗設置章節中所述以及結合特徵向量340討論所列出的16個特徵，3. PPG信號加上HbA1c，以及4. PPG信號、HbA1c及該等16個特徵。In addition, the model of Figure 3B with one convolution module instead of the model of Figure 3C with two parallel convolution modules is used to train and test BG prediction, and its use is determined by the criteria of receiving medical treatment and not receiving medical treatment. Data obtained from two groups of subjects and training groups. Four input types were used, including: 1. PPG signal only, 2. PPG signal plus the 16 features listed in this experimental setup chapter and discussed in conjunction with feature vector 340, 3. PPG signal plus HbA1c , and 4. PPG signal, HbA1c and these 16 characteristics.
此外，圖3D之全連結神經網絡模型係用於訓練以及測試BG預測，其使用藉由有接受醫療及未接受醫療之標準決定之兩組受試者及訓練定群的資料。使用兩個類型之輸入向量，包括：1. 16個特徵，如在該實驗設置章節所述中列出，2. HbA1c及該等16個特徵。 實驗設置 In addition, the fully connected neural network model in Figure 3D was used to train and test BG prediction using data from two groups of subjects determined by the criteria of those who received medical treatment and those who did not receive medical treatment and the training cohort. Two types of input vectors were used, including: 1. 16 features, as listed in the experimental setup section, 2. HbA1c and these 16 features. Experimental setup
本研究之樣本收集已獲台灣中央研究院之研究倫理委員會（Institutional Review Board of Academia Sinica）核准（申請案號：AS-IRB01-16081）。該等樣本收集自總計2538名自願受試者。所有受試者均知情並已同意收集資料及其用途。針對各受試者記錄兩個連續一分鐘長之PPG信號區段，以及其生理資訊，諸如年紀、體形、血壓、HbA1c及BG值。詳細之實驗設置及程序可見於[25]中。The sample collection for this study has been approved by the Institutional Review Board of Academia Sinica, Taiwan (Application No.: AS-IRB01-16081). The samples were collected from a total of 2538 voluntary subjects. All subjects informed and consented to the collection of data and its use. Two continuous one-minute long PPG signal segments were recorded for each subject, as well as their physiological information, such as age, body shape, blood pressure, HbA1c and BG values. Detailed experimental setup and procedures can be found in [25].
每分鐘長之PPG信號首先分段成用於提取型態特徵及心率變化（HRV）特徵兩者之信號窗。首先，使用該Bigger-Fall-Side演算法識別該PPG波形之波谷[26]。然後從各波谷提取含有該脈搏之向後一秒長的區段（總共250個資料點）。對該脈搏作平均，其用於表示整個分鐘之PPG信號以用於深度學習神經網絡。The minute-long PPG signal is first segmented into signal windows for extracting both pattern features and heart rate variability (HRV) features. First, use the Bigger-Fall-Side algorithm to identify the trough of the PPG waveform [26]. Then a segment containing one second of the pulse (a total of 250 data points) is extracted from each wave trough. This pulse is averaged and used to represent the PPG signal over the entire minute for use in deep learning neural networks.
用於本研究之特徵可分類成至少3個類別，即：個人生理特徵、脈搏型態特徵及心率變化。後兩個類別亦可稱為提取特徵，因為此等係可自該PPG信號提取的特徵。在一具體實施例中，該個人生理特徵包括：年紀、腰圍、身體質量指數、收縮壓及舒張壓。在一具體實施例中，取自平均PPG脈搏之脈搏型態特徵包含：50%高度之脈搏寬度、該分鐘之總脈搏面積、平均脈搏面積、該脈搏面積之中位數，及從脈搏波谷至波峰之時間差。在一具體實施例中，提取自信號窗之該心率變化相關特徵可包含來自FFT之低及高頻功率兩者、來自FFT之總功率、脈搏連續間隔變化超過20ms之百分比、連續間隔變化之標準差。以上列出之總共17個特徵（對於包括HbA1c之模型再加上HbA1c）聚結成一特徵向量F以用於饋送至各模型中。 模型架構 The characteristics used in this study can be classified into at least 3 categories, namely: personal physiological characteristics, pulse pattern characteristics and heart rate changes. The latter two categories may also be called extracted features, since these are features that can be extracted from the PPG signal. In a specific embodiment, the personal physiological characteristics include: age, waist circumference, body mass index, systolic blood pressure and diastolic blood pressure. In a specific embodiment, the pulse pattern features taken from the average PPG pulse include: pulse width at 50% height, total pulse area of the minute, average pulse area, median of the pulse area, and pulse trough to The time difference between wave crests. In a specific embodiment, the heart rate change related features extracted from the signal window may include both low and high frequency power from the FFT, the total power from the FFT, the percentage of continuous pulse interval changes exceeding 20 ms, and the standard of continuous interval changes. Difference. The total of 17 features listed above (plus HbA1c for models including HbA1c) are coalesced into a feature vector F for feeding into each model. Model architecture
受到Google之LeNet Inception結構啟發[27]，用於本研究之圖3C模型之模型架構包含具有不同核尺寸（過濾器長度）之兩個平行的訓練模塊（微訓練及巨訓練模塊），然後是如圖3C所示之合併模塊。本設計之想法是希望收集巨觀及微觀兩種輸入資料。在各訓練模塊中，取得相同之經平均之一秒長區段並獨立地訓練。訓練模塊均由三個連續之一維CNN（1dCNN）層組成，每一者接著進行批次正規化及最大池化[28]。接著將1dCNN層之結果攤平並饋送至在各模塊之3個全連結層內。然後在通過最終合併模塊之前將兩訓練模塊之輸出與特徵（特徵向量F 340）合併，該最終合併模塊由三層全連結神經網絡組成。整流線性單元（Rectified Linear Units，ReLU）在整個模型中用作啟動功能，但在最終輸出層使用線性功能。該模型架構之示意圖係如圖3C所示。Inspired by Google's LeNet Inception structure [27], the model architecture of the Figure 3C model used in this study includes two parallel training modules (micro training and macro training modules) with different kernel sizes (filter lengths), and then The merge module is shown in Figure 3C. The idea behind this design is to collect both macroscopic and microscopic input data. In each training module, the same averaged one second long segments are taken and trained independently. The training modules each consist of three consecutive one-dimensional CNN (1dCNN) layers, each followed by batch regularization and max pooling [28]. The results of the 1dCNN layer are then flattened and fed into three fully connected layers in each module. The outputs of the two training modules are then combined with the features (feature vector F 340) before passing through the final merging module, which is composed of a three-layer fully connected neural network. Rectified Linear Units (ReLU) are used as starting functions throughout the model, but linear functions are used in the final output layer. A schematic diagram of the model architecture is shown in Figure 3C.
用於本研究之第二模型係如圖3B所示，其係圖3C之簡化版。特定而言，圖3B之模型不像圖3C之模型包含第二平行CNN模組，以證明當排除有接受醫療之受試者及訓練定群及/或將該受試者及訓練定群的HbA1c資料包括在輸入中時，在無該平行CNN模組結構下亦可達成高精確性。The second model used in this study is shown in Figure 3B, which is a simplified version of Figure 3C. Specifically, the model in Figure 3B does not include a second parallel CNN module like the model in Figure 3C to demonstrate that when subjects who have received medical treatment are excluded from the training cohort and/or the subjects are excluded from the training cohort When HbA1c data is included in the input, high accuracy can be achieved without this parallel CNN module structure.
用於本研究之第三模型係如圖3D所示，其係全連結神經網絡以證明當排除有接受醫療之受試者及訓練定群及/或將該受試者及訓練定群的HbA1c資料包括在輸入中時，就精確性而言，神經網絡模型之多樣性可係有益的。 結果 The third model used in this study is shown in Figure 3D, which is a fully connected neural network to demonstrate HbA1c when excluding subjects who have received medical treatment and training cohorts and/or including such subjects and training cohorts. Diversity in neural network models can be beneficial in terms of accuracy when data is included in the input. result
建立未藉由分開有接受醫療及未接受醫療之標準將受試者及訓練定群，也未包括HbA1c的通用模型以作為檢驗醫療及在其它模型中包括HbA1c之效應的基準線。在本研究中，以下指標係用於評估模型的性能：均方根誤差（RMSE）、平均絕對誤差（MAE）、平均絕對百分比誤差（MAPE）、決定係數（R^2）、10%變化內比例（±10%）及克拉克誤差網格（CEG）A區至E區[29]。對於各模型，總共進行10個獨立訓練及測試，其中隨機分開資料成9:1之訓練及測試組以適當地檢驗學習一致性亦避免過度擬合。關於圖3C模型，各模型之10個獨立訓練的平均性能總結在表2中。圖3B模型之平均性能總結在表3中，及圖3D模型之平均性能總結在表4中。圖3C模型之測試BG預測性能在圖8A-F之CEG圖中說明（包含自10個訓練之所有測試的累計）並總結在表2中，其提供快速且直覺的方式解譯所估計之BG值。圖3B模型之測試BG預測性能在圖10-13之CEG圖中說明並總結在表3，及圖3D模型之測試BG預測性能在圖14-15之CEG圖中說明並總結在表4中。在CEG圖中，所有估計結果基於該估計對臨床決策的影響程度分類成5個區（A至Z），其中A區視為精確， B區視為臨床上可接受（不造成負面影響），及C區、D區及E區由於明顯的預測臨床誤差而視為危險。Establishing a universal model that did not stratify subjects and trainees by separating those who received medical treatment from those who did not, nor did it include HbA1c as a baseline for testing the effects of medical treatment and including HbA1c in other models. In this study, the following indicators are used to evaluate the performance of the model: root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), coefficient of determination (R^2), within 10% change scale (±10%) and Clark Error Grid (CEG) area A to E [29]. For each model, a total of 10 independent training and testing were performed, in which the data were randomly divided into training and testing groups of 9:1 to properly test learning consistency and avoid overfitting. Regarding the Figure 3C model, the average performance of 10 independent trainings of each model is summarized in Table 2. The average performance of the Figure 3B model is summarized in Table 3, and the average performance of the Figure 3D model is summarized in Table 4. The test BG prediction performance of the Figure 3C model is illustrated in the CEG plots of Figure 8A-F (containing the accumulation of all tests from 10 trainings) and summarized in Table 2, which provides a quick and intuitive way to interpret the estimated BG value. The test BG prediction performance of the Figure 3B model is illustrated in the CEG plots of Figures 10-13 and summarized in Table 3, and the test BG prediction performance of the Figure 3D model is illustrated in the CEG plots of Figures 14-15 and summarized in Table 4. In the CEG diagram, all estimation results are classified into 5 areas (A to Z) based on the impact of the estimate on clinical decision-making, where area A is considered accurate, area B is considered clinically acceptable (without negative impact), and areas C, D, and E are considered hazardous due to significant predicted clinical errors.
表2 圖3C模型之模型性能總結
表3圖3B模型之模型性能總結。
表4 圖3D模型之模型性能總結。
藉由基於受試者及訓練定群是否接受任何醫療分開受試者及訓練定群，我們可看到對於未接受醫療之受試者及訓練定群，所有圖3C、圖3B及圖3D模型之性能增加。相反地，對於有接受醫療之受試者及訓練定群，圖3C、圖3B及圖3D模型之性能變得更糟。關於圖3C模型，與該基線比較，有接受醫療但不包括HbA1c之受試者的定群在CEG之A區中具有少了大約7%的預測量；而對於未接受醫療亦不包括HbA1c之受試者之定群，明顯改善大約26%。對於兩種定群，類似之情形在所有其它性能指標均可見到。有人可能會質疑有接受醫療之定群之模型的不佳性能可能是資料不平衡的結果，其偏向較高的BG值。然而，由於有接受醫療之定群的MAPE比未接受醫療之定群大約高12%，這不是這種情況，MAPE係正規化BG值之預測偏差的量值。此證據指出該通用模型之性能因有接受醫療之受試者降低，其證實我們的懷疑，醫療確實對模型性能造成不良影響。類似的趨勢亦可見於圖3B及圖3D之模型。因此，由於就BG預測精確性而言，基於該受試者及訓練定群是否接受醫療來排除受試者及訓練定群實質上有利於神經網絡模型之多樣性，此種排除應有助於所有神經網絡預測BG。By separating subjects and training cohorts based on whether they received any medical treatment, we can see that for subjects and training cohorts who did not receive medical treatment, all Figure 3C, Figure 3B, and Figure 3D models The performance is increased. In contrast, the performance of the models in Figure 3C, Figure 3B, and Figure 3D becomes worse for subjects receiving medical treatment and for the training cohort. Regarding the model in Figure 3C, compared to the baseline, the cohort of subjects who received medical treatment but did not include HbA1c had approximately 7% less predictive power in Region A of the CEG; and for those who did not receive medical treatment and did not include HbA1c The group of subjects improved significantly by about 26%. Similar situations can be seen for all other performance indicators for both fixed groups. One might question that the poor performance of the model in the medically treated cohort may be the result of data imbalance, which is biased toward higher BG values. However, since the MAPE for the cohort that received medical care was approximately 12% higher than that for the cohort that did not receive medical care, this is not the case and MAPE is a measure of the prediction bias of normalized BG values. This evidence indicates that the performance of the general model is reduced by subjects receiving medical treatment, confirming our suspicion that medical treatment does adversely affect model performance. Similar trends can also be seen in the models of Figure 3B and Figure 3D. Therefore, since excluding subjects and training cohorts based on whether they receive medical treatment essentially benefits neural network model diversity in terms of BG prediction accuracy, such exclusions should contribute to All neural networks predict BG.
此外，將HbA1c包括在內，對於所有圖3C、圖3B及圖3D模型，各例均在其整體性能得到明顯改善。對於圖3C模型，包含HbA1c之模型對於有接受醫療之定群之預測精確性在CEG之A區中之預測從53.32%改善至72.25%。有趣的是，儘管藉由增加HbA1c改善，在CEG之A區中該基準線通用模型僅得到16%的改善，至多至大約77%的預測；而有接受醫療之定群僅勉強優於不含HbA1c之該基準線通用模型。兩者仍皆非實際應用所希望的。另一方面，含HbA1之模型對未接受醫療之定群的性能在CEG之A區中達到94%的預測，且無預測落在測試組的C區、D區及E區內。Furthermore, including HbA1c, the overall performance was significantly improved in each case for all Figure 3C, Figure 3B and Figure 3D models. For the Figure 3C model, the prediction accuracy of the model including HbA1c for the group receiving medical treatment improved from 53.32% to 72.25% in Region A of the CEG. Interestingly, despite improvement by adding HbA1c, the baseline general model improved only 16% in CEG region A, to about 77% prediction; while the cohort receiving medical care was only marginally better than the cohort without. This baseline general model of HbA1c. Both are still undesirable for practical applications. On the other hand, the performance of the HbA1-containing model for the untreated cohort reached 94% prediction in area A of the CEG, and no predictions fell in areas C, D, and E of the test group.
將HbA1c包括在輸入中，使用圖3B及3D之簡化模型亦可看到精確性改善，如表3及4中所示。因此，就BG預測精確性而言，由於納入HbA1c作為輸入之部份有益於神經網絡模型的多樣性，納入HbA1c應有益於所有神經網絡預測BG。Accuracy improvements were also seen using the simplified models of Figures 3B and 3D by including HbA1c in the input, as shown in Tables 3 and 4. Therefore, in terms of BG prediction accuracy, since including HbA1c as an input partially benefits the diversity of neural network models, including HbA1c should benefit all neural networks in predicting BG.
表 5.圖3C之各模型的平均訓練及測試損失
在表5中，我們針對基於圖3C之模型的每個情況列出10個訓練之平均訓練及測試損失。從訓練及測試損失之間的差異，我們可定量及客觀地觀察在訓練及測試之間的差異。在一理想情況下，吾人想最小化在訓練期間的損失，並使該測試損失盡可能接近訓練損失。這將確保該模型實際上學習到資料之模式且導致對該測試組的精確預測。我們模型的學習曲線呈現在圖9A-F中。In Table 5, we list the average training and test losses over 10 training sessions for each case based on the model in Figure 3C. From the difference between training and testing losses, we can quantitatively and objectively observe the differences between training and testing. In an ideal situation, we want to minimize the loss during training and make the test loss as close as possible to the training loss. This will ensure that the model actually learns the patterns in the data and results in accurate predictions for the test group. The learning curves of our model are presented in Figures 9A-F.
對於包括所有受試者之通用模型，訓練及測試之損失皆很大。如圖9A所示，該學習曲線在75時期（epoch）後攤平，是所有情況中最短的。這指示與其它情況相較，在基本事實與該資料模式之間存在衝突資訊，從而在學習期間防止進一步收斂。For a general model that includes all subjects, both training and testing losses are large. As shown in Figure 9A, this learning curve flattens out after 75 epochs, which is the shortest of all cases. This indicates that there is conflicting information between the ground truth and the pattern of the data compared to other cases, preventing further convergence during learning.
自總結在表5中之差異，有接受醫療之受試者的定群在其訓練及測試損失之間有顯著較大的差異。從學習曲線（圖9B及圖9C），儘管該損失在訓練期間收斂地較好，但所得模型無法重現良好的預測。此外，該測試損失之曲線在125個時期之後攤平而未彎曲，這指出無過度擬合。因此，很明顯地該模型無法適當地學習以精確地估計。對於有接受醫療之定群，缺少一些主要資訊以進行進一步深度分析。From the differences summarized in Table 5, the cohorts of subjects who received medical treatment had significantly larger differences between their training and test losses. From the learning curve (Figure 9B and Figure 9C), although this loss converged well during training, the resulting model was unable to reproduce good predictions. Furthermore, the test loss curve flattens out after 125 epochs without bending, indicating no overfitting. Therefore, it is obvious that the model cannot learn properly to estimate accurately. For the group receiving medical treatment, some key information is missing for further in-depth analysis.
相反地，未接受醫療之受試者之定群在在其訓練及測試損失之間有極微小的差異，尤其是包括HbA1c者。這進一步指出改善預測性能的關鍵仰賴於減輕醫療的影響。 討論 In contrast, the cohort of untreated subjects showed minimal differences between their training and test losses, particularly those involving HbA1c. This further points out that the key to improving predictive performance lies in mitigating the impact of medical treatment. Discuss
自該CNN架構提取的PPG型態特徵有助於說明該等樣本間之動態葡萄糖變化。然而該PPG波形變化亦可受到許多因素影響，諸如該手指溫度、探針接點、血壓等等。改善NIBG模型的一個主要步驟是為該模型找出一個基礎來區分葡萄糖相關特徵之靜態部份。在此任務中，我們考慮兩個因素：醫療及HbA1c值。The PPG pattern features extracted from the CNN architecture help explain dynamic glucose changes among the samples. However, the PPG waveform change may also be affected by many factors, such as finger temperature, probe contacts, blood pressure, etc. A major step in improving the NIBG model is to find a basis for the model to distinguish the static portion of the glucose-related features. In this task, we consider two factors: medical and HbA1c values.
該結果顯示排除經藥物治療之樣本並包括HbA1c作為特徵改善了預測性能。接受治療之受試者之較差的結果暗示當涉及醫療時，問題變得更加複雜。由於控制新陳代謝穩定性之人體葡萄糖恆定性，外部藥物治療可改變身體生理病理學，接著造成混雜的生物信號出現，使得生理狀態不可預測。更甚者，各種醫療及不同劑量之藥物一起造成複合的效應。在此，我們提出兩個可能之解決方案以減輕在受試者之間不同醫療的效應。1）適當地闡明治療對我們的研究目標的影響，該PPG信號表示成血糖值預測，及將其等併入模型建構中。本研究需要各受試者之醫療途徑更全面的資料，以及此等治療之生物影響的深度知識。2）針對各受試者設計個人化模型。個人化模型基本上避開了所有個人間的差異，無論是否接受醫療，且更特定地針對個人的生理演化。本研究要求長期收集個人資料，但可為未來投資的一種可行方式。The results show that excluding drug-treated samples and including HbA1c as a feature improves predictive performance. The poorer outcomes for those who received the treatment hint that the problem becomes more complicated when medical treatment is involved. Due to the body's glucose constancy, which controls metabolic stability, external drug treatments can alter the body's physiopathology, resulting in the emergence of mixed biological signals that make physiological states unpredictable. What's more, various treatments and different doses of drugs work together to create compound effects. Here, we propose two possible solutions to mitigate the effects of different treatments among subjects. 1) Properly elucidate the impact of treatment on our research goals, express the PPG signal into blood glucose predictions, and incorporate them into model construction. This study requires more comprehensive information on each subject's medical pathways, as well as in-depth knowledge of the biological effects of these treatments. 2) Design a personalized model for each subject. The personalized model essentially eschews all inter-individual differences, whether receiving medical treatment or not, and is more specific to an individual's physiological evolution. This research requires long-term collection of personal data, but is a viable way to invest in the future.
HbA1c在改善我們的模型方面扮演一重要角色。HbA1c已被視為易測量的，且係表示長期身體葡萄糖平衡之一重要的恆定性指標，HbA1c亦反映了一部份的葡萄糖持久存在血流中，其對近紅外光吸收可具有一定的影響。事實上，人們僅需要每三個月測量其HbA1c一次，這對於需要定期血糖值監測的人而言是顯著的進步。因此，將HbA1c用在我們的NIBG建模是可行的。HbA1c plays an important role in improving our model. HbA1c has been regarded as easy to measure and is an important constant indicator of long-term body glucose balance. HbA1c also reflects the persistence of a portion of glucose in the blood stream, which may have a certain impact on near-infrared light absorption. . The fact that people only need to measure their HbA1c every three months is a significant improvement for people who need regular blood glucose monitoring. Therefore, it is feasible to use HbA1c in our NIBG modeling.
本研究的一項限制是資料不平衡的問題。當將該等受試者分成有接受醫療及未接受醫療之定群時，此問題可變得嚴重，有些高葡萄糖資料錯雜，其可造成明顯的預測離群值。這可能不容易解決，因為其取決於本研究可招募到的受試者。為避免由於不平衡之資料而有偏見地預測，可考量資料重新加權或資料擴增以加強訓練來評估該模型。 結論 One limitation of this study is the problem of unbalanced data. This problem can become serious when dividing these subjects into medically treated and non-medically treated cohorts, with some confounding of the high glucose data, which can create significant predictive outliers. This may not be easily resolved as it depends on the number of subjects that can be recruited into the study. To avoid biased predictions due to imbalanced data, consider data reweighting or data augmentation to enhance training to evaluate the model. Conclusion
藉由分析該模型預測精確性及學習曲線，我們證明用於NIBG預測的單一通用模型可受到不可控制之因素（諸如醫療）困擾。藉由將該等受試者分成定群，含有未接受任何種類醫療之受試者之圖3C的模型的預測精確性比其對應者高大約30%。此外，增加HbA1c使得此模型達到12.4 mg/dl之RMSE，8.9 mg/dl之MAE，0.08之MAPE及94.3%之預測精確性，且無失敗之預測落入CEG圖之錯誤區C、D及E中。類似之趨勢亦見於圖3B及3D模型中。我們相信所提出之具定群資料配置及每季量測之HbA1c的模型極有希望可用在未接受任何醫療人群的臨床使用上。By analyzing the model's prediction accuracy and learning curve, we demonstrate that a single general model for NIBG prediction can be plagued by uncontrollable factors, such as medical care. By dividing the subjects into cohorts, the model of Figure 3C containing subjects who did not receive any kind of medical treatment was approximately 30% more accurate in predictions than its counterpart. In addition, adding HbA1c allowed the model to achieve an RMSE of 12.4 mg/dl, an MAE of 8.9 mg/dl, a MAPE of 0.08, and a prediction accuracy of 94.3%, with no failed predictions falling into error zones C, D, and E of the CEG plot. middle. Similar trends are also seen in Figure 3B and the 3D model. We believe that the proposed model with cohort data configuration and quarterly measurement of HbA1c is very promising for clinical use in people who do not receive any medical treatment.
儘管如此，我們也發現醫療藉由改變一個人所表現之生理信號並造成偏差之預測而對NIBG估計具有明顯的影響。藉由輸入HbA1c，我們儘可能地提高有接受醫療之定群的預測精確性，在CEG圖之A區中大約改善20%。對於進一步改善，我們建議個人化模型之方法來消除個體之間的差異，以獲得臨床上可接受的性能。我們相信通過機器學習將NIBG擴展到所有人群將是一個更佳的解決方案。Nonetheless, we also found that medical treatment has a significant impact on NIBG estimates by altering the physiological signals a person exhibits and biasing predictions. By entering HbA1c, we maximized the predictive accuracy of the cohort receiving medical care, improving by approximately 20% in area A of the CEG plot. For further improvement, we suggest a personalized model approach to eliminate inter-individual differences to obtain clinically acceptable performance. We believe extending NIBG to all populations through machine learning will be a better solution.
儘管本發明已針對特定示例性具體實施例及實例描述，熟習本技術之人士可了解，可對上述該等實例進行變化，而不背離其廣泛的發明概念。因此，應了解本發明不限於所揭示之特定實例，但意欲涵蓋在由隨附申請專利範圍所定義之本發明之精神與範疇內之修改。 參考文獻（各以其全文併入本文中）1. Sarkar, K., et al. Design and Implementation of a Noninvasive Blood Glucose Monitoring Device. in 2018 21st International Conference of Computer and Information Technology (ICCIT). 2018. Dhaka, Bangladesh: IEEE. 2. Mekonnen, B.K., et al., Accurate prediction of glucose concentration and identification of major contributing features from hardly distinguishable near-infrared spectroscopy.Biomedical Signal Processing and Control, 2020. 59: p. 101923. 3. Maier, J.S., et al., Possible correlation between blood glucose concentration and the reduced scattering coefficient of tissues in the near infrared.Optics letters, 1994. 19(24): p. 2062-2064. 4. Tamada, J.A., et al., Noninvasive glucose monitoring: comprehensive clinical results. Cygnus Research Team.Jama, 1999. 282(19): p. 1839-44. 5. Klonoff, D.C., Noninvasive blood glucose monitoring.Diabetes care, 1997. 20(3): p. 433-437. 6. Larin, K.V., et al., Noninvasive blood glucose monitoring with optical coherence tomography: a pilot study in human subjects.Diabetes care, 2002. 25(12): p. 2263-2267. 7. Yadav, J., et al., Prospects and limitations of non-invasive blood glucose monitoring using near-infrared spectroscopy.Biomedical signal processing and control, 2015. 18: p. 214-227. 8. Abd Salam, N.A.B., et al., The evolution of non-invasive blood glucose monitoring system for personal application.Journal of Telecommunication, Electronic and Computer Engineering (JTEC), 2016. 8(1): p. 59-65. 9. Monte-Moreno, E., Non-invasive estimate of blood glucose and blood pressure from a photoplethysmograph by means of machine learning techniques.Artificial Intelligence in Medicine, 2011. 53(2): p. 127-138. 10. Blank, T.B., et al. Clinical results from a noninvasive blood glucose monitor. in Optical Diagnostics and Sensing of Biological Fluids and Glucose and Cholesterol Monitoring II. 2002. International Society for Optics and Photonics. 11. Paul, B., M.P. Manuel, and Z.C. Alex. Design and development of non invasive glucose measurement system. in 2012 1st International Symposium on Physics and Technology of Sensors (ISPTS-1). 2012. 12. Ramasahayam, S., et al. FPGA based system for blood glucose sensing using photoplethysmography and online motion artifact correction using adaline. in 2015 9th International Conference on Sensing Technology (ICST). 2015. 13. Rachim, V.P. and W.-Y. Chung, Wearable-band type visible-near infrared optical biosensor for non-invasive blood glucose monitoring.Sensors and Actuators B: Chemical, 2019. 286: p. 173-180. 14. Maruo, K., et al., New methodology to obtain a calibration model for noninvasive near-infrared blood glucose monitoring.Applied spectroscopy, 2006. 60(4): p. 441-449. 15. Alian, A.A. and K.H. Shelley, Photoplethysmography.Best Pract Res Clin Anaesthesiol, 2014. 28(4): p. 395-406. 16. Jain, P., A.M. Joshi, and S.P. Mohanty, iGLU 1.0: An Accurate Non-Invasive Near-Infrared Dual Short Wavelengths Spectroscopy based Glucometer for Smart Healthcare.arXiv preprint arXiv:1911.04471, 2019. 17. Bunescu, R., et al. Blood glucose level prediction using physiological models and support vector regression. in 2013 12th International Conference on Machine Learning and Applications. 2013. IEEE. 18. Georga, E.I., et al. A predictive model of subcutaneous glucose concentration in type 1 diabetes based on random forests. in 2012 Annual International Conference of the IEEE Engineering in Medicine and Biology Society. 2012. IEEE. 19. Altman, N.S., An Introduction to Kernel and Nearest-Neighbor Nonparametric Regression.The American Statistician, 1992. 46(3): p. 175-185. 20. Tomczak, J.M. Gaussian Process Regression with Categorical Inputs for Predicting the Blood Glucose Level. in Advances in Systems Science. 2017. Cham: Springer International Publishing. 21. Yadav, J., et al., Investigations on multisensor-based noninvasive blood glucose measurement system.Journal of Medical Devices, 2017. 11(3). 22. Paneni, F., et al., Diabetes and vascular disease: pathophysiology, clinical consequences, and medical therapy: part I.Eur Heart J, 2013. 34(31): p. 2436-43. 23. Benichou, T., et al., Heart rate variability in type 2 diabetes mellitus: A systematic review and meta-analysis.PLoS One, 2018. 13(4): p. e0195166. 24. Use of Glycated Haemoglobin (HbA1c) in the Diagnosis of Diabetes Mellitus: Abbreviated Report of a WHO Consultation, in WHO Guidelines Approved by the Guidelines Review Committee. 2011, World Health Organization: Geneva. 25. Chu, J., et al., One-Minute Finger Pulsation Measurement for Diabetes Rapid Screening with 1.3% to 13% False-Negative Prediction Rate.Biomedical Statistics and Informatics, 2021. 6: p. 8. 26. Navakatikyan, M.A., et al., A real-time algorithm for the quantification of blood pressure waveforms.IEEE Trans Biomed Eng, 2002. 49(7): p. 662-70. 27. Szegedy, C., et al., Going deeper with convolutions, in 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2015. p. 1-9. 28. Kiranyaz, S., et al., 1D convolutional neural networks and applications: A survey.Mechanical Systems and Signal Processing, 2021. 151. 29. Clarke, W.L., et al., Evaluating clinical accuracy of systems for self-monitoring of blood glucose.1987. 10(5): p. 622-628. Although the present invention has been described with respect to specific illustrative embodiments and examples, those skilled in the art will appreciate that changes may be made in these examples without departing from the broad inventive concept thereof. It is to be understood, therefore, that this invention is not limited to the specific examples disclosed, but it is intended to cover modifications within the spirit and scope of the invention as defined by the appended claims. References (each incorporated herein by its full text) 1. Sarkar, K., et al. Design and Implementation of a Noninvasive Blood Glucose Monitoring Device . in 2018 21st International Conference of Computer and Information Technology (ICCIT) . 2018. Dhaka , Bangladesh: IEEE. 2. Mekonnen, BK, et al., Accurate prediction of glucose concentration and identification of major contributing features from hardly distinguishable near-infrared spectroscopy. Biomedical Signal Processing and Control, 2020. 59 : p. 101923. 3. Maier, JS, et al., Possible correlation between blood glucose concentration and the reduced scattering coefficient of tissues in the near infrared. Optics letters, 1994. 19 (24): p. 2062-2064. 4. Tamada, JA, et al ., Noninvasive glucose monitoring: comprehensive clinical results. Cygnus Research Team. Jama, 1999. 282 (19): p. 1839-44. 5. Klonoff, DC, Noninvasive blood glucose monitoring. Diabetes care, 1997. 20 (3): p. 433-437. 6. Larin, KV, et al., Noninvasive blood glucose monitoring with optical coherence tomography: a pilot study in human subjects. Diabetes care, 2002. 25 (12): p. 2263-2267. 7. Yadav, J., et al., Prospects and limitations of non-invasive blood glucose monitoring using near-infrared spectroscopy. Biomedical signal processing and control, 2015. 18 : p. 214-227. 8. Abd Salam, NAB, et al ., The evolution of non-invasive blood glucose monitoring system for personal application. Journal of Telecommunication, Electronic and Computer Engineering (JTEC), 2016. 8 (1): p. 59-65. 9. Monte-Moreno, E., Non-invasive estimate of blood glucose and blood pressure from a photoplethysmograph by means of machine learning techniques. Artificial Intelligence in Medicine, 2011. 53 (2): p. 127-138. 10. Blank, TB, et al. Clinical results from a noninvasive blood glucose monitor . in Optical Diagnostics and Sensing of Biological Fluids and Glucose and Cholesterol Monitoring II . 2002. International Society for Optics and Photonics. 11. Paul, B., MP Manuel, and ZC Alex. Design and development of non invasive glucose measurement system . in 2012 1st International Symposium on Physics and Technology of Sensors (ISPTS-1) . 2012. 12. Ramasahayam, S., et al. FPGA based system for blood glucose sensing using photoplethysmography and online motion artifact correction using adaline . in 2015 9th International Conference on Sensing Technology (ICST) . 2015. 13. Rachim, VP and W.-Y. Chung, Wearable-band type visible-near infrared optical biosensor for non-invasive blood glucose monitoring. Sensors and Actuators B: Chemical, 2019. 286 : p. 173-180. 14. Maruo, K., et al., New methodology to obtain a calibration model for noninvasive near-infrared blood glucose monitoring. Applied spectroscopy, 2006. 60 (4): p . 441-449. 15. Alian, AA and KH Shelley, Photoplethysmography. Best Pract Res Clin Anaesthesiol, 2014. 28 (4): p. 395-406. 16. Jain, P., AM Joshi, and SP Mohanty, iGLU 1.0: An Accurate Non-Invasive Near-Infrared Dual Short Wavelengths Spectroscopy based Glucometer for Smart Healthcare. arXiv preprint arXiv:1911.04471, 2019. 17. Bunescu, R., et al. Blood glucose level prediction using physiological models and support vector regression . in 2013 12th International Conference on Machine Learning and Applications . 2013. IEEE. 18. Georga, EI, et al. A predictive model of subcutaneous glucose concentration in type 1 diabetes based on random forests . in 2012 Annual International Conference of the IEEE Engineering in Medicine and Biology Society . 2012. IEEE. 19. Altman, NS, An Introduction to Kernel and Nearest-Neighbor Nonparametric Regression. The American Statistician, 1992. 46 (3): p. 175-185. 20. Tomczak, JM Gaussian Process Regression with Categorical Inputs for Predicting the Blood Glucose Level . in Advances in Systems Science . 2017. Cham: Springer International Publishing. 21. Yadav, J., et al., Investigations on multisensor-based noninvasive blood glucose measurement system. Journal of Medical Devices, 2017. 11 (3). 22. Paneni, F., et al., Diabetes and vascular disease: pathophysiology, clinical consequences, and medical therapy: part I. Eur Heart J, 2013. 34 (31): p. 2436-43. 23. Benichou, T., et al., Heart rate variability in type 2 diabetes mellitus: A systematic review and meta-analysis. PLoS One, 2018. 13 (4): p. e0195166. 24. Use of Glycated Haemoglobin (HbA1c) in the Diagnosis of Diabetes Mellitus: Abbreviated Report of a WHO Consultation , in WHO Guidelines Approved by the Guidelines Review Committee . 2011, World Health Organization: Geneva. 25. Chu, J., et al., One- Minute Finger Pulsation Measurement for Diabetes Rapid Screening with 1.3% to 13% False-Negative Prediction Rate. Biomedical Statistics and Informatics, 2021. 6 : p. 8. 26. Navakatikyan, MA, et al., A real-time algorithm for the quantification of blood pressure waveforms. IEEE Trans Biomed Eng, 2002. 49 (7): p. 662-70. 27. Szegedy, C., et al., Going deeper with convolutions , in 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) . 2015. p. 1-9. 28. Kiranyaz, S., et al., 1D convolutional neural networks and applications: A survey. Mechanical Systems and Signal Processing, 2021. 151. 29. Clarke, WL, et al. al., Evaluating clinical accuracy of systems for self-monitoring of blood glucose. 1987. 10 (5): p. 622-628.
100:受試者 110:PPG信號裝置 112:PPG信號讀取器 114:信號發射器 118:信號模組 120:連接器 200:處理器 220:類比至數位轉換器（A/D轉換器） 222:信號處理器 224:特徵萃取器 230:神經網絡 240:顯示器 250:記憶體 251:生理特徵 252:數位化信號 254:信號處理結果 256:擷取特徵 259:計算結果 301:卷積神經網絡 302:卷積神經網絡 303:卷積神經網絡 306:密集層 310:卷積模組 312:卷積層 312a:卷積層 312b:卷積層 312c:卷積層 314:攤平模組 316:密集層 316a:密集層 320:微卷積模組 321:微過濾器 322:微卷積模組 322a:微卷積層 322b:微卷積層 322c:微卷積層 324:攤平模組 326:密集層 330:巨卷積模組 331:巨過濾器 332:巨卷積層 332a:巨卷積層 332b:巨卷積層 332c:巨卷積層 334:卷積子模組 /攤平模組 336:批次正規化模組 /密集層 338:啟動功能模組 339:池化模組 340:特徵向量 344:密集啟動模組 346:漏失模組 348:批次正規化模組 350:合併模塊模組 352:合併層 354:密集層 100: Subject 110:PPG signaling device 112:PPG signal reader 114:Signal transmitter 118:Signal module 120: Connector 200:processor 220: Analog to digital converter (A/D converter) 222:Signal processor 224:Feature Extractor 230:Neural Network 240:Display 250:Memory 251:Physiological characteristics 252: Digitized signal 254:Signal processing results 256: Extract features 259: Calculation results 301: Convolutional Neural Network 302: Convolutional Neural Network 303: Convolutional Neural Network 306: Dense layer 310:Convolution module 312:Convolution layer 312a: Convolutional layer 312b: Convolutional layer 312c: Convolutional layer 314: Flatten module 316: Dense layer 316a: dense layer 320:Microconvolution module 321: Micro filter 322:Microconvolution module 322a:Microconvolutional layer 322b:Microconvolutional layer 322c: Microconvolutional layer 324: Flatten module 326: Dense layer 330: Giant convolution module 331: Giant filter 332: Giant convolutional layer 332a: Giant convolutional layer 332b: Giant convolutional layer 332c: Giant convolutional layer 334: Convolution submodule/flattening module 336: Batch regularization module/dense layer 338: Start function module 339: Pooling module 340: Feature vector 344: Intensive startup module 346:Leakage module 348:Batch normalization module 350: Merge module module 352:Merge layers 354: Dense layer
圖1描繪本發明之基於PPG之NIBG神經網絡預測系統之具體實施例的總體概述。 圖2詳細描繪本發明之基於PPG之NIBG神經網絡預測系統之處理器200的一具體實施例。 圖3A-D描繪本發明之基於PPG之NIBG預測系統之神經網絡（NN）的四個具體實施例。圖3A描繪本發明之基於PPG之NIBG預測系統之NN 230的一般結構。圖3B描繪本發明之基於PPG之NIBG預測系統之卷積神經網絡（CNN）302的一具體實施例，其包含一單一CNN模組。圖3C描繪本發明之基於PPG之NIBG預測系統之CNN 303的一具體實施例，其包含二個平行的CNN模組。圖3D描繪本發明之全連結神經網絡（FCNN）301的一具體實施例。 圖4A-D描繪本發明之各種PPG信號。圖4A描繪本發明之一示例性PPG信號。圖4B描繪本發明之一示例性PPG信號，其藉由一低通過濾器過濾。圖4C描繪本發明之一示例性PPG信號，其藉由一高通過濾器過濾。圖4D描繪一示例性信號窗，其源自藉由低通及高通過濾器促進之本發明之PPG信號。 圖5A描繪本發明之基於PPG之NIBG預測系統及方法的視覺表示。圖5B說明有接受醫療之受試者之HbA1c及血糖值的相關性。圖5C說明未接受任何醫療之受試者之HbA1c及血糖值的相關性。 圖6說明本發明之基於PPG之NIBG神經網絡預測方法之訓練階段之流程的一具體實施例。 圖7說明本發明之基於PPG之NIBG神經網絡預測方法之預測階段之流程的一具體實施例。 圖8A-D描繪本發明之圖3C之模型之各種具體實施例的CEG分析。圖8A描繪圖3C之模型與所有受試者及訓練定群（有接受醫療以及未接受醫療）之CEG分析，且包括PPG信號及該等16種特徵，但未包括HbA1c作為輸入；圖8B描繪圖3C之模型與所有受試者及訓練定群（有接受醫療以及未接受醫療）之CEG分析，且包括PPG信號、16個特徵及HbA1c作為輸入；圖8C描繪圖3C之模型與受試者之CEG分析，且以有接受醫療之訓練定群訓練，及包括PPG信號及16個特徵，但未包括HbA1c作為輸入；圖8D描繪圖3C之模型與受試者之CEG分析，且以有接受醫療之訓練定群訓練，及包括PPG信號、16個特徵及HbA1c作為該輸入之部份；圖8E描繪圖3C之模型與受試者之CEG分析，且以未接受醫療之訓練定群訓練，及包括PPG信號及特徵，但不包括HbA1c作為輸入；及圖8F描繪圖3C之模型與受試者之CEG分析，且以有接受醫療之訓練定群訓練，及包括PPG信號、16個特徵及HbA1c作為輸入。 圖9A-F描繪本發明之圖3C之模型之各種具體實施例的學習曲線。圖9A描繪本發明之圖3C之模型與所有受試者（有接受醫療以及未接受醫療）的學習曲線，且包括PPG信號及16個特徵，但不包括HbA1c作為輸入，圖9B描繪本發明之圖3C之模型與所有受試者（有接受醫療以及未接受醫療）的學習曲線，且包括PPG信號、16個特徵及HbA1c作為輸入，圖9C描繪本發明之圖3C之模型之學習曲線，其僅包括有接受醫療之受試者，及包括PPG信號及16個特徵，但不包括HbA1c作為輸入， 圖9D描繪本發明之圖3C之模型的學習曲線，其僅包括有接受醫療之受試者，及包括PPG信號、16個特徵及HbA1c作為輸入，圖9E描繪本發明之圖3C之模型的學習曲線，其僅包括未接受任何醫療之受試者，及包括PPG信號及16個特徵，但未包括HbA1c作為輸入，圖9F描繪本發明之圖3C之模型的學習曲線，其僅包括未接受醫療之受試者，及包括PPG信號、16個特徵及HbA1c作為輸入。在各情況中，呈現該等訓練中之最佳及最差訓練。 圖10A及10B描繪本發明之圖3B之模型之CEG分析，其僅包括PPG信號且不包括HbA1c或16個特徵作為該輸入之部份，且受試者及訓練定群未接受任何醫療。圖10A使用來自未接受任何醫療之訓練定群之PPG信號作為輸入描繪圖3B之模型之CEG分析；圖10B描繪圖3B之模型之CEG分析，其僅使用來自未接受任何醫療之受試者的PPG信號作為輸入，及模型使用來自未接受任何醫療之訓練定群之PPG信號訓練。 圖10C及10D描繪本發明之圖3B之模型之CEG分析，其僅包括PPG信號且不包括HbA1c或16個特徵作為該輸入之部份，且受試者有接受醫療，並以有接受醫療之訓練定群訓練。圖10C描繪圖3B之模型之CEG分析，其僅使用來自有接受醫療之訓練定群的PPG信號作為輸入；圖10D描繪圖3B之模型之CEG分析，其僅使用來自有接受醫療之受試者的PPG信號作為輸入及模型使用來自有接受醫療之訓練定群之PPG信號訓練。 圖11A及11B描繪本發明之圖3B之模型之CEG分析，其包括PPG信號及16個特徵但不包括HbA1c，且其中受試者及訓練定群未接受任何醫療。圖11A描繪圖3B之模型之CEG分析，其僅使用來自未接受任何醫療之訓練定群之PPG信號及16個特徵作為輸入；圖11B描繪圖3B之模型之CEG分析，其僅使用來自未接受任何醫療之受試者之PPG信號及16個特徵作為輸入，及模型使用來自未接受任何醫療之訓練定群之PPG信號及16個特徵訓練。 圖11C及11D描繪本發明之圖3B之模型之CEG分析，其包括PPG信號及16個特徵但不包括HbA1c，且其中受試者及訓練定群有接受醫療。圖11C描繪圖3B之模型之CEG分析，其僅使用來自有接受醫療之訓練定群之PPG信號及16個特徵作為輸入；圖11D描繪圖3B之模型之CEG分析，其僅使用來自有接受醫療之受試者之PPG信號及16個特徵作為輸入，及模型使用來自有接受醫療之訓練定群之PPG信號及16個特徵訓練。 圖12A及12B描繪本發明之圖3B之模型之CEG分析，其包括PPG信號及HbA1c但不包括16個特徵作為輸入，且其中受試者及訓練定群未接受任何醫療。圖12A描繪圖3B之模型之CEG分析，其僅使用來自未接受任何醫療之訓練定群之PPG信號及HbA1c作為輸入；圖12B描繪圖3B之模型之CEG分析，其僅使用來自未接受任何醫療之受試者之PPG信號及HbA1c作為輸入，及模型使用來自未接受任何醫療之訓練定群之PPG信號及HbA1c訓練。 圖12C及12D描繪本發明之圖3B之模型之CEG分析，其包括PPG信號及HbA1c但不包括16個特徵作為輸入，且其中受試者及訓練定群有接受醫療。圖12C描繪圖3B之模型之CEG分析，其僅使用來自有接受醫療之訓練定群之PPG信號及HbA1c作為輸入；圖12D描繪圖3B之模型之CEG分析，其僅使用來自有接受醫療之受試者之PPG信號及HbA1c作為輸入，及模型使用來自有接受醫療之訓練定群之PPG信號及HbA1c訓練。 圖13A及13B描繪本發明之圖3B之模型之CEG分析，其包括PPG信號、16個特徵及HbA1c作為輸入，且其中受試者及訓練未接受任何醫療之定群。圖13A描繪圖3B之模型之CEG分析，其僅使用來自未接受任何醫療之訓練定群之PPG信號、16個特徵及HbA1c作為輸入；圖13B描繪圖3B之模型之CEG分析，其僅使用來自未接受任何醫療之受試者之PPG信號、16個特徵及HbA1c作為輸入，及模型使用來自未接受任何醫療之訓練定群之PPG信號、16個特徵及HbA1c訓練。 圖13C及13D描繪本發明之圖3B之模型之CEG分析，其包括PPG信號、16個特徵及HbA1c作為輸入，且其中受試者及訓練定群有接受醫療。圖13C描繪圖3B之模型之CEG分析，其僅使用來自有接受醫療之定群之PPG信號、16個特徵及HbA1c作為輸入；圖13D描繪圖3B之模型之CEG分析，其僅使用來自有接受醫療之受試者之PPG信號、16個特徵及HbA1c作為輸入，及模型使用來自有接受醫療之訓練定群之PPG信號、16個特徵及HbA1c訓練。 圖14A及14B描繪本發明之圖3D之模型之CEG分析，其包括PPG信號及16個特徵但不包括HbA1C作為輸入，且其中受試者及訓練定群未接受任何醫療。圖14A描繪圖3D之模型之CEG分析，其僅使用來自未接受任何醫療之訓練定群之PPG信號及HbA1c作為輸入；圖14B描繪圖3D之模型之CEG分析，其僅使用來自未接受任何醫療之受試者之PPG信號及16個特徵作為輸入，及模型使用來自未接受任何醫療之訓練定群之PPG信號及16個特徵訓練。 圖14C及14D描繪本發明之圖3D之模型之CEG分析，其中包括PPG信號及16個特徵但不包括HbA1C作為輸入，及其中受試者及訓練定群有接受醫療。圖14C描繪圖3D之模型之CEG分析，其僅使用來自有接受醫療之訓練定群之PPG信號及16個特徵作為輸入；圖14D描繪圖3D之模型之CEG分析，其僅使用有接受醫療之受試者之PPG信號及16個特徵作為輸入，及模型使用來自有接受醫療之訓練定群之PPG信號及HbA1c訓練。 圖15A及15B描繪本發明之圖3D之模型之CEG分析，其包括PPG信號、16個特徵及HbA1c作為輸入，且其中受試者及訓練定群未接受任何醫療。圖15A描繪圖3B之模型之CEG分析，其僅使用來自有接受醫療之訓練定群之PPG信號、16個特徵及HbA1c作為輸入；圖15B描繪圖3D之模型之CEG分析，其僅使用來自有接受醫療之受試者之PPG信號、16個特徵及HbA1c作為輸入，及模型使用來自未接受醫療之訓練定群之PPG信號、16個特徵及HbA1c訓練。 圖15C及15D描繪本發明之圖3D之模型之CEG分析，其包括PPG信號、16個特徵及HbA1c作為輸入，且其中受試者及訓練定群有接受醫療。圖15C描繪圖3D之模型之CEG分析，其僅使用來自有接受醫療之定群之PPG信號、16個特徵及HbA1c作為輸入；圖15D描繪圖3D之模型之CEG分析，其僅使用來自有接受醫療之受試者之PPG信號、16個特徵及HbA1c作為輸入，及模型使用來自有接受醫療之訓練定群之PPG信號、16個特徵及HbA1c訓練。 Figure 1 depicts a general overview of a specific embodiment of the PPG-based NIBG neural network prediction system of the present invention. FIG. 2 depicts in detail a specific embodiment of the processor 200 of the PPG-based NIBG neural network prediction system of the present invention. 3A-D depict four specific embodiments of the neural network (NN) of the PPG-based NIBG prediction system of the present invention. Figure 3A depicts the general structure of the NN 230 of the PPG-based NIBG prediction system of the present invention. FIG. 3B depicts a specific embodiment of the convolutional neural network (CNN) 302 of the PPG-based NIBG prediction system of the present invention, which includes a single CNN module. Figure 3C depicts a specific embodiment of the CNN 303 of the PPG-based NIBG prediction system of the present invention, which includes two parallel CNN modules. Figure 3D depicts a specific embodiment of the fully connected neural network (FCNN) 301 of the present invention. Figures 4A-D depict various PPG signals of the present invention. Figure 4A depicts an exemplary PPG signal of the present invention. Figure 4B depicts an exemplary PPG signal of the present invention filtered by a low-pass filter. Figure 4C depicts an exemplary PPG signal of the present invention filtered by a high-pass filter. Figure 4D depicts an exemplary signal window derived from the PPG signal of the present invention promoted by low-pass and high-pass filters. Figure 5A depicts a visual representation of the PPG-based NIBG prediction system and method of the present invention. Figure 5B illustrates the correlation between HbA1c and blood glucose values in subjects receiving medical treatment. Figure 5C illustrates the correlation between HbA1c and blood glucose values in subjects who did not receive any medical treatment. Figure 6 illustrates a specific embodiment of the process of the training phase of the PPG-based NIBG neural network prediction method of the present invention. Figure 7 illustrates a specific embodiment of the process of the prediction stage of the PPG-based NIBG neural network prediction method of the present invention. Figures 8A-D depict CEG analysis of various embodiments of the model of Figure 3C of the present invention. Figure 8A depicts the CEG analysis of the model of Figure 3C with all subjects and the training cohort (with and without medical treatment), and includes the PPG signal and these 16 features, but does not include HbA1c as input; Figure 8B depicts Figure 3C shows the model and CEG analysis of all subjects and training cohorts (with and without medical treatment), and includes PPG signal, 16 features and HbA1c as input; Figure 8C depicts the model of Figure 3C and subjects CEG analysis of the model and subjects with medical training, and included PPG signal and 16 features, but did not include HbA1c as input; Figure 8D depicts CEG analysis of the model and subjects in Figure 3C, and with medical acceptance Medical training cohort training, and includes PPG signal, 16 features and HbA1c as part of the input; Figure 8E depicts the CEG analysis of the model and subjects in Figure 3C, and is based on the training cohort training without medical treatment, and includes the PPG signal and features, but does not include HbA1c as input; and Figure 8F depicts the CEG analysis of the model and subjects of Figure 3C, and is trained on a medically trained cohort, and includes the PPG signal, 16 features, and HbA1c as input. Figures 9A-F depict learning curves for various embodiments of the model of Figure 3C of the present invention. Figure 9A depicts the learning curve of the model of Figure 3C of the present invention and all subjects (with and without medical treatment), and includes the PPG signal and 16 features, but does not include HbA1c as input. Figure 9B depicts the learning curve of the present invention. The learning curve of the model of Figure 3C and all subjects (with and without medical treatment), and including PPG signal, 16 features and HbA1c as inputs, Figure 9C depicts the learning curve of the model of Figure 3C of the present invention, which Including only subjects who received medical treatment, and including the PPG signal and 16 features, but excluding HbA1c as input, Figure 9D depicts the learning curve of the model of Figure 3C of the present invention, which only includes subjects who received medical treatment. , and includes the PPG signal, 16 features and HbA1c as inputs, Figure 9E depicts the learning curve of the model of Figure 3C of the present invention, which only includes subjects who did not receive any medical treatment, and includes the PPG signal and 16 features, but Without including HbA1c as input, Figure 9F depicts the learning curve of the model of Figure 3C of the present invention, which only includes subjects who did not receive medical treatment, and includes the PPG signal, 16 features and HbA1c as input. In each case, the best and worst of these drills are presented. Figures 10A and 10B depict CEG analysis of the model of Figure 3B of the present invention, which includes only the PPG signal and does not include HbA1c or 16 features as part of the input, and the subjects and training cohort did not receive any medical treatment. Figure 10A depicts a CEG analysis of the model of Figure 3B using PPG signals from a training cohort that did not receive any medical treatment as input; Figure 10B depicts a CEG analysis of the model of Figure 3B using only PPG signals from subjects who did not receive any medical treatment. PPG signals are taken as input, and the model is trained using PPG signals from a training cohort that did not receive any medical treatment. Figures 10C and 10D depict a CEG analysis of the model of Figure 3B of the present invention, which includes only the PPG signal and does not include HbA1c or 16 features as part of the input, and the subject has received medical treatment, and has received medical treatment. Training group training. Figure 10C depicts a CEG analysis of the model of Figure 3B, which uses only PPG signals from the training cohort who received medical treatment as input; Figure 10D depicts a CEG analysis of the model of Figure 3B, which only uses PPG signals from subjects who received medical treatment. The PPG signal is used as input and the model is trained using the PPG signal from the medical training cohort. Figures 11A and 11B depict CEG analysis of the model of Figure 3B of the present invention, which includes PPG signal and 16 features but does not include HbA1c, and in which the subjects and training cohort did not receive any medical treatment. Figure 11A depicts a CEG analysis of the model of Figure 3B, which uses only PPG signals and 16 features from a training cohort that did not receive any medical treatment as input; Figure 11B depicts a CEG analysis of the model of Figure 3B, which uses only PPG signals from a training cohort that did not receive any medical treatment. The PPG signal and 16 features of subjects who received any medical treatment were used as input, and the model was trained using the PPG signal and 16 features from the training cohort that did not receive any medical treatment. Figures 11C and 11D depict CEG analysis of the model of Figure 3B of the present invention, which includes PPG signal and 16 features but does not include HbA1c, and in which the subject and training cohort received medical treatment. Figure 11C depicts the CEG analysis of the model of Figure 3B, which uses only the PPG signal and 16 features from the training cohort with medical treatment as input; Figure 11D depicts the CEG analysis of the model of Figure 3B, which uses only the PPG signal from the training cohort with medical treatment. The subject's PPG signal and 16 features are used as input, and the model is trained using the PPG signal and 16 features from the training cohort who have received medical treatment. Figures 12A and 12B depict CEG analysis of the model of Figure 3B of the present invention, which includes PPG signal and HbA1c but not 16 features as input, and in which the subjects and training cohort did not receive any medical treatment. Figure 12A depicts a CEG analysis of the model of Figure 3B, which uses only PPG signals and HbA1c from a training cohort that did not receive any medical treatment as input; Figure 12B depicts a CEG analysis of the model of Figure 3B, which uses only PPG signals from a training cohort that did not receive any medical treatment. The PPG signal and HbA1c of the subjects were used as input, and the model was trained using the PPG signal and HbA1c from the training cohort that did not receive any medical treatment. Figures 12C and 12D depict CEG analysis of the model of Figure 3B of the present invention, which includes PPG signal and HbA1c but does not include 16 features as input, and in which subjects and training cohorts receive medical treatment. Figure 12C depicts a CEG analysis of the model of Figure 3B, which uses only PPG signals and HbA1c from the training cohort who received medical treatment as input; Figure 12D depicts a CEG analysis of the model of Figure 3B, which uses only those from the training cohort who received medical treatment. The subject's PPG signal and HbA1c were used as input, and the model was trained using the PPG signal and HbA1c from the medically trained cohort. Figures 13A and 13B depict a CEG analysis of the model of Figure 3B of the present invention, which includes the PPG signal, 16 features and HbA1c as input, and in which subjects and training did not receive any medical treatment. Figure 13A depicts a CEG analysis of the model of Figure 3B, which uses only the PPG signal, 16 features and HbA1c from the training cohort that did not receive any medical treatment as input; Figure 13B depicts a CEG analysis of the model of Figure 3B, which uses only the PPG signal from the training cohort that did not receive any medical treatment. The PPG signal, 16 features and HbA1c of subjects who did not receive any medical treatment were used as input, and the model was trained using the PPG signal, 16 features and HbA1c from the training cohort who did not receive any medical treatment. Figures 13C and 13D depict CEG analysis of the model of Figure 3B of the present invention, which includes PPG signal, 16 features and HbA1c as input, and in which subjects and training cohorts receive medical treatment. Figure 13C depicts a CEG analysis of the model of Figure 3B, which uses only PPG signals, 16 features, and HbA1c from a cohort receiving medical treatment as input; Figure 13D depicts a CEG analysis of the model of Figure 3B, which uses only those from a cohort of patients receiving medical care. The PPG signal, 16 features and HbA1c of medical subjects were used as input, and the model was trained using the PPG signal, 16 features and HbA1c from the training cohort who received medical treatment. Figures 14A and 14B depict a CEG analysis of the model of Figure 3D of the present invention, which includes the PPG signal and 16 features but not HbA1C as input, and in which the subjects and training cohort did not receive any medical treatment. Figure 14A depicts a CEG analysis of the model of Figure 3D, which uses only PPG signals and HbA1c from a training cohort that did not receive any medical treatment as input; Figure 14B depicts a CEG analysis of the model of Figure 3D, which uses only PPG signals from a training cohort that did not receive any medical treatment. The subject's PPG signal and 16 features were used as input, and the model was trained using the PPG signal and 16 features from the training cohort that did not receive any medical treatment. Figures 14C and 14D depict a CEG analysis of the model of Figure 3D of the present invention, which includes the PPG signal and 16 features but not HbA1C as input, and in which subjects and training cohorts receive medical treatment. Figure 14C depicts a CEG analysis of the model of Figure 3D, which uses only PPG signals and 16 features from the training cohort with medical treatment as input; Figure 14D depicts a CEG analysis of the model of Figure 3D, which uses only those with medical treatment. The subject's PPG signal and 16 features were used as input, and the model was trained using the PPG signal and HbA1c from the medically trained training cohort. Figures 15A and 15B depict CEG analysis of the model of Figure 3D of the present invention, which includes PPG signal, 16 features and HbA1c as input, and in which the subjects and training cohort did not receive any medical treatment. Figure 15A depicts a CEG analysis of the model of Figure 3B, which uses only PPG signals, 16 features and HbA1c from a training cohort receiving medical treatment as input; Figure 15B depicts a CEG analysis of the model of Figure 3D, which uses only PPG signals from a training cohort with medical treatment. The PPG signal, 16 features, and HbA1c of subjects who received medical treatment were used as input, and the model was trained using the PPG signal, 16 features, and HbA1c from the training cohort that did not receive medical treatment. Figures 15C and 15D depict CEG analysis of the model of Figure 3D of the present invention, which includes PPG signal, 16 features and HbA1c as input, and in which subjects and training cohorts receive medical treatment. Figure 15C depicts a CEG analysis of the model of Figure 3D, which uses only PPG signals, 16 features, and HbA1c from a cohort receiving medical treatment as input; Figure 15D depicts a CEG analysis of the model of Figure 3D, which uses only PPG signals from a cohort of patients receiving medical care. The PPG signal, 16 features and HbA1c of medical subjects were used as input, and the model was trained using the PPG signal, 16 features and HbA1c from the training cohort who received medical treatment.
100:受試者 100: Subject
110:PPG信號裝置 110:PPG signaling device
112:PPG信號讀取器 112:PPG signal reader
114:信號發射器 114:Signal transmitter
118:信號模組 118:Signal module
120:連接器 120:Connector
200:處理器 200:processor
Claims (25)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| TW111128479A TWI823501B (en) | 2022-07-29 | 2022-07-29 | Photoplethysmography based non-invasive blood glucose prediction by neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| TW111128479A TWI823501B (en) | 2022-07-29 | 2022-07-29 | Photoplethysmography based non-invasive blood glucose prediction by neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| TWI823501B true TWI823501B (en) | 2023-11-21 |
| TW202405824A TW202405824A (en) | 2024-02-01 |
Family
ID=89722718
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| TW111128479A TWI823501B (en) | 2022-07-29 | 2022-07-29 | Photoplethysmography based non-invasive blood glucose prediction by neural network |
Country Status (1)
| Country | Link |
|---|---|
| TW (1) | TWI823501B (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108652640A (en) * | 2018-02-06 | 2018-10-16 | 北京大学深圳研究生院 | A kind of Noninvasive Blood Glucose Detection Methods and system based on electrocardiosignal |
| TW201933159A (en) * | 2018-01-16 | 2019-08-16 | 中央研究院 | A system and method for non-invasively estimating HBA1C and blood glucose level |
| CN113520380A (en) * | 2021-06-15 | 2021-10-22 | 西安电子科技大学 | Noninvasive blood glucose estimation method based on ECG and PPG signals |
| US20220192494A1 (en) * | 2020-12-18 | 2022-06-23 | Movano Inc. | Method for generating training data for use in monitoring the blood glucose level of a person that utilizes a pulse wave signal generated from radio frequency scanning |
-
2022
- 2022-07-29 TW TW111128479A patent/TWI823501B/en active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| TW201933159A (en) * | 2018-01-16 | 2019-08-16 | 中央研究院 | A system and method for non-invasively estimating HBA1C and blood glucose level |
| CN108652640A (en) * | 2018-02-06 | 2018-10-16 | 北京大学深圳研究生院 | A kind of Noninvasive Blood Glucose Detection Methods and system based on electrocardiosignal |
| US20220192494A1 (en) * | 2020-12-18 | 2022-06-23 | Movano Inc. | Method for generating training data for use in monitoring the blood glucose level of a person that utilizes a pulse wave signal generated from radio frequency scanning |
| CN113520380A (en) * | 2021-06-15 | 2021-10-22 | 西安电子科技大学 | Noninvasive blood glucose estimation method based on ECG and PPG signals |
Also Published As
| Publication number | Publication date |
|---|---|
| TW202405824A (en) | 2024-02-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Loh et al. | Application of photoplethysmography signals for healthcare systems: An in-depth review | |
| Delbeck et al. | Non-invasive monitoring of blood glucose using optical methods for skin spectroscopy—Opportunities and recent advances | |
| Monte-Moreno | Non-invasive estimate of blood glucose and blood pressure from a photoplethysmograph by means of machine learning techniques | |
| Nirala et al. | Detection of type-2 diabetes using characteristics of toe photoplethysmogram by applying support vector machine | |
| Maqsood et al. | A survey: From shallow to deep machine learning approaches for blood pressure estimation using biosensors | |
| WO2002087434A1 (en) | Method of judging efficacy of biological state and action affecting biological state, judging apparatus, judging system, judging program and recording medium holding the program | |
| CN114173647A (en) | Method and system for assessing disease using dynamic analysis of cardiac and photoplethysmographic signals | |
| TWI678630B (en) | A system and method for non-invasively estimating hba1c and blood glucose level | |
| Al Fahoum et al. | Identification of coronary artery diseases using photoplethysmography signals and practical feature selection process | |
| Chowdhury et al. | Deep learning via ECG and PPG signals for prediction of depth of anesthesia | |
| Mandal et al. | Detection of Ventricular Arrhythmia by using Heart rate variability signal and ECG beat image | |
| Wu et al. | Multiscale cross‐approximate entropy analysis as a measure of complexity among the aged and diabetic | |
| Scano et al. | NIRS-EMG for clinical applications: A systematic review | |
| Roy et al. | On-device reliability assessment and prediction of missing photoplethysmographic data using deep neural networks | |
| Manurung et al. | Non-invasive blood glucose monitoring using near-infrared spectroscopy based on Internet of Things using machine learning | |
| Kanawade et al. | Photoplethysmography based arrhythmia detection and classification | |
| Liu et al. | Cuffless blood pressure measurement using smartwatches: a large-scale validation study | |
| Susana et al. | Non-Invasive Classification of Blood Glucose Level Based on Photoplethysmography Using Time–Frequency Analysis | |
| Padmavilochanan et al. | Personalized diabetes monitoring platform leveraging IoMT and AI for non-invasive estimation | |
| Cheng et al. | Heart rate variability in different sleep stages is associated with metabolic function and glycemic control in type 2 diabetes mellitus | |
| TWI823501B (en) | Photoplethysmography based non-invasive blood glucose prediction by neural network | |
| Athaya et al. | An efficient fingertip photoplethysmographic signal artifact detection method: A machine learning approach | |
| Li et al. | Noninvasive Blood Glucose Monitoring Using Spatiotemporal ECG and PPG Feature Fusion and Weight-Based Choquet Integral Multimodel Approach | |
| Nampoothiri et al. | Comparison of infrared and red photoplethysmography signals for non-calibrated non-invasive blood glucose monitoring | |
| Abubeker et al. | A Machine Learning Strategy for Internet-of-Things-Enabled Diabetic Prediction to Mitigate Pneumonia Risk |