WO2021038237A1

WO2021038237A1 - Blood pressure measurement device

Info

Publication number: WO2021038237A1
Application number: PCT/GB2020/052064
Authority: WO
Inventors: Dingchang Zheng; Syed Ghufran KHALID
Original assignee: Coventry University
Priority date: 2019-08-30
Filing date: 2020-08-28
Publication date: 2021-03-04
Also published as: GB201912480D0

Abstract

A blood pressure, BP, measurement device. The device comprising: a sensor configured to obtain a pulse measurement from a patient to which the BP measurement device is attached; and a processor. The processor being configured to perform the steps of: (a) receiving the pulse measurement; (b) classifying the pulse measurement as being indicative of one of: a hypotensive BP; a normotensive BP; and a hypertensive BP; (c) selecting an algorithm from a plurality of algorithms, each of the plurality of algorithms being associated with one of: a hypotensive BP; a normotensive BP; and a hypertensive BP, for use in an estimation of a systolic BP, SBP, and a diastolic BP, DBP, of the patient, the selection being based on the result of the classification in step (b); and (d) estimating the SBP and the DBP of the patient using the pulse measurement and the algorithm selected in step (c).

Description

BLOOD PRESSURE MEASUREMENT DEVICE

Field of the Invention

The present invention relates to a blood pressure (BP) measurement device, and specifically to a cuffless BP measurement device.

Background

BP is a vital sign that reflects the state of the cardiovascular system. Abnormal BP can be indicative of various cardiovascular anomalies, including: stroke; peripheral arterial diseases; heart attack (myocardial infarction); kidney failure; and vascular dementia.

BP is typically classified into three BP categories: hypotensive, in which the BP is low; normotensive, which is the typical or ‘healthy’ BP; and hypertensive, in which the BP is high. Hypertension is defined as having a systolic BP, SBP, (maximum BP reached during ventricular contraction) exceeding 140 mmHg and having a diastolic BP, DBP, (the minimum pressure attained in arteries during ventricular relaxation) exceeding 90 mmHg [Tortora, G, J, & Derrickson, B 2012, Principles of anatomy & physiology, 13th edn, Hoboken, NJ] Hypotension is defined as having a SBP lower than 90 mmHg and a DBP lower than 60 mmHg. The SBP and DBP values are typically quoted together e.g. 90/60 mmHg for hypotension and 140/90mmHg for hypertension. Normotensive is a BP value within the normal ranges, i.e. above hypotensive and below hypertensive. This is shown graphically in Figure 1.

Conventional automatic cuff-based BP measurement devices are used in clinics and hospitals in order to both monitor BP and aid diagnosis. However, cuff-based BP measurement procedures cause inconvenience and a degree of discomfort due to the frequent cuff inflation and deflation around the arm or wrist.

To overcome these limitations, so called ‘cuff-less’ BP measurement devices have been proposed. A typical example utilises the pulse transit time (PTT) principles, which is the travel time of the arterial pulse from the heart to a peripheral artery, such as a fingertip artery. However, in addition to a sensor measuring the arrival time at the finger, a separate device is required to measure the reference point from the heart (e.g. an electrocardiogram).

Moreover, patients who require close clinical care are also typically monitored for deviations in pulse or blood oxygen saturation. Typically, these are measured by a photoplethysmograph (PPG) such as that shown in Figure 2. The main components of a PPG are a light emitting diode, and a photodetector. Typically both operate in the infrared region of the electromagnetic spectrum. The PPG produces a photoplethysmogram. The photoplethysmogram, a record of the change in blood volume, can be analysed to measure the pulse and blood oxygen saturation of the patient (in this case a LED within the visible spectrum is required, often red). A typical photoplethysmogram is shown in Figure 3, for 5 complete cardiac cycles. The upper circles indicate the systolic peaks and the lower circles indicate the onset of a cycle.

Attempts have been made to use a photoplethysmogram to derive an indication of BP, for example by Y. Zhang and Z. Feng in “A SVM Method for Continuous BP Estimation from a PPG signal”, ICMLC, Singapore, 24-26 Feb, 2017, pp. 128-132. However, this study does not achieve the AAM l/ISO standards for BP measurement which are a mean difference no greater than ±5 mmHg and a standard deviation of difference no greater than ± 8 mmHg against the reference BP from manual BP measurement.

Accordingly, there is a desire to provide a cuff-less BP measurement device which conforms to these standards and so would be applicable within a clinical context.

Summary

In a general aspect, embodiments of the invention provide a BP measurement device which utilises a two-stage approach to estimate BPs: (i) classification of a photoplethysmogram into one of a number of BP categories; and (ii) estimation using one or more algorithms tailored to that BP category.

Accordingly, in a first aspect, embodiments of the present invention provide a blood pressure, BP, measurement device comprising: a sensor, configured to obtain a pulse measurement from a patient to which the BP measurement device is attached; and a processor, configured to perform the steps of:

(a) receiving the pulse measurement;

(b) classifying the pulse measurement as being indicative of one of: a hypotensive BP; a normotensive BP; and a hypertensive BP;

(c) selecting an algorithm from a plurality of algorithms, each of the plurality of algorithms being associated with one of: a hypotensive BP; a normotensive BP; and a hypertensive BP, for use in an estimation of a systolic blood pressure, SBP, and a diastolic blood pressure, DBP, of the patient, the selection being based on the result of the classification in step (b); and

(d) estimating the SBP and the DBP of the patient using the pulse measurement and the algorithm selected in step (c) to provide a measurement of BP.

This two-stage approach, notably steps (b) and (c), have been shown to result in a significantly more accurate estimation of SBP and DBP. Notably, such a device has been shown to comply with the AAMI/ISO standards for BP measurement.

The BP measurement device may have any one or, to the extent that they are compatible, any combination of the following optional features.

The sensor may be a photoplethysmograph (PPG), configured to obtain a photoplethysmogram. The sensor may be a pressure sensor, configured to obtain an arterial pressure waveform. By pulse measurement, it may be meant a measurement based on parameters of a patient’s pulse or indicative of parameters associated with a patient’s pulse.

By algorithm, it may be meant a set of instructions or equations which represent a trained machine learning model. By hypotensive BP, it may be meant a BP below a lower threshold value which may be indicative of hypotension. By hypertensive BP, it may be meant BP above an upper threshold value which may be indicative of hypertension. By normotensive BP, it may be meant a BP between the lower threshold and upper threshold.

The processor may be configured to perform a step, between steps (a) and (b) of pre processing the pulse measurement using one or more of: a noise filtration routine; a baseline removable routine; and a normalization routine. The processor may be configured to perform all three of these routines before moving to step (b).

The processor may be configured to select a pair of algorithms from the plurality of algorithms, the pair of algorithms being associated with one of: a hypotensive BP; a normotensive BP; and a hypertensive BP, and wherein the first algorithm of the pair is associated with the SBP estimation and the second algorithm of the pair is associated with the DBP estimation.

The algorithms used to estimate the SBPs and DBPs may be one of a multiple linear regression (MLR), support vector machine (SVM), regression tree algorithms. Preferably, the algorithm is a regression tree algorithm. The optimisation hyperparameter of regression tree algorithms during training process for hypotensive BP may have a minimum leaf size of 1 for systolic pressure and 7 for diastolic BPs. The training error of regression tree algorithms for hypotensive BP may have a minimum mean squared error of 2.3 mmHg for systolic pressure and 2.4 mmHg for diastolic BPs. The regression tree algorithms for normotensive BPs may have a minimum leaf size of 9 for SBPs and 12 for DBPs. The regression tree algorithms for normotensive BPs may have a minimum mean squared error of 4.0 mmHg for systolic BPs and 3.68 mmHg for diastolic BPs. The regression tree algorithms for hypertensive BP may have a minimum leaf size of 9 for systolic BPs and 21 for diastolic BPs. The regression tree algorithms for hypertensive BP may have a minimum mean squared error of 4.2 mmHg for systolic BPs and 4.4 mmHg for diastolic BPs.

The minimum leaf size may be used a hyperparameter for the optimisation of the regression tree algorithm. The values given above have been shown to increase the accuracy with which BP is estimated.

The processor may be configured to classify the pulse measurement through use of a classification algorithm. For example, any one of: discriminant analysis; support vector machine classification; decision tree classification; and K-nearest neighbour (KNN) classification.

The photoplethysmogram, measured by the photoplethysmograph, may have a waveform, and the processor may be configured to classify the photoplethysmogram or pressure waveform and estimate the SBP and DBP using at least three of the following characteristics of the waveform: a systolic area; a diastolic area; an index of areas; a peak interval; a 75% signal width; a 50% signal width; a 25% signal width; a 10% signal width; a 30% signal width; a 40% signal width; a 60% signal width; a 70% signal width; an 80% signal width; a 90% signal width; a total area; a rising time; and a signal width of at least 20% and no more than 30%.

The arterial photoplethysmogram or pressure pulse may have a waveform, and the processor may be configured to classify the photoplethysmogram or pressure waveform and estimate the SBP and DBP using the following three properties of the waveform: a total area, a rising time, and a signal width of at least 20% and no more than 30%. Preferably, the signal width used is 25%. These three properties have been found to have the greatest impact to the accuracy of the classification and estimation. Multi collinearity testing was undertaken, by determining the variant inflation factor (VI F), to ascertain which of the properties of the waveform had the largest impact. Total area, rising time, and signal width at 25% were found to have VI Fs of 2.6, 2.0, and 7.5 respectively. A VI F of less than 10 indicates that the factor is uncorrelated with other features, and so is a good candidate.

The area under a PPG waveform or pressure waveform may be calculated using a trapezoid method by dividing the area under the PPG wave into number of equal trapezoids. The systolic area may be calculated between the ascending onset and the peak of the waveform. The diastolic area may be calculated between the peak and the descending onset of the waveform.

The rising time of the waveform may be calculated as the time duration between the ascending onset and the peak point on the PPG waveform.

The processor may be configured to filter the pulse measurement before the classification and estimation steps. The processor may be configured to use a Savitzky-Golay filter. In embodiments where the pulse measurement is a photoplethysmogram or pressure waveform, this filter helps avoid the incorrect detection of local minima near the cardiac notch (a feature of the PPG or pressure waveform) due to the present of high-frequency noise.

The processor may be configured to classify the pulse measurement using a trained KNN algorithm. The processor may be configured to measures distances in the KNN algorithm using the Minkowski metric, which is indicated in the optimisation of the algorithm among other metrics. The BP category may be chosen by majority vote. The K-nearest neighbour algorithm may have been trained using a dataset including: hypotensive, normotensive, and hypertensive examples.

The algorithms used for estimating the SBPs and DBPs may be regression tree algorithms. The regression tree algorithms may have each been trained using a dataset which contained BP measurements only within one BP classification. For example, the one or more regression tree algorithms for use in estimating hypotensive BP may have only been trained using hypotensive BP examples.

The BP measurement device may be a cuff-less device. In a second aspect, embodiments of the invention provide a method of measuring patient’s BP, comprising the steps of:

(a) obtaining a pulse measurement from the patient using a BP measurement device including a sensor which is attached to the patient;

(c) selecting an algorithm from a plurality of algorithms, each of the plurality of algorithms being associated with one of: a hypotensive BP; a normotensive BP; and a hypertensive BP, for use in an estimation of the SBPs and DBPs of the patient, the selection being based on the classification in step (b); and

(d) estimating a SBP and DBP of the patient using the pulse measurement and the algorithm selected in step (c) to provide a measurement of BP.

The BP measurement device of the method may include any, or any combination insofar as they are compatible, of the optional features discussed in relation to the first aspect.

The steps (b) - (d) may be performed using a processor of the BP measurement device.

The sensor may be a photoplethysmograph (PPG), configured to obtain a photoplethysmogram. The sensor may be a pressure sensor, configured to obtain a pressure waveform. By pulse measurement, it may be meant a measurement based on parameters of a patient’s pulse or indicative of parameters associated with a patient’s pulse.

The method may include selecting a pair of algorithms from the plurality of algorithms, the pair of algorithms being associated with one of: a hypotensive BP; a normotensive BP; and a hypertensive BP, and the first algorithms of the pair maybe associated with a systolic BP estimation and the second algorithm of the pair may be associated with diastolic BP estimation.

The algorithms used to estimate the SBP and DBP may be regression tree algorithms. The regression tree algorithms for hypotensive BP may have a minimum leaf size of 1 for systolic pressure and 7 for diastolic BPs. The regression tree algorithms for hypotensive BP may have a minimum mean squared error (^(Estimated - objective function)²) of 2.3 mmHg for systolic pressure and 2.4 mmHg for DBPs . The regression tree algorithms for normotensive BPs may have a minimum leaf size of 9 for SBPs and 12 for DBPs. The regression tree algorithms for normotensive BPs may have a minimum mean squared error of 4.0 mmHg for SBPs and 3.68 mmHg for DBPs. The regression tree algorithms for hypertensive BP may have a minimum leaf size of 9 for systolic BPs and 21 for DBPs. The regression tree algorithms for hypertensive BP may have a minimum mean squared error of 4.2 mmHg for SBPs and 4.4 mmHg for DBPs.

The classification of the pulse measurement may be done through use of a classification algorithm.

The photoplethysmogram or pressure pulse, measured by the photoplethysmograph, may have a waveform, and the method may include classifying the photoplethysmogram and estimating the systolic BP and diastolic BP using at least three of the following properties of the waveform: a systolic area; a diastolic area; an index of areas; a peak interval; a 75% signal width; a 50% signal width; a 25% signal width; a 10% signal width; a 30% signal width; a 40% signal width; a 60% signal width; a 70% signal width; an 80% signal width; a 90% signal width; a total area, a rising time; and a signal width at least 20% and no more than 30%.

The photoplethysmogram or pressure pulse may have a waveform, and the method may include classifying the photoplethysmogram and estimating the systolic BP and diastolic BP using the following three properties of the waveform: a total area, a rising time, and a signal width of at least 20% and no more than 30%.

The method may include filtering the pulse measurement before classifying and estimating the BPs. The filtering may be performed using a Savitzky-Golay filter.

Classifying the pulse measurement may be done using a trained K-nearest neighbour algorithm. The distances in the KNN algorithm may be measured using the Minkowski metric. The BP category may be chosen by majority vote. The K-nearest neighbour algorithm may have been trained using a dataset including: hypotensive, normotensive, and hypertensive examples.

The algorithms used for estimating the SBPs and DBPs may be regression tree algorithms. The regression tree algorithms may be trained using a dataset which contained BP measurements only within one BP classification.

Herein, patient may refer to any subject to whom the device is attached or on whom method is performed. The subject/patient may be an animal, mammal, a placental mammal, a marsupial (e.g., kangaroo, wombat), a monotreme (e.g., duckbilled platypus), a rodent (e.g., a guinea pig, a hamster, a rat, a mouse), murine (e.g., a mouse), a lagomorph (e.g., a rabbit), avian (e.g., a bird), canine (e.g., a dog), feline (e.g., a cat), equine (e.g., a horse), porcine (e.g., a pig), ovine (e.g., a sheep), bovine (e.g., a cow), a primate, simian (e.g., a monkey or ape), a monkey (e.g., marmoset, baboon), an ape (e.g., gorilla, chimpanzee, orangutang, gibbon), or a human. Furthermore, the subject/patient may be any of its forms of development.

Further aspects of the present invention provide: a computer program comprising code which, when run on a computer (for example a mobile phone, tablet PC, smart device, small- form device etc.), causes the computer to perform the method of the second aspect; a computer readable medium storing a computer program comprising code which, when run on a computer, causes the computer to perform the method of the second aspect; and a computer system programmed to perform the method of the second aspect.

Brief Description of the Drawings

Embodiments of the invention will now be described by way of example with reference to the accompanying drawings in which:

Figure 1 shows a plot showing the clinical classifications of BP;

Figure 2 is a figure showing a typical photoplethysmograph;

Figure 3 is a plot of a typical PPG signal waveform;

Figure 4 is a schematic showing components of a BP measurement device according to embodiments of the present invention;

Figures 5A and 5B respectively show two filtering stages used in a photoplethysmograph which is an embodiment of the present invention;

Figures 6A and 6B show a series of systolic and diastolic BPs measured by a BP measurement device according to embodiments of the present invention plotted against a reference BP measurement;

Figure 7 is a flow diagram showing a measurement process of a BP measurement device according to embodiments of the present invention; Figures 8A-8C show, respectively, the composition of three training sets for machine learning algorithms;

Figure 9 is a flow diagram showing a training process for BP category specific machine learning algorithms;

Figure 10 is a flow diagram showing a KNN algorithm used for categorising a PGG waveform;

Figure 11 is a confusion matrix corresponding to the KNN algorithm of Figure 10;

Figure 12 is a flow diagram of a regression tree algorithm used for estimating BP from a PPG waveform;

Figures 13A - 13D show, respectively, Bland-Altman plots for a generic algorithm (13A) and three BP category specific algorithms (13B-13D);

Figures 14A - 14D show, respectively, regression plots for a generic algorithm (14A) and three BP category specific algorithms (14B-14D);

Figures 15A and 15B are a Bland-Altman plot and regression plot respectively of SBP measured using a BP measurement device according to embodiments of the present invention;

Figures 16A and 16B are a Bland-Altman plot and a regression plot respectively of DBP measurement using a BP measurement device according to embodiments of the present invention; and

Figure 17 is a plot showing a comparison between the estimated BP accuracy between a generic algorithm and a method of measuring BP according to embodiments of the present invention.

Detailed Description and Further Optional Features

Aspects and embodiments of the present invention will now be discussed with reference to the accompanying figures. Further aspects and embodiments will be apparent to those skilled in the art. All documents mentioned in this text are incorporated herein by reference. Broadly, a BP measurement device according to some embodiments of the present invention takes the form shown in Figure 4. Broadly, the device includes a photoplethysmograph which is connected to a processor. The processor is connected to a display; a storage medium; and an input/output interface. In use, the photoplethysmograph takes a photoplethysmogram from the patient, which in some examples is then filtered. The photoplethysmogram is then provided to the processor which may further filter it, and then classifies it and estimates a systolic and diastolic BP. This estimation can then be provided to one or more of the display, storage medium, and input/output interface. For example, the estimation may be displayed on the display for a direct reading. It may be stored in the storage medium for subsequent retrieval, and/or it may be transmitted through the input/output interface to another device (e.g. a hospital management system). The input/output interface may be, for example, a Bluetooth (RTM) interface or Wi-Fi (RTM) interface.

A photoplethysmograph according to embodiments of the present invention contains a light source, preferably infrared light emitting diode, and a light photodetector, preferably a light dependent resistor (LDR) tuned for use in the infrared area of the electromagnetic spectrum. The LED and LDR may be on the same side of photoplethysmograph, in that the photoplethysmograph may operate by monitoring the reflection of light from a finger adjacent to the LED and LDR. Alternatively, or additionally, the LED and LDR may be on opposite sides of the photoplethysmograph with a cavity there between for the insertion of a finger. Such a photoplethysmograph may operate by monitoring the transmission of light through the finger. In embodiments where the sensor is a pressure sensor, a pressure applicator may apply pressure on a peripheral part of the body to which it is attached, and the pressure sensor may detect an arterial pulse pressure. The pressure sensor may, for example, take the form of a bracelet or wristband. The sensing component in the pressure sensor may use a piezoresistive sensor to provide blood pulse-wave measurements.

In some examples, a photoplethysmograph or pressure sensor according to the present invention contains the circuits shown in Figures 5A and 5B. The original signal from the sensor (e.g. LDR) may have a low amplitude, and so in a first stage it is passed through an RC high-pass filter (HPF). In this example the HPF has a cut-off frequency of 0.5 Hz which blocks any DC component of the signal. The signal is then passed to the operational amplifier-based low pass filter (LPF) which in this example has a gain of 48 and cut-off frequency of 3.4 Hz. The signal is then passed to the second stage, shown in Figure 4B. Here, the amplification and filtration part (HPF and LFP) are identical in form to those in Figure 4A. After a second filtration pass, the signal exits through a non-inverting buffer stage. The signal is then converted from analogue to digital (e.g. via an ADC), and is processed in a manner discussed below.

Figures 6A and 6B show a series of SBP and DBP measurements taken by a BP measurement device according to embodiments of the present invention. The measurements are plotted as circles with a fit line passing through them. The number of estimations is representative of the duration over which the measurements were taken. Also plotted is a reference BP, obtained through a calibrated cuff-based BP device. The total time (0 - 60 estimations) was five minutes. As can be seen, the device performs well in estimating the systolic and diastolic BPs as compared to the calibrated device. Indeed, the device performs within the expectations set by the AAM l/ISO standards.

Figure 7 is a flow diagram showing the measurement process of measuring BP using a BP measurement device according to the present invention, where the sensor is a photoplethysmograph. In a first step, 701, the photoplethysmogram (PPG) is acquired.

Next, in step 702, the PPG is analysed in order to classify it as representing one of: a hypotensive BP; a normotensive BP; and a hypertensive BP. This is done by using the classification algorithm which has been optimised. In this example, the classification algorithm is a K-nearest neighbours algorithm.

Next, depending on the classification found in step 702, the method proceeds to one of 703a-703c. In step 703a, the specific algorithm(s) for hypotensive BPs is used to estimate the systolic and diastolic BPs. In step 703b, the specific algorithm(s) for normotensive BPs is used to estimate the systolic and diastolic BPs. In step 703c, the specific algorithm(s) for hypertensive BPs is used to estimate the systolic and diastolic BPs. In this example, the algorithms used for estimating BPs are regressive tree algorithms.

After this estimation is complete, a systolic (SBP) and diastolic (DBP) BP is provided.

In embodiments where the sensor is a pressure sensor, the same steps are performed however a pressure waveform is acquired and classified.

The classification and regression tree algorithms used in some embodiments of the present invention, and particularly those used to produce the results in Figures 6A and 6B, were trained using a dataset. The dataset was formed of a combination of the University of Queensland database (Liu et al. 2012) and the Multiparameter Intelligent Monitoring in Intensive Care (MIMIC II) database (User guide, MIMIC II, database version. 2.6, pp.1-70.). Both databases provided PPGs with accompanying systolic and diastolic BPs. The Queensland database comprises data recorded in 4 operating theatres in the Royal Adelaide Hospital, and originally comprised 23,617 entries. The database was filtered to remove those where the PPG was of insufficient quality, and after filtering 8,133 entries remained. The remaining entries were categorised by BP category as shown in Figure 8A. The MIMIC II database comprises data collected from patients within the intensive care units of Beth Israel Deaconess Medical Centre between the years 2001 and 2008. After a similar filtering process as applied to the Queensland database, 9877 entries were retained and categorised as shown in Figure 8B. The two filtered datasets were combined, to produce the dataset shown in Figure 8C, comprising 636 entries corresponding to hypotensive BPs, 11,245 entries corresponding to normotensive BPs, and 6,129 entries corresponding to hypertensive BPs.

The training of KNN algorithm provides the optimal K value at which BP categories are separated with minimal prediction error. During each iteration of training, the K value of the algorithm changes and a check is made of the prediction error. In the KNN algorithm optimisation, a K-value was selected (K=3) which produced the least prediction error (8.3%) in 30 iteration.

After the combined dataset was created, a training process as shown in Figure 9 was used to train the BP specific machine learning algorithms. The training set was subdivided into the three BP categories: hypotensive, normotensive, and hypertensive, in accordance with the definitions given previously. The categorised datasets were then split 50/50 to provide a training and testing dataset. For each category, the specific BP algorithm included two independent algorithms for systolic BP and diastolic BP. The regression tree (Cart) algorithms were used.

Figure 10 shows a flow diagram of the K-nearest neighbours (KNN) algorithm used for categorising the PGG waveforms. In this example the algorithm takes, as its input, the total area, the rising time, and the 25% signal width of the PPG waveform. The KNN algorithm then measures the distances between the inputs and the neighbouring points within the training data. The resulting minimum distances are arranged in ascending order, and a majority check is used to select the BP category. The distance is measured using the Minkowski metric (K=3), per the equation:

Where d_si is the distance, ¾ represents the closest points to a query or set of points y_t}, p represents the order of the distance.

Figure 11 shows the confusion matrix of the KNN algorithm. Class 1 (hypotensive category) had a prediction accuracy of 76% out of the total 508 samples. Class 2 (normotensive category) had a prediction accuracy of 92% out of the total 5,622 samples. Class 3 (hypertensive category) had a prediction accuracy of 93% out of the total 2,875 samples.

As shown in Table 1 below, the KNN classification algorithm had an overall accuracy of 91.70% with a misclassification rate of 8.3%. A maximum specificity of 98.1 % was achieved in the Hypotensive category, with comparable values of 91.4% and 94.3% in Normotensive and Hypertensive categories. The precision (93.5%, 91.4%, and 72.4%) and F-score (92.0%, 91.4%, and 73.0%) indicated identical order from Normotensive, Hypertensive to Hypotensive categories.

Table 1 MC=Misclassification rate; Sens=Sensitivity; FN Rate=False negative rate; Spec=Specificity; FP

Rate=False Positive Rate; Prec=Precision; F-s=F-score; Hypo=Hypotension; Normo=Normotensive; Hyper=Hypertensive

Figure 12 is a flow diagram of a regression tree algorithm used for estimating BP from a PPG waveform. A regression tree algorithm is a non-parametric machine learning approach. One of the advantages of the algorithm is the ability to train and estimate the results with minimal processing time, as compared to other machine learning algorithms. A regression tree is developed with the recursive division of binary data iteratively; and the tree includes branches, internal nodes, and terminal nodes. This is shown in Figure 12. The decision typically presents on the terminal node, and the algorithm picks the decision (estimated BP(s)) from the root node to the leaf node based on the PPG signal features provided.

The CART algorithm was used as the particular implementation of a regression tree algorithm. Next, for each independent variable, the dataset was split at multiple points. At each split data point, the difference between the reference and estimated BPs as calculated as a mean square error, using the equation:

Where W_j is the weight that is 1/n, where k is the sample size, ¾ is the reference BP, y_t is the estimated BP, and T represents the total observations included in the node.

The algorithm calculates the probability for individual observations within the node £ using the equation:

Each node t of data points is split into left and right child nodes (t_L and t_R) and saved in T_L and ¾ sets. Then, all of the errors are compared and the point where the lowest sum of square error occurred is selected as the root node as shown in this equation:

MS£ = P{T^')s_t — P(T_L}s_t, — F(¾)s_tg

These steps are recursively repeated until the algorithm converged. The input parameters (Sig.1 - Sig. 3) in this example were: total area; rising time; and width at 25% signal.

A regression tree’s depth is controlled by the maximum number of splits. The algorithm compares the number of split nodes and maximum number of split value. When the number of split nodes exceeds the limit, the algorithm reduces the least successful nodes from the Regression tree in order to remain within the set limit of a maximum number of splits. This reduction is also referred to as pruning.

In order to compare the performance of the two-stage approach (classify, and estimate based on the classification) a generic regression tree algorithm was trained on the combined dataset without having divided it into BP categories. An overall summary of the performance of the generic algorithm as compared to the specific SBP algorithms is shown in Table 2 below:

Table 2

An overall summary of the performance of the generic algorithm as compared to the specific DBP algorithms is shown in Table 3 below:

Table 3 Figures 13A - 13D are Bland-Altman plots for: the generic algorithm (13A); the hypotensive DBP algorithm (13B); the normotensive DBP algorithm (13C); and the hypertensive DBP algorithm (13D).

The generic algorithm (Figure 13A) displayed a bias of -O.lmmHg which conforms to the ISO standard requirements (±5 mmHg). However, the width of the confidence interval (±18.4 mmHg) was outside of the limits set by the ISO standard (±16 mmHg).

The ISO standards were achieved for all of the specific DBP algorithms. For hypotensive, normotensive, and hypertensive DBP algorithms (Figures 13B-D) the biases were: 0.04mmHg, -0.03mmHg, and -0.03mmHg. The widths of the confidence intervals were 5.5mmHg to 5.6mmHg, ±11.4mmHg, and -8.5mmHg to 8.4mmHg. Figures 14A - 14D are regression plots for: the generic algorithm (14A); the hypotensive DBP algorithm (14B); the normotensive DBP algorithm (14C); and the hypertensive DBP algorithm (14D).

The generic algorithm (Figure 14A) achieved a slope of 0.9 and R-squared value of 0.6 (p<0.001)) indicating the general effectiveness of this algorithm. In contrast, the regression plot for the hypotensive DBP algorithm (Figure 14B) shows an increased accordance with reference and estimated DBP values (slope 0.8) with the R-squared value (0.7, p<0.001).

The regression plot for the normotensive DBP algorithm (Figure 14C) showed similar accordance with the generic algorithm, between the reference and estimated DBP values (slope: 0.8) with the R-squared value (0.6, p<0.001). Finally, the regression plot for the hypertensive DBP algorithm in Figure 14D showed the best accordance between the reference and estimated DBP values (slope 0.9) with the highest R-squared value (0.8, pO.001).

Further evaluation was performed of the two-step approach, i.e. classifying and subsequently applying a BP specific algorithm to estimate the BP. Figure 15A is a Bland- Altman plot of the estimation of systolic BP as compared to a reference (directly measured) systolic BP taken at the same time. The solid line shows the bias, and the dashed lines show the upper (14.12 mmHg) and lower (-14.27 mmHg) limits of agreement. Of the 9005 segments, 3% of the total distribution lay above the limits of the ISO standard (±16mmHg) and 4% of the total distribution lay below the limits of the ISO standard. The bias was found to be -0.07 mmHg and the width of the confidence interval was found to be ±14.2 mmHg which conforms to the ISO standards (±5 mmHg for bias and ±16 mmHg for confidence interval).

Figure 15B is a regression plot between the reference SBP and estimated SBP. The plot shows a slope which is close to 1 , which indicates an excellent level of consistency between estimated and measured values. The outliers outside of the 95% interval are generally symmetric in distribution. The R-squared value (0.9, p=0.4) indicates considerable effectiveness of fitting.

Figure 16A is a Bland-Altman plot of the estimation of diastolic BP as compared to a reference (directly measured) systolic BP taken at the same time. The solid line shows the bias, and the dashed lines show the upper (12.2 mmHg) and lower (-12.0 mmHg) limits of agreement. The bias was found to be 0.08 mmHg, and the limit of agreement (i.e. confidence width) was found to be ±12.1 mmHg which are both acceptable with the ISO standards.

Figure 16B is a regression plot between reference DBP and estimated DBP. The plot shows a slope which is close to 0.7, indicating excellent consistency between estimated and measured values. The outliers outside the 95% interval are generally symmetric in distribution. The R-squared value (0.7, p<0.001) indicates considerable effectiveness of fitting.

Figure 17 is a comparison of the two-step approach over the generic algorithm. As can be seen, the p-values show that the difference between the two-step and generic estimation accuracies is significant in both SBP and DBP. It can therefore be concluded that the two- step approach provides a considerable improvement in BP prediction accuracy as compared to the generic algorithm. Moreover, as has been discussed previously, the two-step approach results in a device which conforms to the ISO standards. Whereas the generic approach was found to be beyond the accept range defined by the ISO standards.

In addition to these statistically based performance metrics, data was obtained from 53 normotensive participants in three different postures (supine, sitting, and standing-up). A finger photoplethysmograph was attached to each participant, as was a cuff-based Omron BP monitor. The data was paired, such that a reference BP existed for each PPG recorded by the photoplethysmograph.

The two-step approach algorithms for systolic BP and diastolic BP were then validated according to the British Hypertension Society (BHS) standards, as shown in Table 4 below. The two-step approach achieved an A grade for DBP measured in the supine position with the 64% estimation errors being less than or equal to 5 mmHg which means the majority of errors are within the standard criteria of AAMI/ISO. The supine and stand-up postures the DBP algorithm achieved a B grade, as the 50% errors were less than or equal to 5 mmHg. Finally, the best grade of B for SBP estimation was achieved in the supine position. Whereas sitting and standing postures achieved a C grade.

Table 4

While the invention has been described in conjunction with the exemplary embodiments described above, many equivalent modifications and variations will be apparent to those skilled in the art when given this disclosure. Accordingly, the exemplary embodiments of the invention set forth above are considered to be illustrative and not limiting. Various changes to the described embodiments may be made without departing from the spirit and scope of the invention.

All references referred to above are hereby incorporated by reference.

Claims

1. A blood pressure, BP, measurement device, comprising: a sensor, configured to obtain a pulse measurement from a patient to which the BP measurement device is attached; and a processor, configured to perform the steps of:

(a) receiving the pulse measurement;

(c) selecting an algorithm from a plurality of algorithms, each of the plurality of algorithms being associated with one of: a hypotensive BP; a normotensive BP; and a hypertensive BP, for use in an estimation of a systolic BP, SBP, and a diastolic BP, DBP, of the patient, the selection being based on the result of the classification in step (b); and

(d) estimating the SBP, and the DBP of the patient using the pulse measurement and the algorithm selected in step (c) to provide a measurement of BP.

2. The BP measurement device of claim 1 , wherein the sensor is a photoplethysmograph, PPG, configured to obtain a photoplethysmogram from the patient.

3. The BP measurement device of claim 1 or 2, wherein the processor is configured to select a pair of algorithms from the plurality of algorithms, the pair of algorithms being associated with one of: a hypotensive BP; a normotensive BP; and a hypertensive BP, and wherein a first algorithm of the pair is associated with systolic BP estimation and the second algorithm of the pair is associated with diastolic BP estimation.

4. The BP measurement device of any of claims 1 to 3, wherein the algorithms used to estimate the SBP and DBP are regression tree algorithms.

5. The BP measurement device of claim 4 as dependent on claim 3, wherein the regression tree algorithms for hypotensive BPs have a minimum leaf size of 1 for systolic BPs and 7 for diastolic BPs.

6. The BP measurement device of either claim 4 or claim 5 as dependent on claim 3, wherein the regression tree algorithms for normotensive BPs have a minimum leaf size of 9 for SBPs and 12 for DBPs.

7. The BP measurement device of any of claims 4 - 6 as dependent on claim 3, wherein the regression tree algorithms for hypertensive BP have a minimum leaf size of 9 for SBPs and 21 for DBPs.

8. The BP measurement device of any preceding claim, wherein the processor is configured to classify the pulse measurement though use of a classification algorithm.

9. The BP measurement device of any preceding claim, wherein the pulse measurement is a photoplethysmogram and has a waveform, and the processor is configured to classify the photoplethysmogram or pressure waveform and estimate the SBP and DBP using at least three of the following properties of the waveform: a systolic area; a diastolic area; an index of areas; a peak interval; a 75% signal width; a 50% signal width; a 25% signal width; a 10% signal width; a 30% signal width; a 40% signal width; a 60% signal width; a 70% signal width; an 80% signal width; a 90% signal width; a total area, a rising time; and a signal width at least 20% and no more than 30%.

10. The BP measurement device of any preceding claim, wherein the pulse measurement is a photoplethysmogram and has a waveform, and the processor is configured to classify the photoplethysmogram or pressure waveform and estimate the SBP and DBP using the following three properties of the waveform: a total area, a rising time, and a signal width of at least 20% and no more than 30%.

11. The BP measurement device of any preceding claim, wherein the processor is configured to filter the pulse measurement before the classification and estimation steps.

12. The BP measurement device of claim 11 , wherein the processor is configured to filter the pulse measurement using a Savitzky-Golay filter.

13. The BP measurement device of any preceding claim, wherein the processor is configured to classify the pulse measurement using a trained K-nearest neighbour algorithm.

14. The BP measurement device of claim 13, wherein the processor is configured to measure distances in the K-nearest neighbour algorithm using the Minkowski metric.

15. The BP measurement device of either claim 13 or claim 14, wherein the BP category is chosen by majority vote.

16. The BP measurement device of any of claims 13 to 15, wherein the K-nearest neighbour algorithm was trained using a dataset including: hypotensive, normotensive, and hypertensive examples.

17. The BP measurement device of any preceding claim, wherein the algorithms used for estimating the SBPs and DBPs are regression tree algorithms.

18. The BP measurement device of claim 17, wherein the regression tree algorithms were each trained using a dataset which contained BP measurements only within one BP classification.

19. The BP measurement device of any preceding claim, wherein the BP measurement device is a cuff-less device.

20. A method of measuring a patient’s blood pressure, BP, comprising the steps of:

21. The method of claim 20, wherein the sensor of the BP measurement device is a photoplethysmograph.

22. A computer readable storage medium containing instructions which, when executed on a computer, cause the computer to carry out the method of claim 19.