CN109875546A - A kind of depth model classification results method for visualizing towards ECG data - Google Patents
A kind of depth model classification results method for visualizing towards ECG data Download PDFInfo
- Publication number
- CN109875546A CN109875546A CN201910067724.0A CN201910067724A CN109875546A CN 109875546 A CN109875546 A CN 109875546A CN 201910067724 A CN201910067724 A CN 201910067724A CN 109875546 A CN109875546 A CN 109875546A
- Authority
- CN
- China
- Prior art keywords
- model
- result
- section
- heartbeat
- value
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Landscapes
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
Abstract
The invention discloses a kind of depth model classification results method for visualizing towards ECG data, comprising: by the trained depth model of electrocardiogram sequence inputting, obtain benchmark result;By blocking the information in the heartbeat section for erasing selected in section, depth model output result when by not selected heartbeat block information compares with the benchmark result that depth model exports, and calculates and obtains heartbeat each time for the impact factor Δ O of depth model;The impact factor Δ O visable representation of heartbeat each time is come out using gradient color band, realizes the visualization of depth model classification results.The present invention exports depth model by ECG data under analysis two kinds of granularities of both macro and micro the influence of result, can show to obtain the key evidence of category of model result, can enhance the interpretation of the classification results of model output.
Description
Technical field
It is the invention belongs to depth model classification results visualization technique field, in particular to a kind of towards ECG data
Depth model classification results method for visualizing.
Background technique
According to the definition of wikipedia, ECG data refers to a kind of transthoracic electricity that heart is recorded as unit of the time
Physiological activity, and the data for capturing and recording by the electrode on skin.In practice, in order to improve efficiency, mitigate doctor
Raw burden and working strength, some models based on deep learning are applied to the feature extraction and classifying on ECG data
On.But these existing models can only provide last classification results, can not generation to the classification results according to making explanations;
And the classification results prediction that do not explain clearly in practice is difficult to be received and apply, so that application scenarios are significantly limited,
It is unfavorable for the classification results that doctor utilizes model output.
To sum up, a kind of depth model classification results method for visualizing towards ECG data is needed.
Summary of the invention
The purpose of the present invention is to provide a kind of depth model classification results method for visualizing towards ECG data, with
Solve above-mentioned technical problem.The present invention can show to obtain the key evidence of final result, can enhance model output
The interpretation of classification results.
In order to achieve the above objectives, the invention adopts the following technical scheme:
A kind of depth model classification results method for visualizing towards ECG data, comprising the following steps:
Step 1, the ECG data of acquisition is handled as electrocardio graphic sequence, by the trained depth of electrocardiogram sequence inputting
In model, benchmark result is obtained;
Step 2, using eartbeat interval as basic unit, according to the heartbeat message dynamic adjustment blocked area in ECG data
Between, the depth model output by blocking the information in the heartbeat section for erasing selected in section, when by without the heartbeat block information
Compared with the benchmark result that depth model exports when as a result with comprising the heartbeat message, calculates and obtain heartbeat each time for depth
The impact factor Δ O of model;
Step 3, the impact factor Δ O visable representation of heartbeat each time is come out using gradient color band, realizes depth mould
The visualization of type classification results.
Further, further includes:
Step 4, setting movably blocks section, successively blocks each point in ECG data;By ECG data
The depth model for blocking each point exports result respectively compared with the benchmark result of depth model output, on acquisition ECG data
Each pair of point is in the impact factor of depth model output result;
Step 5, the impact factor of each point step 4 obtained carries out visable representation.
Further, step 2 specifically includes:
Step 2.1, according to original electrocardiographicdigital diagram data, the length in each heartbeat section is obtained, is set dynamically according to the length
Section is blocked, each heartbeat section is successively blocked;
Step 2.2, the electrocardiogram sequence vector for blocking section will be added to be separately input in depth model, is obtained new
Depth model exports result;
Step 2.3, each new the depth model output result and step 1 for calculating separately step 2.2 acquisition obtain benchmark
As a result difference obtains each heartbeat section to the impact factor of depth model output result.
Further, step 3 specifically includes:
Step 3.1, the corresponding Δ O value in each heartbeat section is encoded, obtains a corresponding colour sequential;Rule
Are as follows: as Δ O > 0, it is encoded to a kind of pre-set color, the value is bigger, then color depth is deeper;As Δ O < 0, compiled
Code is another different pre-set color, and the value is smaller, then color depth is deeper;
Step 3.2, using each heartbeat siding-to-siding block length as rectangle width, using the height at the peak highest R on electrocardiogram as rectangle
ECG data sequence is divided into several rectangles by length, and each rectangle includes a heartbeat section;Step 3.1 is obtained every
The color filling that a heartbeat Interval Coding generates is into the corresponding rectangle in each heartbeat section;
Step 3.3, the corresponding coloured rectangle of filling in each heartbeat section that step 3.2 obtains is added to electrocardiogram
In data background, the visualization of depth model classification results is realized.
Further, in step 3.2, rectangular centre is set as transparent, and both ends are set as Fill Color, and rectangle is adjusted to
Gradient color band.
Further, step 1 specifically includes:
ECG data processing is the representation after electrocardio graphic sequence are as follows:
S=[s1,s2,…,si,…,sn]
In formula, S is n-dimensional vector, i=1,2 ..., n, siIndicate i-th point in sequence of data;
By electrocardiogram sequence inputting into trained depth model, obtained result data format are as follows:
Y=[y1,y2,…,yj,…,yN]
In formula, Y is N-dimensional vector, and N indicates the number of labels of category of model;J=1,2 ..., N, yjIndicate model in label
Classification value on j, 0≤yj≤1;
Wherein, yjCorresponding label is the prediction classification results of depth model when being maximized, by the corresponding y of the labelj
Value is set to a reference value O, and label sequence number is set as I, the expression formula of a reference value O are as follows:
O=max { y1,y2,…,yj,…,yN}
In formula, yjIndicate classification value of the model on label j, 0≤yj≤1。
Further, step 2.1, it is dynamically determined and blocks siding-to-siding block length;
From original electrocardiographicdigital diagram data, the peak position the R label of each heartbeat is obtained, is considered primary between two peaks R
The section RR of heartbeat;It is arranged k-th and blocks siding-to-siding block length are as follows:
Lengthk=xk+1-xk
In formula, LengthkIndicate the length for blocking section being arranged on k-th of section RR, xkIndicate k-th of peak position R
Abscissa, 0≤xkThe total length of≤Len, Len expression electrocardio graphic sequence;
Step 2.2, each heartbeat block information is calculated for the impact factor of depth model output result;
Step 2.2.1 will block the R peak position alignment of section starting position and kth time heartbeat, and siding-to-siding block length is set as
Lengthk, so that blocking section covering kth time heartbeat block information;
The vector value blocked in section is uniformly assigned a value of 0 by step 2.2.2, and the vector value of remaining position remains unchanged, and repairs
Electrocardio graphic sequence after changing are as follows:
Sk=[s1, s2..., 0 ..., 0 ..., sn]
Wherein, siIt indicates i-th point of data in sequence, is assigned a value of 0 region since the peak R of kth time heartbeat, length
For Lengthk;
Step 2.2.3 will be added to the electrocardio graphic sequence S for blocking sectionkVector is input in depth model, is obtained new
Depth model exports result Yk, YkIt is N-dimensional vector, expression formula are as follows:
Yk=[y '1,y′2..., y 'N]
In formula, y '1, y '2..., y 'N1,2 are illustrated respectively in ..., the output valve on N label;
Step 2.2.4 calculates the depth model result O for obtaining and blocking kth time heartbeat block informationkWith the difference of benchmark result
It is worth Δ Ok;ΔOkFor the impact factor in k-th of heartbeat section, expression formula are as follows:
ΔOk=yI-y′I
In formula, I indicates the label sequence number for a reference value O that step 1 is calculated, yIWith y 'IIt indicates in the label sequence number
Depth model output valve;ΔOkIndicate k-th of heartbeat section for the impact factor of depth model output result;ΔOk> 0 indicates
The heartbeat section has positive influences to category of model result, is the supporting evidence of model, and the value is bigger, expression and category of model
As a result more agree with;ΔOk< 0 indicates that the heartbeat section has negative effect to final classification result, is the opposition evidence of model, should
Value is negative value, and the smaller expression of value more deviates from category of model result;
By the numerical value of Δ O, influence of the different heartbeat sections for category of model result is distinguished, is realized to category of model knot
The explanation of fruit.
Further, step 4 specifically includes:
Step 4.1, since first data of electrocardio graphic sequence S vector, 0 will be set to by L vector Value Data later,
The vector value of remaining position remains unchanged, and section is blocked in formation;Section is blocked since first data, moves backward one every time
Lattice, until all data in traversal ECG data;
The electrocardio graphic sequence S for blocking section is added to when the m times circulationmVector data expression formula are as follows:
Sm=[s1,s2..., sm-1, 0,0 ..., 0, sm+L,…,sn]
Wherein s1,s2,…,snThe individual data for indicating composition electrocardio graphic sequence, by the formula it is found that sm,sm+1...,
sm+L-1It is added to block section, the data in section are all assigned 0;
Step 4.2, the difference of depth model output result and benchmark result behind section is blocked in node-by-node algorithm setting, obtains the heart
The impact factor Δ O numerical value that each is put on electrograph;
Specific steps include:
Step 4.2.1 is added to the electrocardio graphic sequence S for blocking section when by the m times circulationmVector is input to depth model
In, obtain the output result Y of modelm, expression formula are as follows:
Ym=[y '1, y '2..., y 'N]
In formula, y '1,y′2,…,y′N1,2 are illustrated respectively in ..., the output valve on N label;
Step 4.2.2 calculates the difference between the new model output result and model reference result that step 4.2.1 is obtained
ΔOm, value reflection individually influence of the point for model output result, calculation formula are as follows:
ΔOm=yI-y′I
In formula, I indicates the label sequence number for a reference value O being calculated, yIWith y 'IIndicate the depth in the label sequence number
Model output value;ΔOmIndicate that than the m-th data is for the impact factor of depth model output result in heart sequence;ΔOm> 0 table
Show that the point has positive influences to final classification result, be the supporting evidence of model, the value is bigger, indicates and model final result
More agree with;ΔOm< 0 indicates that the point has negative effect to final classification result, is the opposition evidence of model, which is negative value,
It is worth smaller expression more to deviate from final result;
By the numerical value of Δ O, obtains each pair of point on electrocardiogram and realize electrocardiogram in the impact factor of category of model result
The explanation of detailed information in data.
Further, step 5 specific steps include:
Step 5.1, the Δ O numeric coding of each point step 4.2 obtained is height, and passes through the position of the point and height
Spend determine electrocardio plan on a point P, Δ O > 0, indicate point P electrocardiogram upper area, and will on electrocardiogram correspondence
Point is shown as a kind of pre-set color;Δ O=0 indicates that point P is fallen in zero axle, and corresponding points on electrocardiogram are shown as another pre-
If color;Δ O < 0 indicates that corresponding points on electrocardiogram in the lower zone of electrocardiogram, are shown as another pre-set color by point P;
Pre-set color is all different;
It step 5.2, the use of smoothed curve will be that the point that ordinate is formed connects with serial number abscissa, Δ O numerical value,
And surround out several regions jointly with zero axle;The size of the height reflection Δ O absolute value of curve, spike and the low ebb reflection of curve
Support model result and the crucial foundation for violating model result;
Step 5.3, depth model classification is realized in the region surrounded using preset 5.2 curve of different colours filling step
Result visualization.
Further, in step 4, the range for blocking the length L in section is 10≤L≤20.
Compared with prior art, the invention has the following advantages:
Depth model classification results method for visualizing towards ECG data of the invention, devises from the overall situation to details
Visualization result show process, can completely show influences the crucial foundation obtained a result of model.Method of the invention first will
The original electrocardiographicdigital diagram data of acquisition is input in depth model, is obtained the output data of depth model, is analyzed according to output data
It determines the classification results of prediction, and result and participate in subsequent comparison on the basis of output data is saved, obtains impact factor;So
It combines heartbeat message dynamic setting to block interval parameter afterwards, obtains the shadow that model final result is predicted in each heartbeat section
It rings, and is intuitively shown with visualization method;Further design movably blocks section, calculates the deviation of each point and benchmark
Value, which is superimposed with original electrocardiographicdigital diagram data, shows the minutia in ECG data, side by peak value and region area
Just it searches and there is abnormal details area.The present invention blocks interval computation specific region for the shadow of final result by setting
It rings, by influence of the ECG data for final mask result under analysis two kinds of granularities of both macro and micro, can show model
The key evidence of final result is obtained, the interpretation of model result can be enhanced.
Method for visualizing of the invention can enhance the interpretation of model result;Conventional method drag result is one
A specific classification results label has no idea to explain the foundation for obtaining the result, and such result is more difficult to be adopted and use.
Method of the invention is made that explanation for model result, has found supporting evidence and oppose evidence that model is obtained a result, exhibition
The influence that each details obtains model final result is shown, the interpretation of model result can be greatly promoted.
The present invention visualizes interpretation process from the angle of both macro and micro;Electrocardiogram number under conventional method
According to mixed and disorderly tediously long, it is time-consuming and laborious for therefrom differentiating key message.It is likely to be crucial to the region that model result is affected
Property abnormal area, such as there are P wave disappear etc. abnormal phenomenon.Method of the invention excavates such area from ECG data
Domain, and show it by visualized elements such as color, height from two kinds of granularities of both macro and micro, so that model be made to transport
Row process is more intuitive, further improves the interpretation of model result.
Method of the invention is suitable for various deep learning models, and scalability is strong;Interpretation model result under conventional method
Model structure is needed to refer to, can not be expanded on other models.Method of the invention is not rely on particular model, all to be applicable in
This method can be used in the depth model classification results of ECG data to explain and show, and can easily expand to
On the improved model to emerge one after another at present.
Detailed description of the invention
Fig. 1 is a kind of process signal of depth model classification results method for visualizing towards ECG data of the invention
Block diagram;
Fig. 2 is heartbeat section in a kind of depth model classification results method for visualizing towards ECG data of the invention
Influence the schematic process flow diagram of method for visualizing;
Fig. 3 is influenced point by point in a kind of depth model classification results method for visualizing towards ECG data of the invention
The schematic process flow diagram of method for visualizing;
Fig. 4 is heartbeat section in a kind of depth model classification results method for visualizing towards ECG data of the invention
Influence visualization result schematic diagram;
Fig. 5 is influenced point by point in a kind of depth model classification results method for visualizing towards ECG data of the invention
Visualization result schematic diagram.
Specific embodiment
Invention is further described in detail in the following with reference to the drawings and specific embodiments.
A kind of depth model classification results method for visualizing towards ECG data of the invention, specifically includes following step
It is rapid:
Step 1, acquisition obtains the ECG data after diagnosing of preset quantity, and each ECG data is handled as the heart
Electrograph sequence, the trained depth model that each electrocardiogram sequence inputting is selected obtain depth model output as a result, by this
When output result be set to depth model output benchmark result.
The processing of original electrocardiographicdigital diagram data is the representation after electrocardio graphic sequence are as follows:
S=[s1,s2,…,si,…,sn]
In formula, S is n-dimensional vector, i=1,2 ..., n, siIndicate i-th point in sequence of data, the data sequence is defeated
Enter the result data format obtained to presetting in trained depth model are as follows:
Y=[y1,y2,…,yj,…,yN]
In formula, Y is N-dimensional vector, and N indicates the number of labels of category of model;J=1,2 ..., N, yjIndicate model in label
Classification value on j, 0≤yj≤ 1, wherein yjCorresponding label is the prediction classification results of depth model when being maximized, will
The corresponding y of the labeljValue is set to a reference value O, and label sequence number is set as I, the expression formula of a reference value O are as follows:
O=max { y1,y2,…,yj,…,yN}
Y in formulajIndicate classification value of the model on label j, 0≤yj≤1。
Step 2, influence of the different heartbeat sections for depth model output result is macroscopically shown.
Using eartbeat interval as basic unit, section is blocked according to the heartbeat message dynamic adjustment in ECG data, and
Calculate impact factor of the heartbeat section for final mask each time.Then use gradient color band by this influence visable representation
Out.
Step 2 specifically includes the following steps:
Step 2.1, it is dynamically determined and blocks siding-to-siding block length.
From original electrocardiographicdigital diagram data, the peak position the R label of each available heartbeat is considered between two peaks R
The section RR of heartbeat.Therefore kth time heartbeat is arranged blocks siding-to-siding block length are as follows:
Lengthk=xk+1-xk
In formula, LengthkIndicate the length for blocking section being arranged on k-th of section RR, xkIndicate k-th of peak position R
Abscissa, 0≤xkThe total length of≤Len, Len expression electrocardio graphic sequence;
Step 2.2, influence of each heartbeat section for depth model output result is calculated.
From the length for obtaining kth time heartbeat section in step 2.1, next set according to the length in heartbeat section dynamic
It sets and blocks section;Specific steps include:
Step 2.2.1 will block the R peak position alignment of section starting position and kth time heartbeat, block siding-to-siding block length setting
For Lengthk, so that block section covers kth time heartbeat RR block information just;
The vector value blocked in section is uniformly assigned a value of 0 by step 2.2.2, and the vector value of remaining position remains unchanged, and repairs
Electrocardio graphic sequence after changing are as follows:
Sk=[s1,s2,…,0,…,0,…,sn]
Wherein, siIt indicates i-th point of data in sequence, is assigned a value of 0 region since the peak R of kth time heartbeat, length
For Lengthk;
Step 2.2.3, we are provided on kth time heartbeat section and block section in step 2.2.2, will add now
Block the electrocardio graphic sequence S in sectionkVector is input in depth model, obtains new depth model output result Yk, YkIt is N
Dimensional vector, expression formula are as follows:
Yk=[y '1,y′2,…,y′N]
In formula, y '1,y′2,…,y′N1,2 are illustrated respectively in ..., the output valve on N label;
Step 2.2.4 calculates the depth model result O for obtaining and blocking kth time heartbeat block informationkWith the difference of benchmark result
It is worth Δ Ok;ΔOkFor the impact factor in k-th of heartbeat section, expression formula are as follows:
ΔOk=yI-y′I
In formula, I indicates the label sequence number for a reference value O being calculated in step 1, yIWith y 'IIt indicates in the label sequence number
Depth model output valve;ΔOkIndicate k-th of heartbeat section for the impact factor of depth model output result;ΔOk> 0 table
Show that the heartbeat section has positive influences to category of model result, be the supporting evidence of model, the value is bigger, indicates and model point
Class result is more agreed with;ΔOk< 0 indicates that the heartbeat section has negative effect to final classification result, is the opposition evidence of model,
The value is negative value, and the smaller expression of value more deviates from category of model result;
By the numerical value of Δ O, influence of the different heartbeat sections for category of model result is distinguished, is realized to category of model knot
The explanation of fruit.
Step 2.2.5, it is mobile to block section, above procedure is repeated, until influence of all heartbeat sections for result
The factor has all calculated completion.
It is the information of the heartbeat of erasing that the purpose for blocking section, which is arranged, the result of model and includes this when by without the heartbeat
The result of model compares when heartbeat, and the heartbeat can be calculated for the impact factor of model.The operation is executed repeatedly, i.e.,
Each heartbeat can be obtained for the impact factor of model result.
Step 2.3, the impact factor of each heartbeat is visualized.
In step 2.2, obtained difference DELTA O can be used to indicate the heartbeat section for the shadow of model final result
It rings.But ECG data is tediously long, includes multiple heartbeat sections, it is not intuitive enough using numerical approach, therefore also need to design phase
The method for visualizing answered.By the way that numerical value to be mapped to the color of rectangle, each heartbeat can be intuitively shown in ECG data
The performance in section.
The specific method is as follows for step 2.3:
(1) Δ O is encoded to color;Each heartbeat section corresponds to a Δ O value, can be obtained one after coding
Colour sequential
In order to intuitively show the meaning of Δ O, it is encoded to color in the present invention, rule are as follows:
As Δ O > 0, it is encoded to red, the value is bigger, then red depth is deeper;
As Δ O < 0, it is encoded to blue, the value is smaller, then blue depth is deeper.
(2) gradual change rectangle is generated
It, can be with using the height at the peak highest R on electrocardiogram as rectangle length using each heartbeat siding-to-siding block length as rectangle width
ECG data sequence is divided into several rectangles, each rectangle includes a heartbeat section.The heartbeat Interval Coding is generated
Color filling is into rectangle.In order not to block ECG information, rectangular centre is set as transparent, and both ends are set as Fill Color,
Rectangle is adjusted to gradient color band.
(3) rectangle is added to electrocardiogram background
The corresponding gradual change rectangle in each heartbeat section is added in the background of electrocardiogram, that is, produces effect of visualization.
The supporting evidence of model is obtained by the color checked on each heartbeat section and opposes evidence, passes through color depth
Degree, it can be determined that influence intensity of the evidence to final classification result;This method through the invention, the classification results of model can
It is explained in heartbeat level.
Step 3, influence of the microcosmic upper details for showing ECG data for model result.
In step 2, we have found influence of the different heartbeats for model result, primary explanation using heartbeat as interval
The classification results of model.But certain details within eartbeat interval have important influence similarly for category of model result, such as
Fruit is untreated, then is easy to cause details to lack.Therefore it also needs to carry out visualization exhibition to the details in electrocardiogram sequence data
Show, the foundation of category of model is explained in greater detail, reinforces the interpretation of model.
Step 3 specifically includes the following steps:
Step 3.1, setting is removable blocks section.
Due to needing to calculate point-by-point impact factor, section is blocked since first point, moves backward one every time
Lattice.The range for blocking the length L in section is 10≤L≤20, takes L=15 in this patent.This is carried out to many experiments result
The empirical value compared.Because blocking, section is too short to will lead to model output result difference very little, can not embody independent one
Influence of a point for whole result;It is too long, the influence of each point can be obscured.The present invention entirely will block section to model knot
Influences that fruit generates is considered as the impact factor of first point in section, in this way by move point by point block section can be obtained it is each
Individually influence of the point for model result.
Step 3.2, node-by-node algorithm difference.
After step 3.1, the length for blocking section has determined.It is following then point-by-point using interval computation is blocked
Difference, specific step is as follows for step 3.2:
Step 3.2.1 will be set to 0 by L vector Value Data later since first data of electrocardio graphic sequence S vector,
The vector value of remaining position remains unchanged, and section is blocked in formation;Section is blocked since first data, moves backward one every time
Lattice, until all data in traversal ECG data;
The electrocardio graphic sequence S for blocking section is added to when the m times circulationmVector data expression formula are as follows:
Sm=[s1,s2,…,sm-1,0,0,…,0,sm+L,…,sn]
Wherein s1,s2,…,snThe individual data for indicating composition electrocardio graphic sequence, by the formula it is found that sm,sm+1,…,
sm+L-1It is added to block section, the data in section are all assigned 0;
Step 3.2.2 is added to the electrocardio graphic sequence S for blocking section when by the m times circulationmVector is input to depth model
In, obtain the output result Y of modelm, expression formula are as follows:
Ym=[y '1,y′2,…,y′N]
In formula, y '1,y′2..., y 'N1,2 are illustrated respectively in ..., the output valve on N label;
Step 3.2.3 calculates the difference between the new model output result and model reference result obtained in step 3.2.2
It is worth Δ Om, value reflection individually influence of the point for model output result, calculation formula are as follows:
ΔOm=yI-y′I
In formula, I indicates the label sequence number for a reference value O being calculated in step 1, yIWith y 'IIt indicates in the label sequence number
Depth model output valve;ΔOmIndicate that than the m-th data is for the impact factor of depth model output result in heart sequence;Δ
Om> 0 indicates that the point has positive influences to final classification result, is the supporting evidence of model, and the value is bigger, indicates with model most
Termination fruit more agrees with;ΔOm< 0 indicates that the point has negative effect to final classification result, is the opposition evidence of model, which is
Negative value, the smaller expression of value more deviate from final result;
By the numerical value of Δ O, obtains each pair of point on electrocardiogram and realize electrocardiogram in the impact factor of category of model result
The explanation of detailed information in data.
Step 3.2.4 will block section and move backward a lattice, repeats above procedure, and to the last a point, which calculates, completes.
It may finally obtain the Δ O numerical value that each is put on electrocardiogram.
Step 3.3, point-by-point contribution is visualized.
In step 3.2, by being calculated the Δ O numerical value of every bit, the value can reflect out individually point for
The influence of the last classification results of model.But the numerical value for being to look at each point is non-intuitive, therefore also needs to design for point-by-point
Method for visualizing.Point value is different from heartbeat section numerical value, and individually point is difficult to find out its color, therefore cannot use upper
The method for visualizing of one link, it is necessary to be shown for the characteristics of point-by-point data.
Specific step is as follows for step 3.3:
The Δ O numeric coding of each point is height by step 3.3.1.
By step 3.2, in ECG data sequence, each data has corresponded to Δ O numerical value, further by Δ O number
Value is encoded to height, and a point P on electrocardio plan is determined by the abscissa of the data and by the height that Δ O is encoded:
Δ O > 0, indicate point P in the upper area of electrocardiogram, and it is corresponding points on electrocardiogram are shown in red;Δ O=0 indicates that point P is fallen
Black is shown as in zero axle, and by corresponding points on electrocardiogram;Δ O < 0 indicates point P in the lower zone of electrocardiogram, and by the heart
Corresponding points are shown as blue on electrograph.
Each point has divided color in this way on electrocardiogram, shows their contributions for category of model result.Together
When ECG data sequence in each data corresponded to the point P generated by Δ O.
Step 3.3.2 uses the corresponding point P of data each in smoothed curve connection ECG data sequence.
Since point P is excessively dense, it can not intuitively reflect its information by color, height, it is therefore desirable to using smooth bent
Line connects point P, and surrounds out several regions jointly with zero axle.The height of curve reflects the size of Δ O absolute value, bent
The spike and low ebb of line reflect support model result and violate the crucial foundation of model result.
Step 3.3.3, the region surrounded using color filling curve.
In order to keep the information in local detail region more intuitive, the Fill Color in several regions that step 3.3.2 is formed,
Keep its attribute more obvious.Area filling above zero axle is red, represents the final classification knot of the regional area support model
Fruit;Area filling blue below zero axle, represents the final classification result that the regional area violates model.Original electrocardiographicdigital figure is bent
Line has been divided into several paragraphs, is indicated respectively using different colours.Meanwhile it can be with according to the filling region near zero axle
Solve electrocardiogram local detail information, region is bigger, spike is higher, and a possibility that being abnormal is bigger, represent the region for
The formation of model final result influences bigger.Category of model is further illustrated for the visual presentation of ECG data details
As a result formation foundation, enhances the interpretation of model.
To sum up, the present invention provides a kind of depth model classification results method for visualizing towards ECG data, for solving
Certainly existing depth model result simple abstract, the insufficient defect of interpretation are main to block interval computation given zone by setting
Influence of the domain for final result, and separately design scheme from the angle of both macro and micro and visualize out by the influence.
Compared with prior art, invention enhances the interpretations of model result;Conventional method drag result is one specific
Classification results label has no idea to explain that the foundation for obtaining the result, such result are difficult to be adopted by doctor in medical field.
This method is made that explanation for model result, has found supporting evidence and oppose evidence that model is obtained a result, shows every
One details obtains model the influence of final result, greatly improves the interpretation of model result;The present invention is from macroscopic view
Visualized with microcosmic angle to interpretation process: ECG data is tediously long in a jumble under conventional method, therefrom differentiates
Key message is a time-consuming and laborious job.It is likely to be critical exceptions area to the region that model result is affected
Domain, for example there are the abnormal phenomenon such as P wave disappearance.This method excavates such region from ECG data, and from macroscopic view and micro-
It sees and shows it by visualized elements such as color, height in two kinds of granularities, to keep model running process more intuitive, mention
The interpretation of model result is risen;Method of the invention is suitable for various models, and scalability is strong: explaining mould under conventional method
Type result needs to refer to model structure, can not expand on other models.This method is not rely on particular model, all to be applicable in
This method can be used in the depth model classification results of ECG data to explain and show, and can easily expand to
On the improved model to emerge one after another at present.
Embodiment
Referring to Fig. 1, in order to realize final effect of visualization, method for visualizing of the invention the following steps are included:
S101 determines benchmark result.
In the present embodiment, the processing of original electrocardiographicdigital diagram data is the representation after electrocardio graphic sequence are as follows:
S=[s1, s2..., si..., sn]
In formula, S is n-dimensional vector, i=1,2 ..., n, siIndicate i-th point in sequence of data, the data sequence is defeated
Enter the result data format obtained to presetting in trained depth model are as follows:
Y=[y1, y2..., yj,…,yN]
In formula, Y is N-dimensional vector, and N indicates the number of labels of category of model;J=1,2 ..., N, yjIndicate model in label
Classification value on j, 0≤yj≤ 1, wherein yjCorresponding label is the prediction classification results of depth model when being maximized, will
The corresponding y of the labeljValue is set to a reference value O, and label sequence number is set as I, the expression formula of a reference value O are as follows:
O=max { y1,y2,…,yj,…,yN}
Y in formulajIndicate classification value of the model on label j, 0≤yj≤1;
S102, the method for visualizing that design heartbeat section influences model result.
Referring to Fig. 2, the method for visualizing that design heartbeat section influences model result, specific steps include:
1) it is dynamically determined and blocks siding-to-siding block length.
From original electrocardiographicdigital diagram data, the peak position the R label of each available heartbeat is considered between two peaks R
The section RR of heartbeat.Therefore kth time heartbeat is arranged blocks siding-to-siding block length are as follows:
Lengthk=xk+1-xk
In formula, LengthkIndicate the length for blocking section being arranged on k-th of section RR, xkIndicate k-th of peak position R
Abscissa, 0≤xkThe total length of≤Len, Len expression electrocardio graphic sequence;
2) influence of each heartbeat for model result is calculated.
We obtain the length in kth time heartbeat section from previous step, next need to be blocked according to length setting
Section.
S1 will block the R peak position alignment of section starting position and kth time heartbeat, and siding-to-siding block length is set as Lengthk,
So that block section covers kth time heartbeat just.
The vector value blocked in section is uniformly assigned a value of 0 by S2, and the vector value of remaining position remains unchanged, modified
Electrocardio graphic sequence are as follows:
Sk=[s1,s2,…,0,…,0,…,sn]
Wherein, siIt indicates i-th point of data in sequence, is assigned a value of 0 region since the peak R of kth time heartbeat, length
For Lengthk;
S3, we are provided on kth time heartbeat section and block section in S2, will be added to the heart for blocking section now
Electrograph sequence SkVector is input in depth model, obtains new depth model output result Yk, YkIt is N-dimensional vector, expression formula
Are as follows:
Yk=[y '1,y′2,…,y′N]
In formula, y '1,y′2,…,y′N1,2 are illustrated respectively in ..., the output valve on N label;
S4 calculates the depth model result O for obtaining and blocking kth time heartbeat block informationkWith the difference DELTA O of benchmark resultk;
ΔOkFor the impact factor in k-th of heartbeat section, expression formula are as follows:
ΔOk=yI-y′I
In formula, I indicates the label sequence number for a reference value O being calculated in step 1, yIWith y 'IIt indicates in the label sequence number
Depth model output valve;ΔOkIndicate k-th of heartbeat section for the impact factor of depth model output result;ΔOk> 0 table
Show that the heartbeat section has positive influences to category of model result, be the supporting evidence of model, the value is bigger, indicates and model point
Class result is more agreed with;ΔOk< 0 indicates that the heartbeat section has negative effect to final classification result, is the opposition evidence of model,
The value is negative value, and the smaller expression of value more deviates from category of model result;
By the numerical value of Δ O, influence of the different heartbeat sections for category of model result is distinguished, is realized to category of model knot
The explanation of fruit.
S5, it is mobile to block section, above procedure is repeated, influence of+1 heartbeat of kth for result is calculated.
It is the information of the heartbeat of erasing that the purpose in section is blocked in setting in the embodiment of the present invention, model when will be without the heartbeat
Result with comprising the heartbeat when model result compared with, influence numerical value of the heartbeat for model can be calculated.Instead
The operation is executed again, and influence numerical value of each heartbeat for model result can be obtained.
3) influence of each heartbeat is visualized.
In step 2), obtained difference DELTA O can be used to indicate the influence of the heartbeat section for model final result.
But ECG data is very long, includes multiple heartbeat sections, it is not intuitive enough using numerical approach, therefore also need to design corresponding
Method for visualizing.By the way that numerical value to be mapped to the color of rectangle, each heartbeat section can be intuitively shown in ECG data
Performance.The specific method is as follows:
Δ O is encoded to color by S1.In order to intuitively show the meaning of Δ O, it can be encoded to color, rule are as follows:
As Δ O > 0, it is encoded to red, the value is bigger, then red depth is deeper;As Δ O < 0, it is encoded to blue, it should
It is worth smaller, then blue depth is deeper.Each heartbeat section corresponds to a Δ O value, and a color can be obtained after coding
Sequence.
S2 generates gradual change rectangle.
It, can be with using the height at the peak highest R on electrocardiogram as rectangle length using each heartbeat siding-to-siding block length as rectangle width
Electrocardiogram is divided into several rectangles, each rectangle includes a heartbeat section.The color filling that the heartbeat Interval Coding is generated
Into rectangle.
Meanwhile in order not to block ECG information, rectangular centre is set as transparent, and both ends are set as Fill Color, by square
Shape is adjusted to gradient color band.
Rectangle is added to electrocardiogram background by S3.
Finally the corresponding gradual change rectangle in each heartbeat section is added in the background of electrocardiogram, that is, produces visualization effect
Fruit.According to the color on each heartbeat section obtain model supporting evidence and oppose evidence, and to category of model result into
Row is explained.
In the present embodiment, we choose the practical ECG data donated by AliveCor come the implementation of illustration method
Journey.It should be pointed out that as an example, this example only lists a data slot to illustrate the implementation procedure of this method, actually
ECG data will be considerably beyond enumerating range.
In the embodiment of the present invention, ECG data segment are as follows:
S=[... 0bff 02ff fbfe f7fe f4fe f4fe f5fe f7fe f9fe fcfe 00ff 03ff
07ff 09ff 0bff 0dff...];
S is input in model by a reference value in order to obtain, category of model result are as follows:
Y=[0.1215,0.9877,0.1010];
It can be seen that, the corresponding classification value of AF label is maximum in all classification values from the classification results, is 0.9877,
Think that category of model result is AF, represents Atrial Fibrillation (auricular fibrillation).According to the definition of front, we
Available a reference value yI=0.9877;
Data in the section RR are changed to 0 by the label that two peaks electrocardiogram R are obtained from data, and section is blocked in formation.It repairs
Change rear S are as follows:
S=[... 0,000 0,000 0,000 0,000 0000 0000f5fe f7fe f9fe fcfe 00ff 03ff
07ff 09ff 0bff 0dff...]
It is re-entered into model, obtains new classification results are as follows:
Y=[0.2011,0.6856,0.1317]
The corresponding classification value y ' of label A F at this timeI=0.6856, impact factor Δ O=y can be obtained by formulaI-y′I=
0.3021。
Due to Δ O > 0, i.e., after blocking the heartbeat segment, the conspicuousness of category of model result declines, and thus we can be with
Think that supporting function is played for category of model result in the heartbeat section, is the positive foundation that model obtains the result.
Above procedure is repeated, to its impact factor of each heartbeat interval computation.Then it will affect factor value to be encoded to
Color generates gradual change rectangle and is added on electrocardiographic wave.
Referring to Fig. 4, finally obtained effect of visualization is as shown in figure 4, from the figure, it can be seen that be directed to each heartbeat
Section, the color of gradual change rectangle show influence of the section for model final result, and RED sector represents support model point
Class opposes category of model as a result, the depth of color then reflects the size of influence as a result, blue portion represents.The visualization result
Explain each effect of heartbeat section to model final classification result.
S103, the method for visualizing that design point by point influences model result.
Referring to Fig. 3, design method for visualizing implementing procedure such as Fig. 3 institute that individually point influences category of model result
Show, specific steps include:
1) setting is removable blocks section.
Due to needing to calculate point-by-point difference, section is blocked since first point, moves backward a lattice every time.It hides
Keep off the length L=15 in section.
2) node-by-node algorithm difference.
After first step, the length for blocking section has determined.It is following then using blocking interval computation
Point-by-point difference, the specific steps are as follows:
S1 will be set to 0 by L vector Value Data later, remaining position since first data of electrocardio graphic sequence S vector
The vector value set remains unchanged, and section is blocked in formation;Section is blocked since first data, moves backward a lattice every time, directly
All data into traversal ECG data;
The electrocardio graphic sequence S for blocking section is added to when the m times circulationmVector data expression formula are as follows:
Sm=[s1,s2,…,sm-1,0,0,…,0,sm+L,…,sn]
Wherein s1,s2,…,snThe individual data for indicating composition electrocardio graphic sequence, by the formula it is found that sm,sm+1,…,
sm+L-1It is added to block section, the data in section are all assigned 0;
S2 is added to the electrocardio graphic sequence S for blocking section when by the m times circulationmVector is input in depth model, is obtained
The output result Y of modelm, expression formula are as follows:
Ym=[y '1,y′2,…,y′N]
In formula, y '1,y′2..., y 'N1,2 are illustrated respectively in ..., the output valve on N label;
S3 calculates the difference DELTA O between the new model output result and model reference result obtained in step S2m, should
Value reflection individually influence of the point for model output result, calculation formula are as follows:
ΔOm=yI-y′I
In formula, I indicates the label sequence number for a reference value O being calculated in step 1, yIWith y 'IIt indicates in the label sequence number
Depth model output valve;ΔOmIndicate that than the m-th data is for the impact factor of depth model output result in heart sequence;Δ
Om> 0 indicates that the point has positive influences to final classification result, is the supporting evidence of model, and the value is bigger, indicates with model most
Termination fruit more agrees with;ΔOm< 0 indicates that the point has negative effect to final classification result, is the opposition evidence of model, which is
Negative value, the smaller expression of value more deviate from final result;
By the numerical value of Δ O, obtains each pair of point on electrocardiogram and realize electrocardiogram in the impact factor of category of model result
The explanation of detailed information in data.
S4 will block section and move backward a lattice, repeats above procedure, each point on electrocardiogram is finally calculated
Δ O numerical value.
3) point-by-point contribution is visualized.
In previous step, through being calculated each point Δ O numerical value, the value can reflect out individually point for
The influence of the last classification results of model.But the numerical value for being to look at each point is non-intuitive, therefore also needs to design for point-by-point
Method for visualizing.Point value is different from heartbeat section numerical value, and individually point is difficult to find out its color, therefore cannot use upper
The method for visualizing of one link, it is necessary to be shown for the characteristics of point-by-point data.Specific step is as follows:
The Δ O numeric coding of each point is height by S1.
It has passed through previous step, in ECG data sequence, each data has corresponded to Δ O numerical value, further will
Δ O numeric coding is height, and determines by the abscissa of the data and by the height that Δ O is encoded one on electrocardio plan
Point P: Δ O > 0, indicate point P in the upper area of electrocardiogram, and it is corresponding points on electrocardiogram are shown in red;Δ O=0 is indicated
Point P is fallen in zero axle, and corresponding points on electrocardiogram are shown as black;Δ O < 0, indicate point P electrocardiogram lower zone, and
Corresponding points on electrocardiogram are shown as blue.
Each point has divided color in this way on electrocardiogram, shows their contributions for category of model result.Together
When ECG data sequence in each data corresponded to the point P generated by Δ O.
S2 uses the corresponding point P of data each in smoothed curve connection ECG data sequence.
It is excessively dense due to putting, it can not intuitively reflect its information by color, height, it is therefore desirable to use smoothed curve
Point P is connected, and surrounds out several regions jointly with zero axle.The height of curve reflects the size of Δ O absolute value, curve
Spike and low ebb reflect and support model result and violate the crucial foundation of model result.
S3, the region surrounded using color filling curve.
In order to keep the information in local detail region more intuitive, the Fill Color in several regions that previous step is formed makes
Its attribute is more obvious.Area filling above zero axle is red, represents the final classification result of the regional area support model;
Area filling blue below zero axle, represents the final classification result that the regional area violates model.Original electrocardiographicdigital figure curve
Several paragraphs have been divided into, have been indicated respectively using different colours.Meanwhile it will be seen that according to the filling region near zero axle
A possibility that electrocardiogram local detail information, region is bigger, spike is higher, is abnormal is bigger, for model final result
Formation influence it is bigger.For the visual presentation of ECG data details further illustrate the formation of category of model result according to
According to enhancing the interpretation of model.
In the present embodiment, as an example, still continuing exemplary thinking in S102.The difference is that in S102,
Each influence of heartbeat section for classification results in order to obtain, the length and moving distance of shielding window are according to heartbeat section
Length dynamic adjust, effect is just to block a heartbeat section.In this step, in order to obtain each point for minute
The influence of class result is blocked section and is needed using regular length, while moving backward a point every time, until the shadow of all points
The factor is rung all to have calculated.
For example, a true ECG data segment are as follows:
S=[... 2cff 2dff 2eff 2fff 32ff 35ff 37ff 3aff...];
It is 15 that siding-to-siding block length is blocked in setting, moving distance 1, and for the impact factor for finding out first point, we can be from
The point, which starts setting up, blocks section:
S=[... 0,000 0,000 0000 000f 32ff 35ff 37ff 3aff...]
The data for blocking section will be provided with to be input in model, according to model, new output result obtains new classification value
y′I=0.9903.
The impact factor Δ O=y of first point can be obtained by formulaI-y′I=-0.0026.
Because of Δ O < 0, one can consider that the point opposes category of model as a result, obtaining the classification results for model
Negative foundation.
It is constant that siding-to-siding block length is blocked later, moves backward a point:
S=[... 2,000 0,000 0,000 0000 32ff 35ff 37ff 3aff...];
Above procedure is repeated, the impact factor of each point can be obtained.
Then the point on electrocardio plan is generated according to method for visualizing, is connected with curve, the curve and zero axle can
To form several enclosing regions.Enclosing region is filled corresponding color by positive and negative according to impact factor.
Referring to Fig. 5, finally formed effect of visualization is as shown in Figure 5.From the figure, it can be seen that the original wave of electrocardiogram
Shape is divided into red, blue, black three kinds of colors, represents influence of this section of waveform for model final classification result.Meanwhile by influence because
The point that son determines is connected as a curve, the curve and zero axle surround jointly and form several regions, these region interpretations model
Obtain the detailed foundation of final classification result.For example, upward spike region occurs in the part outlined in Fig. 5, prompt in the area
There may be exceptions in domain.According to medical knowledge, be in fact here occurred P wave disappear it is abnormal, just because of being concerned about this
Details is so model has finally obtained the classification results of auricular fibrillation (AF).It is difficult to intuitively find the area in ordinary electrocardiogram
Domain, but then occur strong spike herein by the method for visualizing in the present invention, illustrate compared with the ordinary method, this method
It can more intuitively make explanations to depth model classification results, that is, improve the interpretation of depth model classification results.
S104 forms final visualization result.
It is overlapped, is finally established from macroscopic view to thin by above step, and by macroscopic view and details effect of visualization
The integrated visualization effect of section.Macro-effect is as shown in figure 4, details effect is as shown in Figure 5.Effect of visualization completely explains mould
Type classification results, protrusion illustrate the crucial foundation that model makes classification results, strengthen the interpretation of category of model result.
In practical applications, doctor can be likely to occur abnormal heartbeat section according to macroscopic information determination, navigate to rapidly specific
Details is checked in heartbeat section;The abnormal phenomenon being likely to occur can also be judged according to detailed information, search out from waveform details
Key message, to improve diagnosis efficiency.
To sum up, the invention discloses a kind of depth model classification results method for visualizing towards ECG data, including
Original electrocardiographicdigital diagram data: being input in model by following steps first, obtains the original output data of model, analyzes final prediction
As a result, and participating in subsequent comparison on the basis of saving initial data;Then section is blocked according to the dynamic setting of heartbeat section, obtained
Each heartbeat and will affect the factor and be encoded to color for the impact factor of final classification result out, and it is folded to generate gradual change rectangle
It is added in original electrocardiographicdigital figure information, intuitively shows each heartbeat section for model final classification result with visualization method
Influence.Next it resets to move and blocks interval parameter, section is blocked in movement, calculates the deviation of each point and benchmark
Value, which is superimposed with initial data, and the minutia of electrocardiogram is shown by peak value and region area, shows tiny area pair
In the influence of category of model result, the crucial foundation that model obtains the result is disclosed.The present invention is by showing macroscopic view and details two
Influence of the ECG data for final mask classification results under kind granularity, explains category of model result, illustrates mould
Type obtains the key evidence of final result, solves the problems, such as that model result interpretation is insufficient;Visual presentation method simultaneously
The key message in ECG data is deeply excavated, model running process is intuitively showed, depth mould is further improved
The interpretation of type classification results.
The above embodiments are merely illustrative of the technical scheme of the present invention and are not intended to be limiting thereof, although referring to above-described embodiment pair
The present invention is described in detail, those of ordinary skill in the art still can to a specific embodiment of the invention into
Row modification perhaps equivalent replacement these without departing from any modification of spirit and scope of the invention or equivalent replacement, applying
Within pending claims of the invention.
Claims (10)
1. a kind of depth model classification results method for visualizing towards ECG data, which comprises the following steps:
Step 1, the ECG data of acquisition is handled as electrocardio graphic sequence, by the trained depth model of electrocardiogram sequence inputting
In, obtain benchmark result;
Step 2, using eartbeat interval as basic unit, section is blocked according to the heartbeat message dynamic adjustment in ECG data,
By blocking the information in the heartbeat section for erasing selected in section, depth model when by without the heartbeat block information exports result
Compared with the benchmark result that depth model exports when with comprising the heartbeat message, calculates and obtain heartbeat each time for depth model
Impact factor Δ O;
Step 3, the impact factor Δ O visable representation of heartbeat each time is come out using gradient color band, realizes depth model point
The visualization of class result.
2. a kind of depth model classification results method for visualizing towards ECG data according to claim 1, special
Sign is, further includes:
Step 4, setting movably blocks section, successively blocks each point in ECG data;ECG data is blocked
The depth model output result of each point compared with the benchmark result of depth model output, obtains each on ECG data respectively
Impact factor of the point for depth model output result;
Step 5, the impact factor of each point step 4 obtained carries out visable representation.
3. a kind of depth model classification results method for visualizing towards ECG data according to claim 1, special
Sign is that step 2 specifically includes:
Step 2.1, according to original electrocardiographicdigital diagram data, the length in each heartbeat section is obtained, is blocked according to length dynamic setting
Each heartbeat section is successively blocked in section;
Step 2.2, the electrocardiogram sequence vector for blocking section will be added to be separately input in depth model, obtains new depth
Model exports result;
Step 2.3, each new the depth model output result and step 1 for calculating separately step 2.2 acquisition obtain benchmark result
Difference, obtain each heartbeat section to depth model output result impact factor.
4. a kind of depth model classification results method for visualizing towards ECG data according to claim 1, special
Sign is that step 3 specifically includes:
Step 3.1, the corresponding Δ O value in each heartbeat section is encoded, obtains a corresponding colour sequential;Rule are as follows: when
When Δ O > 0, it is encoded to a kind of pre-set color, the value is bigger, then color depth is deeper;As Δ O < 0, it is encoded to another
A kind of different pre-set color, the value is smaller, then color depth is deeper;
Step 3.2, using each heartbeat siding-to-siding block length as rectangle width, using the height at the peak highest R on electrocardiogram as rectangle length,
ECG data sequence is divided into several rectangles, each rectangle includes a heartbeat section;Each heartbeat that step 3.1 is obtained
The color filling that Interval Coding generates is into the corresponding rectangle in each heartbeat section;
Step 3.3, the corresponding coloured rectangle of filling in each heartbeat section that step 3.2 obtains is added to ECG data
In background, the visualization of depth model classification results is realized.
5. a kind of depth model classification results method for visualizing towards ECG data according to claim 4, special
Sign is, in step 3.2, rectangular centre is set as transparent, and both ends are set as Fill Color, and rectangle is adjusted to gradient color band.
6. a kind of depth model classification results method for visualizing towards ECG data according to claim 1, special
Sign is that step 1 specifically includes:
ECG data processing is the representation after electrocardio graphic sequence are as follows:
S=[s1,s2,…,si,…,sn]
In formula, S is n-dimensional vector, i=1,2 ..., n, siIndicate i-th point in sequence of data;
By electrocardiogram sequence inputting into trained depth model, obtained result data format are as follows:
Y=[y1,y2,…,yj,…,yN]
In formula, Y is N-dimensional vector, and N indicates the number of labels of category of model;J=1,2 ..., N, yjIndicate model on label j
Classification value, 0≤yj≤1;
Wherein, yjCorresponding label is the prediction classification results of depth model when being maximized, by the corresponding y of the labeljValue is fixed
On the basis of value O, label sequence number is set as I, the expression formula of a reference value O are as follows:
O=max { y1,y2,…,yj,…,yN}
In formula, yjIndicate classification value of the model on label j, 0≤yj≤1。
7. a kind of depth model classification results method for visualizing towards ECG data according to claim 6, special
Sign is,
Step 2.1, it is dynamically determined and blocks siding-to-siding block length;
From original electrocardiographicdigital diagram data, the peak position the R label of each heartbeat is obtained, is considered a heartbeat between two peaks R
The section RR;It is arranged k-th and blocks siding-to-siding block length are as follows:
Lengthk=xk+1-xk
In formula, LengthkIndicate the length for blocking section being arranged on k-th of section RR, xkIndicate the horizontal seat of k-th of peak position R
Mark, 0≤xkThe total length of≤Len, Len expression electrocardio graphic sequence;
Step 2.2, each heartbeat block information is calculated for the impact factor of depth model output result;
Step 2.2.1 will block the R peak position alignment of section starting position and kth time heartbeat, and siding-to-siding block length is set as
Lengthk, so that blocking section covering kth time heartbeat block information;
The vector value blocked in section is uniformly assigned a value of 0 by step 2.2.2, and the vector value of remaining position remains unchanged, after modification
Electrocardio graphic sequence are as follows:
Sk=[s1,s2,…,0,…,0,…,sn]
Wherein, siIt indicates i-th point of data in sequence, is assigned a value of 0 region since the peak R of kth time heartbeat, length is
Lengthk;
Step 2.2.3 will be added to the electrocardio graphic sequence S for blocking sectionkVector is input in depth model, obtains new depth
Model exports result Yk, YkIt is N-dimensional vector, expression formula are as follows:
Yk=[y '1,y′2,…,y′N]
In formula, y '1,y′2,…,y′N1,2 are illustrated respectively in ..., the output valve on N label;
Step 2.2.4 calculates the depth model result O for obtaining and blocking kth time heartbeat block informationkWith the difference DELTA of benchmark result
Ok;ΔOkFor the impact factor in k-th of heartbeat section, expression formula are as follows:
ΔOk=yI-y′I
In formula, I indicates the label sequence number for a reference value O that step 1 is calculated, yIWith y 'IIndicate the depth in the label sequence number
Model output value;ΔOkIndicate k-th of heartbeat section for the impact factor of depth model output result;ΔOk> 0 indicates the heart
Jumping section has positive influences to category of model result, is the supporting evidence of model, and the value is bigger, indicates and category of model result
More agree with;ΔOk< 0 indicates that the heartbeat section has negative effect to final classification result, is the opposition evidence of model, which is
Negative value, the smaller expression of value more deviate from category of model result;
By the numerical value of Δ O, influence of the different heartbeat sections for category of model result is distinguished, is realized to category of model result
It explains.
8. a kind of depth model classification results method for visualizing towards ECG data according to claim 2, special
Sign is that step 4 specifically includes:
Step 4.1, since first data of electrocardio graphic sequence S vector, 0 will be set to by L vector Value Data later, remaining position
The vector value set remains unchanged, and section is blocked in formation;Section is blocked since first data, moves backward a lattice every time, directly
All data into traversal ECG data;
The electrocardio graphic sequence S for blocking section is added to when the m times circulationmVector data expression formula are as follows:
Sm=[s1,s2,…,sm-1,0,0,…,0,sm+L,…,sn]
Wherein s1,s2,…,snThe individual data for indicating composition electrocardio graphic sequence, by the formula it is found that sm,sm+1,…,sm+L-1Quilt
It is added to and blocks section, the data in section are all assigned 0;
Step 4.2, the difference of depth model output result and benchmark result behind section is blocked in node-by-node algorithm setting, obtains electrocardiogram
The impact factor Δ O numerical value of each upper point;
Specific steps include:
Step 4.2.1 is added to the electrocardio graphic sequence S for blocking section when by the m times circulationmVector is input in depth model, is obtained
To the output result Y of modelm, expression formula are as follows:
Ym=[y '1,y′2,…,y′N]
In formula, y '1,y′2,…,y′N1,2 are illustrated respectively in ..., the output valve on N label;
Step 4.2.2 calculates the difference DELTA O between the new model output result and model reference result that step 4.2.1 is obtainedm,
Value reflection individually influence of the point for model output result, calculation formula are as follows:
ΔOm=yI-y′I
In formula, I indicates the label sequence number for a reference value O being calculated, yIWith y 'IIndicate the depth model in the label sequence number
Output valve;ΔOmIndicate that than the m-th data is for the impact factor of depth model output result in heart sequence;ΔOm> 0 indicates to be somebody's turn to do
Point has positive influences to final classification result, is the supporting evidence of model, and the value is bigger, indicates to get over contract with model final result
It closes;ΔOm< 0 indicates that the point has negative effect to final classification result, is the opposition evidence of model, which is negative value, and value is got over
Small expression more deviates from final result;
By the numerical value of Δ O, obtains each pair of point on electrocardiogram and realize ECG data in the impact factor of category of model result
The explanation of middle detailed information.
9. a kind of depth model classification results method for visualizing towards ECG data according to claim 8, special
Sign is that step 5 specific steps include:
Step 5.1, the Δ O numeric coding of each point step 4.2 obtained is height, and true by the position and height of the point
A point P in electrograph plane of feeling relieved, Δ O > 0 indicate that point P in the upper area of electrocardiogram, and corresponding points on electrocardiogram is shown
It is shown as a kind of pre-set color;Δ O=0 indicates that point P is fallen in zero axle, and corresponding points on electrocardiogram are shown as another default face
Color;Δ O < 0 indicates that corresponding points on electrocardiogram in the lower zone of electrocardiogram, are shown as another pre-set color by point P;It is default
Color is all different;
Step 5.2, it will be connected using smoothed curve with the point that serial number abscissa, Δ O numerical value are ordinate formation, and with
Zero axle surrounds out several regions jointly;The size of the height reflection Δ O absolute value of curve, the spike and low ebb of curve, which reflect, to be supported
Model result and the crucial foundation for violating model result;
Step 5.3, depth model classification results are realized in the region surrounded using preset 5.2 curve of different colours filling step
Visualization.
10. a kind of depth model classification results method for visualizing towards ECG data according to claim 8, special
Sign is, in step 4, the range for blocking the length L in section is 10≤L≤20.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910067724.0A CN109875546B (en) | 2019-01-24 | 2019-01-24 | Depth model classification result visualization method for electrocardiogram data |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910067724.0A CN109875546B (en) | 2019-01-24 | 2019-01-24 | Depth model classification result visualization method for electrocardiogram data |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109875546A true CN109875546A (en) | 2019-06-14 |
CN109875546B CN109875546B (en) | 2020-07-28 |
Family
ID=66926716
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910067724.0A Active CN109875546B (en) | 2019-01-24 | 2019-01-24 | Depth model classification result visualization method for electrocardiogram data |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109875546B (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111161789A (en) * | 2019-12-11 | 2020-05-15 | 深圳先进技术研究院 | Analysis method and device for key region of model prediction |
CN112587148A (en) * | 2020-12-01 | 2021-04-02 | 上海数创医疗科技有限公司 | Template generation method and device comprising fuzzification similarity measurement method |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2004032741A1 (en) * | 2002-10-09 | 2004-04-22 | Bang & Olufsen Medicom A/S | A procedure for extracting information from a heart sound signal |
CN101263510A (en) * | 2004-11-08 | 2008-09-10 | 依德西亚有限公司 | Method and apparatus for electro-biometric identity recognition |
CN102542283A (en) * | 2010-12-31 | 2012-07-04 | 北京工业大学 | Optimal electrode assembly automatic selecting method of brain-machine interface |
CN105960200A (en) * | 2014-02-25 | 2016-09-21 | 圣犹达医疗用品心脏病学部门有限公司 | Systems and methods for using electrophysiology properties for classifying arrhythmia sources |
CN108478209A (en) * | 2018-02-24 | 2018-09-04 | 乐普(北京)医疗器械股份有限公司 | Ecg information dynamic monitor method and dynamic monitor system |
US20180249921A1 (en) * | 2015-09-10 | 2018-09-06 | Nihon Kohden Corporation | Electrocardiogram analyzing method, electrocardiogram analyzing apparatus, electrocardiogram analyzing program, and computer-readable medium stored with the electrocardiogram analyzing program |
CN108937912A (en) * | 2018-05-12 | 2018-12-07 | 鲁东大学 | A kind of automatic arrhythmia analysis method based on deep neural network |
-
2019
- 2019-01-24 CN CN201910067724.0A patent/CN109875546B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2004032741A1 (en) * | 2002-10-09 | 2004-04-22 | Bang & Olufsen Medicom A/S | A procedure for extracting information from a heart sound signal |
CN101263510A (en) * | 2004-11-08 | 2008-09-10 | 依德西亚有限公司 | Method and apparatus for electro-biometric identity recognition |
CN102542283A (en) * | 2010-12-31 | 2012-07-04 | 北京工业大学 | Optimal electrode assembly automatic selecting method of brain-machine interface |
CN105960200A (en) * | 2014-02-25 | 2016-09-21 | 圣犹达医疗用品心脏病学部门有限公司 | Systems and methods for using electrophysiology properties for classifying arrhythmia sources |
US20180249921A1 (en) * | 2015-09-10 | 2018-09-06 | Nihon Kohden Corporation | Electrocardiogram analyzing method, electrocardiogram analyzing apparatus, electrocardiogram analyzing program, and computer-readable medium stored with the electrocardiogram analyzing program |
CN108478209A (en) * | 2018-02-24 | 2018-09-04 | 乐普(北京)医疗器械股份有限公司 | Ecg information dynamic monitor method and dynamic monitor system |
CN108937912A (en) * | 2018-05-12 | 2018-12-07 | 鲁东大学 | A kind of automatic arrhythmia analysis method based on deep neural network |
Non-Patent Citations (2)
Title |
---|
ZIQI LIU ET AL.: "Nonparametric models for characterizing the topical communities in social network", 《NEUROCOMPUTING》 * |
郝亚洲等: "面向网络舆情数据的异常行为识别", 《计算机研究与发展》 * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111161789A (en) * | 2019-12-11 | 2020-05-15 | 深圳先进技术研究院 | Analysis method and device for key region of model prediction |
CN111161789B (en) * | 2019-12-11 | 2023-10-31 | 深圳先进技术研究院 | Analysis method and device for key areas of model prediction |
CN112587148A (en) * | 2020-12-01 | 2021-04-02 | 上海数创医疗科技有限公司 | Template generation method and device comprising fuzzification similarity measurement method |
CN112587148B (en) * | 2020-12-01 | 2023-02-17 | 上海数创医疗科技有限公司 | Template generation method and device comprising fuzzification similarity measurement method |
Also Published As
Publication number | Publication date |
---|---|
CN109875546B (en) | 2020-07-28 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN105263405B (en) | Multi-parameter physiology maps | |
JP5161217B2 (en) | How to reconstruct a voxel cluster map in an image | |
US20210272235A1 (en) | Two-dimensional scalar field data visualization method and system based on colormap optimization | |
CN107423700A (en) | The method and device of testimony verification | |
CN109875546A (en) | A kind of depth model classification results method for visualizing towards ECG data | |
CN102496023B (en) | Region of interest extraction method of pixel level | |
JP6288676B2 (en) | Visualization device, visualization method, and visualization program | |
CN106901718A (en) | Many active regions are shown on electro-anatomical map | |
CN104798101A (en) | Makeup support device, makeup support method, and makeup support program | |
CN102519395B (en) | Color response calibration method in colored structure light three-dimensional measurement | |
Fiorini et al. | Automatic Generation of Synthetic Retinal Fundus Images. | |
CN105787948A (en) | Quick graph cutting method based on multiple deformation resolutions | |
CN109948454A (en) | The Enhancement Method in expression data library, calculates equipment and storage medium at training method | |
CN108364356A (en) | A kind of automatic division method of tooth three-dimensional grid model | |
CN106716491A (en) | Image color calibration with multiple color scales | |
CN109242760A (en) | Processing method, device and the electronic equipment of facial image | |
CN115035058A (en) | Self-coding network medical image anomaly detection method | |
Lichtenberg et al. | Concentric Circle Glyphs for Enhanced Depth-Judgment in Vascular Models. | |
CN104899908B (en) | The method and apparatus for generating event group evolution diagram | |
CN109116988A (en) | Steady-state induced current potential brain-computer interface method based on apparent motion perception | |
CN107330948A (en) | A kind of fMRI data two-dimensional visualization methods based on popular learning algorithm | |
CN113298754B (en) | Method for detecting control points of outline of prostate tissue | |
CN105894561B (en) | Color Mapping Approach and system based on curvature distribution in a kind of Discrete Surfaces | |
CN115568941A (en) | Tumor ablation path evaluation method | |
CN110009082A (en) | Optimization method, medium, computer equipment and the device of three-dimension code |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |