CN109875546B - Depth model classification result visualization method for electrocardiogram data - Google Patents

Abstract

The invention discloses a deep model classification result visualization method oriented to electrocardiogram data, comprising: inputting electrocardiogram sequences into a trained depth model to obtain reference results; The depth model output results when determining the heartbeat interval information are compared with the benchmark results output by the depth model, and the impact factor ΔO of each heartbeat on the depth model is calculated and obtained; the impact factor ΔO of each heartbeat is visualized by using a gradient color band to achieve Visualization of deep model classification results. By analyzing the influence of electrocardiogram data in macroscopic and microscopic granularities on the output result of the depth model, the present invention can display the key evidence for obtaining the classification result of the model, and can enhance the interpretability of the classification result output by the model.

Description

A visualization method of deep model classification results for ECG data

technical field

The invention belongs to the technical field of deep model classification result visualization, and particularly relates to a deep model classification result visualization method oriented to electrocardiogram data.

Background technique

According to the definition of Wikipedia, ECG data refers to a time-based recording of the electrophysiological activity of the heart through the chest cavity, which is captured and recorded by electrodes on the skin. In practice, in order to improve efficiency and reduce the burden and work intensity of doctors, some deep learning-based models are applied to feature extraction and classification on ECG data. However, these existing models can only give the final classification result, and cannot explain the basis of the classification result; and the classification result prediction without a clear explanation in practice is difficult to be accepted and applied, which makes the application scenario greatly affected. It is not conducive to doctors using the classification results output by the model.

In conclusion, there is an urgent need for a visualization method of deep model classification results for ECG data.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a visualization method of deep model classification results oriented to electrocardiogram data, so as to solve the above-mentioned technical problems. The present invention can show the key evidence of obtaining the final result, and can enhance the interpretability of the classification result output by the model.

To achieve the above object, the present invention adopts the following technical solutions:

A deep model classification result visualization method for electrocardiogram data, comprising the following steps:

Step 1: Process the collected electrocardiogram data into an electrocardiogram sequence, and input the electrocardiogram sequence into the trained deep model to obtain a benchmark result;

Step 2, take the heartbeat interval as the basic unit, dynamically adjust the occlusion interval according to the heartbeat information in the electrocardiogram data, erase the information of the selected heartbeat interval through the occlusion interval, and output the result of the depth model without the heartbeat interval information and include the heartbeat interval information. Comparing the benchmark results output by the depth model when heartbeat information is used, the impact factor ΔO of each heartbeat on the depth model is calculated and obtained;

In step 3, the influence factor ΔO of each heartbeat is visualized by using a gradient color band, so as to realize the visualization of the classification result of the deep model.

Further, it also includes:

Step 4: Set a movable occlusion interval to occlude each point in the ECG data in turn; compare the depth model output result of each point occluded by the ECG data with the benchmark result output by the depth model, and obtain the corresponding value of each point on the ECG data. Influence factor of the output result of the deep model;

Step 5: Visually represent the impact factor of each point obtained in Step 4.

Further, step 2 specifically includes:

Step 2.1, according to the original electrocardiogram data, obtain the length of each heartbeat interval, dynamically set the blocking interval according to the length, and sequentially block each heartbeat interval;

Step 2.2, input the ECG sequence vector with the added occlusion interval into the depth model respectively, and obtain the output result of the new depth model;

Step 2.3: Calculate the difference between each new depth model output result obtained in step 2.2 and the reference result obtained in step 1, and obtain the influence factor of each heartbeat interval on the depth model output result.

Further, step 3 specifically includes:

Step 3.1, encode the ΔO value corresponding to each heartbeat interval to obtain a corresponding color sequence; the rule is: when ΔO>0, encode it as a preset color, the larger the value, the darker the color depth ; when ΔO<0, encode it as another different preset color, the smaller the value, the deeper the color depth;

Step 3.2, take the length of each heartbeat interval as the width of the rectangle, and take the height of the highest R peak on the electrocardiogram as the length of the rectangle, divide the electrocardiogram data sequence into several rectangles, and each rectangle contains a heartbeat interval; The color generated by the encoding is filled into the rectangle corresponding to each heartbeat interval;

In step 3.3, the colored rectangles corresponding to each heartbeat interval obtained in step 3.2 are superimposed on the background of the electrocardiogram data, so as to realize the visualization of the classification result of the depth model.

Further, in step 3.2, the center of the rectangle is set to transparent, the two ends are set to fill color, and the rectangle is adjusted to a gradient color band.

Further, step 1 specifically includes:

The representation of the ECG data after processing it into an ECG sequence is:

S=[s ₁ ,s ₂ ,…,s _i ,…,s _n ]

In the formula, S is an n-dimensional vector, i=1,2,...,n, s _i represents the data of the i-th point in the sequence;

Input the ECG sequence into the trained deep model, and the resulting data format is:

Y=[y ₁ ,y ₂ ,...,y _j ,...,y _N ]

In the formula, Y is an N-dimensional vector, N represents the number of labels of the model classification; j=1,2,...,N, y _j represents the classification value of the model on the label _j , 0≤yj≤1;

Among them, the label corresponding to the maximum value of y _j is the predicted classification result of the depth model, the value of y _j corresponding to the label is set as the reference value O, the label serial number is set as I, and the expression of the reference value O is:

O=max{y ₁ ,y ₂ ,…,y _j ,…,y _N }

In the formula, y _j represents the classification value of the model on the label _j , 0≤yj≤1.

Further, step 2.1, dynamically determine the length of the occlusion interval;

From the original ECG data, the R peak position label of each heartbeat is obtained, and the interval between two R peaks is considered to be the RR interval of a heartbeat; the length of the kth occlusion interval is set as:

Length _k = x _k+1 -x _k

In the formula, Length _k represents the length of the occlusion interval set on the kth RR interval, x _k represents the abscissa of the _kth R peak position, 0≤xk ≤Len, and Len represents the total length of the ECG sequence;

Step 2.2, calculate the influence factor of each heartbeat interval information on the output result of the depth model;

Step 2.2.1, align the start position of the occlusion interval with the position of the R peak of the kth heartbeat, and set the interval length to Length _k , so that the occlusion interval covers the information of the kth heartbeat interval;

Step 2.2.2, uniformly assign the vector values in the occlusion interval to 0, and keep the vector values in other positions unchanged. The modified ECG sequence is:

_Sk = [s ₁ , s ₂ , ..., 0, ..., 0, ..., s _n ]

Wherein, _si represents the data of the ith point in the sequence, and the area assigned as 0 starts from the R peak of the kth heartbeat, and the length is Length _k ;

Step 2.2.3, input the electrocardiogram sequence _Sk vector with the added occlusion interval into the depth model, and obtain the new depth model output result Y _k , where Y _k is an N-dimensional vector, and the expression is:

Y _k = [y' _1, y' ₂ , ..., y' _N ]

In the formula, y′ ₁ , y′ ₂ , …, y′ _N represent the output values on the labels 1, 2, …, N, respectively;

Step 2.2.4, calculate and obtain the difference ΔO _k between the depth model result O _k that blocks the information of the kth heartbeat interval and the reference result; ΔOk is the influence factor of the _kth heartbeat interval, and the expression is:

ΔO _k =y _I -y' _I

In the formula, I represents the label number of the reference value O calculated in step 1, y _I and y′ _I represent the depth model output value on the label number; ΔO _k represents the influence of the kth heartbeat interval on the output result of the depth model factor; ΔO _k >0 indicates that the heartbeat interval has a positive impact on the model classification result, which is the supporting evidence of the model. The larger the value, the better the fit with the model classification result; _ΔOk <0 indicates that the heartbeat interval has a positive effect on the final classification result. Negative influence is the evidence against the model, the value is negative, and the smaller the value, the greater the deviation from the model classification result;

Through the value of ΔO, the influence of different heartbeat intervals on the model classification results can be distinguished, and the interpretation of the model classification results can be realized.

Further, step 4 specifically includes:

Step 4.1, starting from the first data of the S vector of the electrocardiogram sequence, set the following L vector value data to 0, and the vector values of the remaining positions remain unchanged to form an occlusion interval; the occlusion interval starts from the first data, and each time Move backward one grid until all the data in the ECG data are traversed;

The vector data expression of the ECG sequence S _m with the occlusion interval added in the mth cycle is:

S _m =[s ₁ ,s ₂ ,...,s _m-1 ,0,0,...,0,s _m+L ,...,s _n ]

Among them, s ₁ , s ₂ ,…,s _n represent the single data constituting the ECG sequence. It can be seen from this formula that s _m , s _m+1 ,…, s _m+L-1 are added with occlusion intervals, and the data in the interval are assigned to 0;

Step 4.2: Calculate the difference between the output result of the depth model and the reference result after setting the occlusion interval point by point, and obtain the value of the influence factor ΔO of each point on the ECG;

Specific steps include:

Step 4.2.1, input the electrocardiogram sequence S _m vector with the occlusion interval added in the mth cycle into the depth model, and obtain the output result Y _m of the model, the expression is:

Y _m =[y' ₁ , y' ₂ , ..., y' _N ]

In the formula, y′ ₁ , y′ ₂ ,…,y′ _N represent the output values on the labels 1, 2,…,N, respectively;

Step 4.2.2, calculate the difference ΔO _m between the new model output result obtained in step 4.2.1 and the model benchmark result, which reflects the influence of a single point on the model output result. The calculation formula is:

ΔO _m =y _I -y' _I

In the formula, I represents the label number of the calculated reference value O, y _I and y′ _I represent the depth model output value on the label number; ΔO _m represents the influence of the mth data in the heartbeat sequence on the output result of the depth model factor; ΔO _m >0 indicates that the point has a positive impact on the final classification result, which is the supporting evidence for the model. The larger the value, the better the final result of the model; ΔO _m <0 indicates that the point has a negative impact on the final classification result. , is the evidence against the model, the value is negative, the smaller the value, the more deviation from the final result;

Through the value of ΔO, the influence factor of each point on the ECG on the model classification result is obtained, and the detailed information in the ECG data can be explained.

Further, the specific steps of step 5 include:

Step 5.1, encode the ΔO value of each point obtained in step 4.2 as the height, and determine a point P on the ECG plane by the position and height of the point, ΔO>0, indicating that point P is in the upper area of the ECG, and the The corresponding point on the ECG is displayed as a preset color; ΔO=0, it means that the point P falls on the zero axis, and the corresponding point on the ECG is displayed as another preset color; ΔO<0, it means that the point P is below the ECG area, and display the corresponding point on the ECG as another preset color; the preset colors are all different;

Step 5.2, use a smooth curve to connect the points formed with the serial number as the abscissa and the ΔO value as the ordinate, and together with the zero axis to enclose several areas; the height of the curve reflects the absolute value of ΔO, and the peaks and valleys of the curve reflect Key evidence supporting and violating model results;

In step 5.3, use preset different colors to fill the area surrounded by the curve in step 5.2 to realize the visualization of the classification result of the deep model.

Further, in step 4, the range of the length L of the occlusion interval is 10≤L≤20.

Compared with the prior art, the present invention has the following beneficial effects:

The visualization method of the deep model classification result for electrocardiogram data of the present invention designs a visualization result display process from the overall situation to the details, and can completely display the key basis affecting the results obtained by the model. The method of the present invention firstly inputs the collected original electrocardiogram data into the depth model, obtains the output data of the depth model, determines the predicted classification result according to the output data analysis, saves the output data as the benchmark result and participates in the subsequent comparison, and obtains Influence factor; then dynamically set the parameters of the occlusion interval combined with the heartbeat information, obtain the influence of each heartbeat interval on the prediction of the final result of the model, and visualize it in a visual way; further design a movable occlusion interval, and calculate the difference between each point and the benchmark. The deviation value is superimposed with the original ECG data, and the detailed features in the ECG data are displayed through the peak value and area area, which is convenient for finding abnormal detailed areas. The invention calculates the influence of a specific area on the final result by setting the occlusion interval, and analyzes the influence of the electrocardiogram data in macroscopic and microscopic granularities on the final model result, which can display the key evidence that the model obtains the final result, and can enhance the interpretability of the model result. sex.

The visualization method of the present invention can enhance the interpretability of the model result; under the traditional method, the model result is a specific classification result label, and there is no way to explain the basis for the result, and such a result is difficult to be adopted and used. The method of the invention explains the model results, finds the supporting evidence and the opposing evidence for the model results, and shows the influence of every detail on the model results, which can greatly improve the interpretability of the model results.

The present invention visualizes the interpretation process from the macroscopic and microscopic perspectives; under the traditional method, the electrocardiogram data is cluttered and lengthy, and it is time-consuming and laborious to distinguish key information from it. The areas that have a greater impact on the model results are likely to be critical abnormal areas, such as the presence of abnormal phenomena such as the disappearance of P waves. The method of the present invention excavates such regions from the electrocardiogram data, and displays them through visual elements such as color and height from both macroscopic and microscopic granularities, thereby making the model running process more intuitive and further improving the interpretability of the model results. sex.

The method of the present invention is suitable for various deep learning models, and has strong expansibility; under the traditional method, the interpretation of the model results needs to refer to the model structure and cannot be extended to other models. The method of the present invention does not depend on a specific model, and all deep model classification results applicable to electrocardiogram data can be interpreted and displayed by this method, and can be easily extended to the endlessly emerging improved models.

Description of drawings

Fig. 1 is a kind of flow schematic block diagram of the visualization method of deep model classification result oriented to electrocardiogram data of the present invention;

Fig. 2 is a schematic flow diagram of a method for visualizing the influence of a heartbeat interval in a deep model classification result visualization method for electrocardiogram data of the present invention;

3 is a schematic flow diagram of a point-by-point influence visualization method in a deep model classification result visualization method oriented to electrocardiogram data of the present invention;

4 is a schematic diagram of the visualization result of a heartbeat interval in a method for visualizing the results of deep model classification for electrocardiogram data according to the present invention;

FIG. 5 is a schematic diagram of a point-by-point influence visualization result in a method for visualizing a result of deep model classification for electrocardiogram data according to the present invention.

Detailed ways

The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

A method for visualizing the results of deep model classification for electrocardiogram data of the present invention specifically includes the following steps:

Step 1: Collect and obtain a preset number of diagnosed electrocardiogram data, process each electrocardiogram data into an electrocardiogram sequence, and input each electrocardiogram sequence into the selected trained deep model to obtain the output result of the deep model. The output result is set as the benchmark result output by the deep model.

The representation of the raw ECG data after processing it into an ECG sequence is:

S=[s ₁ ,s ₂ ,…,s _i ,…,s _n ]

In the formula, S is an n-dimensional vector, i=1,2,...,n, s _i represents the data of the ith point in the sequence, input the data sequence into the preset trained depth model, and get the result data The format is:

Y=[y ₁ ,y ₂ ,...,y _j ,...,y _N ]

In the formula, Y is an N-dimensional vector, N represents the number of labels of the model classification; j=1,2,...,N, y _j represents the classification value of the model on the label j, 0≤y _j ≤1, where, The label corresponding to the maximum value of y _j is the predicted classification result of the deep model. The value of y _j corresponding to the label is set as the reference value O, the label serial number is set as I, and the expression of the reference value O is:

O=max{y ₁ ,y ₂ ,…,y _j ,…,y _N }

In the formula, y _j represents the classification value of the model on the label _j , 0≤yj≤1.

Step 2: Macroscopically display the impact of different heartbeat intervals on the output results of the depth model.

Taking the heartbeat interval as the basic unit, the occlusion interval is dynamically adjusted according to the heartbeat information in the ECG data, and the impact factor of each heartbeat interval on the final model is calculated. This effect is then visualized using a gradient color ramp.

Step 2 specifically includes the following steps:

Step 2.1, dynamically determine the length of the occlusion interval.

From the original ECG data, the R-peak position label of each heartbeat can be obtained, and the interval between two R-peaks is considered to be the RR interval of a heartbeat. Therefore, the length of the occlusion interval of the kth heartbeat is set as:

Length _k = x _k+1 -x _k

In the formula, Length _k represents the length of the occlusion interval set on the kth RR interval, x _k represents the abscissa of the _kth R peak position, 0≤xk ≤Len, and Len represents the total length of the ECG sequence;

Step 2.2: Calculate the impact of each heartbeat interval on the output of the deep model.

The length of the kth heartbeat interval is obtained from step 2.1, and then the occlusion interval is dynamically set according to the length of the heartbeat interval; the specific steps include:

Step 2.2.1, align the start position of the occlusion interval with the R peak position of the kth heartbeat, and set the length of the occlusion interval to Length _k , so that the occlusion interval just covers the information of the RR interval of the kth heartbeat;

Step 2.2.2, uniformly assign the vector values in the occlusion interval to 0, and keep the vector values in other positions unchanged. The modified ECG sequence is:

S _k =[s ₁ ,s ₂ ,…,0,…,0,…,s _n ]

Wherein, _si represents the data of the ith point in the sequence, and the area assigned as 0 starts from the R peak of the kth heartbeat, and the length is Length _k ;

Step 2.2.3, in step 2.2.2, we set the occlusion interval on the kth heartbeat interval, and now the ECG sequence S _k vector with the occlusion interval added is input into the depth model, and the new output result Y of the depth model is obtained. _k , Y _k is an N-dimensional vector whose expression is:

Y _k =[y′ ₁ ,y′ ₂ ,…,y′ _N ]

In the formula, y′ ₁ , y′ ₂ ,…,y′ _N represent the output values on the labels 1, 2,…,N, respectively;

Step 2.2.4, calculate and obtain the difference ΔO _k between the depth model result O _k that blocks the information of the kth heartbeat interval and the reference result; ΔOk is the influence factor of the _kth heartbeat interval, and the expression is:

ΔO _k =y _I -y' _I

In the formula, I represents the label number of the reference value O calculated in step 1, y _I and y′ _I represent the depth model output value on the label number; ΔO _k represents the kth heartbeat interval for the depth model output result. Impact factor; ΔO _k >0 indicates that the heartbeat interval has a positive impact on the model classification result, which is the supporting evidence of the model. The larger the value, the better the model classification result; ΔO _k <0 indicates that the heartbeat interval has a positive impact on the final classification result. It has a negative impact and is the evidence against the model. The value is negative, and the smaller the value, the greater the deviation from the model classification result;

Through the value of ΔO, the influence of different heartbeat intervals on the model classification results can be distinguished, and the interpretation of the model classification results can be realized.

Step 2.2.5, move the occlusion interval, and repeat the above process until all the influence factors of the heartbeat interval on the result have been calculated.

The purpose of setting the occlusion interval is to erase the information of the heartbeat. By comparing the results of the model without the heartbeat with the results of the model with the heartbeat, the impact factor of the heartbeat on the model can be calculated. Repeatedly perform this operation, you can get the influence factor of each heartbeat on the model result.

Step 2.3, visualize the impact factor of each heartbeat.

In step 2.2, the obtained difference ΔO can be used to represent the influence of the heartbeat interval on the final result of the model. However, the ECG data is lengthy and contains multiple heartbeat intervals, and the numerical method is not intuitive enough, so a corresponding visualization method needs to be designed. By mapping the values to the colors of the rectangles, the performance of each heartbeat interval can be visualized in the ECG data.

The specific method of step 2.3 is as follows:

(1) Code ΔO as a color; each heartbeat interval corresponds to a ΔO value, and a color sequence can be obtained after encoding

In order to visually display the meaning of ΔO, it is coded as a color in the present invention, and its rules are:

When ΔO>0, encode it as red, the larger the value, the deeper the red depth;

When ΔO < 0, it is encoded as blue, the smaller the value, the deeper the blue depth.

(2) Generate a gradient rectangle

Taking the length of each heartbeat interval as the rectangle width, and taking the height of the highest R peak on the electrocardiogram as the rectangle length, the electrocardiogram data sequence can be divided into several rectangles, and each rectangle contains a heartbeat interval. Fill the rectangle with the color generated by the heartbeat interval encoding. In order not to block the ECG information, the center of the rectangle is set to transparent, the ends are set to fill color, and the rectangle is adjusted to a gradient color band.

(3) Superimpose the rectangle to the ECG background

The visualization can be generated by superimposing the gradient rectangle corresponding to each heartbeat interval on the background of the electrocardiogram.

The supporting evidence and the opposing evidence of the model can be obtained by checking the color on each heartbeat interval, and the influence strength of the evidence on the final classification result can be judged by the color depth; through this method of the present invention, the classification result of the model can be obtained at the heartbeat level explain.

Step 3: Microscopically display the influence of the details of the ECG data on the model results.

In step 2, we used the heartbeat as an interval to find the impact of different heartbeats on the model results, and preliminarily explained the classification results of the model. However, some details within the heartbeat interval also have an important impact on the model classification results. If not processed, it will easily lead to the loss of details. Therefore, it is also necessary to visualize the details in the ECG sequence data, explain the basis of the model classification in more detail, and strengthen the interpretability of the model.

Step 3 specifically includes the following steps:

Step 3.1, set the movable occlusion interval.

Since the point-by-point influence factor needs to be calculated, the occlusion interval starts from the first point and moves backward one grid at a time. The range of the length L of the blocking interval is 10≤L≤20, and L=15 in this patent. This is an empirical value obtained by comparing the results of multiple experiments. Because the occlusion interval is too short, the difference between the output results of the model will be small, which cannot reflect the influence of a single point on the overall result; if it is too long, the influence of each point will be confused. The present invention regards the influence of the entire occlusion interval on the model result as the influence factor of the first point in the interval, so that the influence of each individual point on the model result can be obtained by moving the occlusion interval point by point.

Step 3.2, calculate the difference point by point.

After step 3.1, the length of the occlusion interval has been determined. Next, the occlusion interval is used to calculate the point-by-point difference. The specific steps of step 3.2 are as follows:

Step 3.2.1, starting from the first data of the S vector of the ECG sequence, set the following L vector value data to 0, and the vector values of the remaining positions remain unchanged to form an occlusion interval; the occlusion interval starts from the first data, Move backward one grid at a time until all the data in the ECG data are traversed;

The vector data expression of the ECG sequence S _m with the occlusion interval added in the mth cycle is:

S _m =[s ₁ ,s ₂ ,...,s _m-1 ,0,0,...,0,s _m+L ,...,s _n ]

Among them, s ₁ , s ₂ ,…,s _n represent the single data constituting the ECG sequence. It can be seen from this formula that s _m , s _m+1 ,…, s _m+L-1 are added with occlusion intervals, and the data in the interval are assigned to 0;

Step 3.2.2, input the electrocardiogram sequence S _m vector with the occlusion interval added in the mth cycle into the depth model, and obtain the output result Y _m of the model, the expression is:

Y _m =[y′ ₁ ,y′ ₂ ,...,y′ _N ]

In the formula, y′ ₁ , y′ ₂ ,…, y′ _N represent the output values on the labels 1, 2,…, N, respectively;

Step 3.2.3, calculate the difference ΔO _m between the new model output result obtained in step 3.2.2 and the model benchmark result, which reflects the influence of a single point on the model output result. The calculation formula is:

ΔO _m =y _I -y' _I

In the formula, I represents the label number of the reference value O calculated in step 1, y _I and y′ _I represent the depth model output value on the label number; ΔO _m represents the mth data in the heartbeat sequence for the depth model output. The impact factor of the result; ΔO _m >0 indicates that the point has a positive impact on the final classification result, which is the supporting evidence of the model. The larger the value, the better the final result of the model; ΔO _m <0 indicates that the point has a positive impact on the final classification result. It has a negative impact and is the evidence against the model. The value is negative, and the smaller the value, the more deviation from the final result;

Through the value of ΔO, the influence factor of each point on the ECG on the model classification result is obtained, and the detailed information in the ECG data can be explained.

Step 3.2.4, move the occlusion interval backward by one grid, and repeat the above process until the calculation of the last point is completed. Finally, the ΔO value of each point on the ECG can be obtained.

Step 3.3, visualize the point-by-point contribution.

In step 3.2, the ΔO value of each point is calculated, which can reflect the influence of individual points on the final classification result of the model. But looking at the value of each point is not intuitive, so it is also necessary to design a point-by-point visualization method. The point value is different from the heartbeat interval value. It is difficult to see the color of a single point. Therefore, the visualization method of the previous link cannot be used, and it must be displayed according to the characteristics of point-by-point data.

Step 3.3 The specific steps are as follows:

Step 3.3.1, encode the ΔO value of each point as height.

After step 3.2, in the electrocardiogram data sequence, each data corresponds to the ΔO value, and the ΔO value is further encoded as the height, and a point P on the ECG plane is determined by the abscissa of the data and the height encoded by ΔO: ΔO> 0, indicates that point P is in the upper area of the ECG, and the corresponding point on the ECG is displayed in red; ΔO=0, indicates that the point P falls on the zero axis, and the corresponding point on the ECG is displayed in black; ΔO<0, indicates that the point P is in the lower area of the ECG, and the corresponding point on the ECG is shown in blue.

In this way, each point on the ECG is divided into colors, showing their contribution to the classification results of the model. At the same time, each data in the ECG data sequence corresponds to the point P generated by ΔO.

Step 3.3.2, use a smooth curve to connect the points P corresponding to each data in the ECG data series.

Since the point P is too dense to reflect its information intuitively through color and height, it is necessary to use a smooth curve to connect the point P, and together with the zero axis to enclose several areas. The height of the curve reflects the magnitude of the absolute value of ΔO, and the peaks and valleys of the curve reflect the key evidence supporting and violating the model results.

Step 3.3.3, fill the area enclosed by the curve with color.

In order to make the information of the local detail area more intuitive, fill the color in the several areas formed in step 3.3.2 to make its attributes more obvious. The area above the zero axis is filled with red, indicating that the local area supports the final classification result of the model; the area below the zero axis is filled with blue, indicating that the local area violates the final classification result of the model. The original ECG curve has been divided into several paragraphs, which are represented by different colors. At the same time, the local detailed information of the ECG can be learned according to the filled area near the zero axis. The larger the area and the higher the peak, the greater the possibility of abnormality, which means that the area has a greater influence on the final result of the model. The visual display of the details of the ECG data further illustrates the basis for the formation of the model classification results and enhances the interpretability of the model.

To sum up, the present invention provides a method for visualizing the classification results of a depth model for ECG data, which is used to solve the defects of simple abstraction and insufficient interpretability of the existing depth model results. , and design solutions from the macro and micro perspectives to visualize the impact. Compared with the prior art, the present invention enhances the interpretability of the model results; under the traditional method, the model result is a specific classification result label, and there is no way to explain the basis for the result, and such a result is difficult to be used in the medical field. Doctor accepts. The method explains the results of the model, finds supporting evidence and evidence against the results obtained by the model, shows the influence of every detail on the final results obtained by the model, and greatly improves the interpretability of the model results; The interpretation process is visualized from the macro and micro perspectives: under the traditional method, the ECG data is messy and lengthy, and it is a time-consuming and laborious task to distinguish key information from it. The areas that have a greater impact on the model results are likely to be critical abnormal areas, such as the presence of abnormal phenomena such as the disappearance of P waves. This method excavates such regions from the ECG data, and displays them through visual elements such as color and height from both macroscopic and microscopic granularities, thereby making the model running process more intuitive and improving the interpretability of the model results. The invented method is suitable for various models and has strong scalability: the interpretation of the model results under the traditional method requires reference to the model structure and cannot be extended to other models. This method does not depend on a specific model, and all classification results of deep models applicable to ECG data can be interpreted and displayed by this method, and can be easily extended to an endless stream of improved models.

Example

Please refer to Fig. 1, in order to realize the final visualization effect, the visualization method of the present invention comprises the following steps:

S101, determining a benchmark result.

In this embodiment, the representation form after processing the original ECG data into an ECG sequence is:

S=[s ₁ , s ₂ ,...,s _i ,...,s _n ]

In the formula, S is an n-dimensional vector, i=1,2,...,n, s _i represents the data of the ith point in the sequence, input the data sequence into the preset trained depth model, and get the result data The format is:

Y=[y ₁ , y ₂ ,...,y _j ,...,y _N ]

In the formula, Y is an N-dimensional vector, N represents the number of labels of the model classification; j=1,2,...,N, y _j represents the classification value of the model on the label j, 0≤y _j ≤1, where, The label corresponding to the maximum value of y _j is the predicted classification result of the deep model. The value of y _j corresponding to the label is set as the reference value O, the label serial number is set as I, and the expression of the reference value O is:

O=max{y ₁ ,y ₂ ,…,y _j ,…,y _N }

In the formula, y _j represents the classification value of the model on the label _j , 0≤yj≤1;

S102, designing a visualization method for the influence of the heartbeat interval on the model result.

Please refer to Figure 2 to design a visualization method for the impact of the heartbeat interval on the model results. The specific steps include:

1) Dynamically determine the length of the occlusion interval.

From the original ECG data, the R-peak position label of each heartbeat can be obtained, and the interval between two R-peaks is considered to be the RR interval of a heartbeat. Therefore, the length of the occlusion interval of the kth heartbeat is set as:

Length _k = x _k+1 -x _k

In the formula, Length _k represents the length of the occlusion interval set on the kth RR interval, x _k represents the abscissa of the _kth R peak position, 0≤xk ≤Len, and Len represents the total length of the ECG sequence;

2) Calculate the impact of each heartbeat on the model results.

From the previous step, we obtained the length of the kth heartbeat interval, and then we need to set the occlusion interval according to this length.

S1, align the start position of the occlusion interval with the position of the R peak of the kth heartbeat, and set the interval length to Length _k , so that the occlusion interval just covers the kth heartbeat.

S2, uniformly assign the vector values in the occlusion interval to 0, and keep the vector values in other positions unchanged. The modified ECG sequence is:

S _k =[s ₁ ,s ₂ ,…,0,…,0,…,s _n ]

Wherein, _si represents the data of the ith point in the sequence, and the area assigned as 0 starts from the R peak of the kth heartbeat, and the length is Length _k ;

S3, in S2, we set the occlusion interval on the kth heartbeat interval, and now the ECG sequence _Sk vector with the occlusion interval added is input into the depth model, and the new depth model output result _Yk is obtained, _Yk is N dimensional vector, the expression is:

Y _k =[y′ ₁ ,y′ ₂ ,…,y′ _N ]

In the formula, y′ ₁ , y′ ₂ ,…,y′ _N represent the output values on the labels 1, 2,…,N, respectively;

S4, calculate and obtain the difference ΔO _k between the depth model result O _k that blocks the information of the kth heartbeat interval and the reference result; ΔOk is the influence factor of the _kth heartbeat interval, and the expression is:

ΔO _k =y _I -y' _I

In the formula, I represents the label number of the reference value O calculated in step 1, y _I and y′ _I represent the depth model output value on the label number; ΔO _k represents the kth heartbeat interval for the depth model output result. Impact factor; ΔO _k >0 indicates that the heartbeat interval has a positive impact on the model classification result, which is the supporting evidence of the model. The larger the value, the better the model classification result; ΔO _k <0 indicates that the heartbeat interval has a positive impact on the final classification result. It has a negative impact and is the evidence against the model. The value is negative, and the smaller the value, the greater the deviation from the model classification result;

Through the value of ΔO, the influence of different heartbeat intervals on the model classification results can be distinguished, and the interpretation of the model classification results can be realized.

S5, move the occlusion interval, repeat the above process, and calculate the influence of the k+1th heartbeat on the result.

The purpose of setting the occlusion interval in the embodiment of the present invention is to erase the information of the heartbeat. By comparing the results of the model without the heartbeat with the results of the model including the heartbeat, the impact value of the heartbeat on the model can be calculated. By repeating this operation, you can get the value of the impact of each heartbeat on the model results.

3) Visualize the impact of each heartbeat.

In step 2), the obtained difference ΔO can be used to represent the influence of the heartbeat interval on the final result of the model. However, the ECG data is very long and contains multiple heartbeat intervals. The numerical method is not intuitive enough. Therefore, a corresponding visualization method needs to be designed. By mapping the values to the colors of the rectangles, the performance of each heartbeat interval can be visualized in the ECG data. The specific method is as follows:

S1, encode ΔO as a color. In order to visualize the meaning of ΔO, it can be encoded as a color. The rule is: when ΔO>0, encode it as red, the larger the value, the deeper the red depth; when ΔO<0, encode it as red is blue, the smaller the value, the deeper the blue depth. Each heartbeat interval corresponds to a ΔO value, and after encoding, a color sequence can be obtained.

S2, generate a gradient rectangle.

Taking the length of each heartbeat interval as the width of the rectangle, and taking the height of the highest R peak on the electrocardiogram as the length of the rectangle, the electrocardiogram can be divided into several rectangles, and each rectangle contains a heartbeat interval. Fill the rectangle with the color generated by the heartbeat interval encoding.

At the same time, in order not to block the ECG information, the center of the rectangle is set to transparent, the two ends are set to fill color, and the rectangle is adjusted to a gradient color band.

S3, superimpose the rectangle to the ECG background.

Finally, the gradient rectangle corresponding to each heartbeat interval is superimposed on the background of the electrocardiogram to generate a visualization effect. According to the color on each heartbeat interval, the supporting evidence and the negative evidence for the model are obtained, and the classification results of the model are explained.

In this embodiment, we select the actual ECG data donated by AliveCor to illustrate the implementation process of the method. It should be pointed out that, as an example, this example only enumerates a data segment to illustrate the execution process of the method, and the actual electrocardiogram data is far beyond the enumeration range.

In the embodiment of the present invention, the electrocardiogram data segment is:

S=[...0bff 02ff fbfe f7fe f4fe f4fe f5fe f7fe f9fe fcfe 00ff 03ff07ff 09ff 0bff 0dff...];

In order to get the benchmark value, S is input into the model, and the model classification result is:

Y=[0.1215, 0.9877, 0.1010];

It can be seen from the classification results that the classification value corresponding to the AF label is the largest among all classification values, which is 0.9877, that is, the model classification result is considered to be AF, representing Atrial Fibrillation (atrial fibrillation). According to the previous definition, we can get the reference value y _I = 0.9877;

The labels of two ECG R peaks are obtained from the data, and the data in the RR interval is changed to 0 to form an occlusion interval. After modification, S is:

S=[...0000 0000 0000 0000 0000 0000f5fe f7fe f9fe fcfe 00ff 03ff07ff 09ff 0bff 0dff...]

Re-input it into the model to get a new classification result as:

Y=[0.2011, 0.6856, 0.1317]

At this time, the classification value corresponding to the label AF is y' _I =0.6856, and the influence factor ΔO=y _I -y' _I =0.3021 can be obtained from the formula.

Since ΔO>0, that is, after the heartbeat segment is occluded, the significance of the model classification result decreases. Therefore, we can think that the heartbeat interval supports the model classification result and is a positive basis for the model to obtain the result.

Repeat the above process, and calculate its impact factor for each heartbeat interval. The impact factor values are then encoded as colors, and gradient rectangles are generated and superimposed on the ECG waveform.

Please refer to Figure 4. The final visualization effect is shown in Figure 4. As can be seen from the figure, for each heartbeat interval, the color of the gradient rectangle shows the impact of the interval on the final result of the model, and the red part represents the support model classification As a result, the blue part represents the anti-model classification results, and the depth of the color reflects the magnitude of the effect. This visualization explains how each heartbeat interval contributes to the final classification of the model.

S103, designing a visualization method of point-by-point influence on the model result.

Referring to Figure 3, the implementation process of the visualization method for designing the influence of individual points on the model classification results is shown in Figure 3. The specific steps include:

1) Set the movable occlusion interval.

Since the point-by-point difference needs to be calculated, the occlusion interval starts from the first point and moves backward one grid at a time. The length of the occlusion interval is L=15.

2) Calculate the difference point by point.

After the first step, the length of the occlusion interval has been determined. Next, the occlusion interval is used to calculate the point-by-point difference. The specific steps are as follows:

S1, starting from the first data of the S vector of the electrocardiogram sequence, set the following L vector value data to 0, and the vector values of the remaining positions remain unchanged to form an occlusion interval; the occlusion interval starts from the first data, and each direction Then move a grid until all the data in the ECG data are traversed;

The vector data expression of the ECG sequence S _m with the occlusion interval added in the mth cycle is:

S _m =[s ₁ ,s ₂ ,...,s _m-1 ,0,0,...,0,s _m+L ,...,s _n ]

Among them, s ₁ , s ₂ ,…,s _n represent the single data constituting the ECG sequence. It can be seen from this formula that s _m , s _m+1 ,…, s _m+L-1 are added with occlusion intervals, and the data in the interval are assigned to 0;

S2, the electrocardiogram sequence S _m vector with the occlusion interval added in the mth cycle is input into the depth model, and the output result Y _m of the model is obtained, and the expression is:

Y _m =[y′ ₁ ,y′ ₂ ,...,y′ _N ]

In the formula, y′ ₁ , y′ ₂ , …, y′ _N represent the output values on the labels 1, 2, …, N, respectively;

S3, calculate the difference ΔO _m between the new model output result obtained in step S2 and the model reference result, this value reflects the influence of a single point on the model output result, and the calculation formula is:

ΔO _m =y _I -y' _I

In the formula, I represents the label number of the reference value O calculated in step 1, y _I and y′ _I represent the depth model output value on the label number; ΔO _m represents the mth data in the heartbeat sequence for the depth model output. The impact factor of the result; ΔO _m >0 indicates that the point has a positive impact on the final classification result, which is the supporting evidence of the model. The larger the value, the better the final result of the model; ΔO _m <0 indicates that the point has a positive impact on the final classification result. It has a negative impact and is the evidence against the model. The value is negative, and the smaller the value, the more deviation from the final result;

Through the value of ΔO, the influence factor of each point on the ECG on the model classification result is obtained, and the detailed information in the ECG data can be explained.

S4, move the occlusion interval back by one grid, repeat the above process, and finally calculate the ΔO value of each point on the electrocardiogram.

3) Visual display of point-by-point contributions.

In the previous step, the ΔO value of each point was calculated, which can reflect the influence of individual points on the final classification result of the model. But looking at the value of each point is not intuitive, so it is also necessary to design a point-by-point visualization method. The point value is different from the heartbeat interval value. It is difficult to see the color of a single point. Therefore, the visualization method of the previous link cannot be used, and it must be displayed according to the characteristics of point-by-point data. Specific steps are as follows:

S1, encode the ΔO value of each point as height.

After the previous step, in the ECG data sequence, each data corresponds to the ΔO value, and the ΔO value is further encoded as the height, and a point P on the ECG plane is determined by the abscissa of the data and the height encoded by ΔO: ΔO>0, it means that point P is in the upper area of the ECG, and the corresponding point on the ECG is displayed in red; ΔO=0, it means that point P falls on the zero axis, and the corresponding point on the ECG is displayed in black; ΔO<0, Indicates that point P is in the lower area of the ECG, and displays the corresponding point on the ECG in blue.

In this way, each point on the ECG is divided into colors, showing their contribution to the classification results of the model. At the same time, each data in the ECG data sequence corresponds to the point P generated by ΔO.

S2, use a smooth curve to connect the points P corresponding to each data in the ECG data sequence.

Since the points are too dense to directly reflect their information through color and height, it is necessary to use a smooth curve to connect the points P, and together with the zero axis to enclose several areas. The height of the curve reflects the magnitude of the absolute value of ΔO, and the peaks and valleys of the curve reflect the key evidence supporting and violating the model results.

S3, fill the area surrounded by the curve with color.

In order to make the information of the local detail area more intuitive, fill the color in the several areas formed in the previous step to make the properties more obvious. The area above the zero axis is filled with red, indicating that the local area supports the final classification result of the model; the area below the zero axis is filled with blue, indicating that the local area violates the final classification result of the model. The original ECG curve has been divided into several paragraphs, which are represented by different colors. At the same time, according to the filling area near the zero axis, the local details of the ECG can be learned. The larger the area and the higher the peak, the greater the possibility of abnormality, and the greater the impact on the final result of the model. The visual display of the details of the ECG data further illustrates the basis for the formation of the model classification results and enhances the interpretability of the model.

In this embodiment, as an example, the idea of the example in S102 is still continued. The difference is that in S102, in order to obtain the influence of each heartbeat interval on the classification result, the length and moving distance of the shielding window are dynamically adjusted according to the length of the heartbeat interval, and the effect is to shield exactly one heartbeat interval. In this step, in order to obtain the influence of each point on the classification result, the occlusion interval needs to adopt a fixed length, and at the same time move backward one point at a time, until the influence factors of all points have been calculated.

For example, a real piece of ECG data is:

S=[...2cff 2dff 2eff 2fff 32ff 35ff 37ff 3aff...];

Set the length of the occlusion interval to 15 and the moving distance to 1. In order to find the influence factor of the first point, we can set the occlusion interval from this point:

S=[...0000 0000 0000 000f 32ff 35ff 37ff 3aff...]

Input the data with the occlusion interval set into the model, and obtain a new classification value y′ _I =0.9903 according to the new output result of the model.

From the formula, the influence factor of the first point ΔO=y _I -y' _I =-0.0026 can be obtained.

Because ΔO<0, we can think that this point is against the classification result of the model, which is a negative basis for the model to obtain the classification result.

After that, the length of the occlusion interval remains unchanged, and it moves one point backward:

S=[...2000 0000 0000 0000 32ff 35ff 37ff 3aff...];

Repeat the above process to get the impact factor of each point.

Then, the points on the ECG plane are generated according to the visualization method, and are connected by a curve, and the curve and the zero axis can form several surrounding areas. Fill the surrounding area with the corresponding color according to the positive or negative of the influence factor.

Please refer to Figure 5, the final visualization effect is shown in Figure 5. As can be seen from the figure, the original waveform of the ECG is divided into three colors: red, blue and black, which represent the influence of this waveform on the final classification result of the model. At the same time, the points determined by the impact factor are connected into a curve, which together with the zero axis forms several regions, which explain the detailed basis for the model to obtain the final classification result. For example, an upward spike region appears in the boxed portion in Figure 5, suggesting that there may be anomalies in this region. According to medical knowledge, there is actually an abnormal disappearance of the P wave here. It is precisely because of this detail that the model finally obtained the classification result of atrial fibrillation (AF). It is difficult to intuitively find this area in an ordinary electrocardiogram, but with the help of the visualization method in the present invention, a strong peak appears here, indicating that compared with the ordinary method, this method can explain the classification results of the deep model more intuitively, that is, Improved interpretability of deep model classification results.

S104, a final visualization result is formed.

After the above steps and superimposing the macro and detail visualization effects, a comprehensive visualization effect from macro to details is finally established. The macro effect is shown in Figure 4, and the detail effect is shown in Figure 5. The visualization effect fully explains the model classification results, highlights the key basis for the model to make the classification results, and strengthens the interpretability of the model classification results. In practical applications, doctors can determine the heartbeat interval that may be abnormal according to the macro information, and quickly locate the specific heartbeat interval to view the details; they can also judge the possible abnormal phenomenon according to the detailed information, and find the key information from the waveform details, so as to Improved diagnostic efficiency.

In summary, the present invention discloses a deep model classification result visualization method oriented to electrocardiogram data, which includes the following steps: firstly, inputting original electrocardiogram data into a model, obtaining the original output data of the model, analyzing the final prediction result, and converting the original data into Save it as a benchmark to participate in subsequent comparisons; then dynamically set the occlusion interval according to the heartbeat interval, obtain the impact factor of each heartbeat on the final classification result, encode the impact factor as a color, and generate a gradient rectangle to be superimposed on the original ECG information for visualization. The method visually shows the impact of each heartbeat interval on the final classification result of the model. Next, reset the parameters of the movable occlusion interval, move the occlusion interval, calculate the deviation value of each point from the benchmark, superimpose this value with the original data, and display the detailed characteristics of the ECG through the peak value and area area, showing that the small area is useful for the model classification results. , revealing the key basis for the model to arrive at this result. The present invention interprets the model classification result by showing the influence of the electrocardiogram data in macroscopic and detailed granularities on the final model classification result, shows the key evidence that the model obtains the final result, and solves the problem of insufficient interpretability of the model result; The visual display method digs deeply into the key information in the ECG data, visualizes the running process of the model, and further improves the interpretability of the classification results of the deep model.

The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art can still modify or equivalently replace the specific embodiments of the present invention. , any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention are all within the protection scope of the claims of the present invention for which the application is pending.

Claims (10)

1. A depth model classification result visualization method oriented to electrocardiogram data is characterized by comprising the following steps:

step 1, processing acquired electrocardiogram data into an electrocardiogram sequence, and inputting the electrocardiogram sequence into a trained depth model to obtain a reference result;

step 2, taking the heartbeat interval as a basic unit, dynamically adjusting an occlusion interval according to heartbeat information in electrocardiogram data, erasing information of a selected heartbeat interval through the occlusion interval, comparing an output result of the depth model without the heartbeat interval information with a reference result output by the depth model containing the heartbeat information, and calculating to obtain an influence factor delta O of each heartbeat on the depth model;

and 3, visually expressing the influence factor delta O of each heartbeat by adopting a gradient color band to realize the visualization of the depth model classification result.

2. The method for visualizing the classification result of the depth model based on the electrocardiographic data according to claim 1, further comprising:

step 4, setting a movable shielding interval, and sequentially shielding each point in the electrocardiogram data; comparing the depth model output result of each point shielded by the electrocardiogram data with the reference result output by the depth model respectively to obtain the influence factor of each point on the depth model output result on the electrocardiogram data;

and 5, visually representing the influence factors of each point obtained in the step 4.

3. The method for visualizing the classification result of the depth model oriented to the electrocardiographic data according to claim 1, wherein the step 2 specifically comprises:

step 2.1, acquiring the length of each heartbeat interval according to the original electrocardiogram data, dynamically setting a shielding interval according to the length, and sequentially shielding each heartbeat interval;

step 2.2, the electrocardiogram sequence vectors added with the shielding intervals are respectively input into the depth model to obtain a new depth model output result;

and 2.3, respectively calculating the difference value between each new depth model output result obtained in the step 2.2 and the reference result obtained in the step 1, and obtaining the influence factor of each heartbeat interval on the depth model output result.

4. The method for visualizing the classification result of the depth model oriented to the electrocardiographic data according to claim 1, wherein the step 3 specifically comprises:

step 3.1, encoding the delta O value corresponding to each heartbeat interval to obtain a corresponding color sequence; the rule is as follows: when the delta O is more than 0, the delta O is coded into a preset color, and the larger the value is, the deeper the color depth is; when the delta O is less than 0, the delta O is coded into another different preset color, and the smaller the value is, the deeper the color depth is;

step 3.2, dividing the electrocardiogram data sequence into a plurality of rectangles by taking the length of each heartbeat interval as the width of the rectangle and the height of the highest R peak on the electrocardiogram as the length of the rectangle, wherein each rectangle comprises a heartbeat interval; filling the color generated by each heartbeat interval code obtained in the step 3.1 into a rectangle corresponding to each heartbeat interval;

and 3.3, overlaying the color-filled rectangles corresponding to the heartbeat intervals obtained in the step 3.2 on an electrocardiogram data background, so as to realize the visualization of the depth model classification result.

5. The method for visualizing the classification result of the depth model based on the electrocardiographic data according to claim 4, wherein in step 3.2, the center of the rectangle is set to be transparent, the two ends are set to be filled with color, and the rectangle is adjusted to be a gradient color band.

6. The method for visualizing the classification result of the depth model oriented to the electrocardiographic data according to claim 1, wherein the step 1 specifically comprises:

the representation form after processing the electrocardiogram data into an electrocardiogram sequence is as follows:

S＝[s₁，s₂，...，s_i，...，s_n]

wherein S is an n-dimensional vector, i is 1,2_iData representing the ith point in the sequence;

inputting the electrocardiogram sequence into the trained depth model, and obtaining the result data with the format as follows:

Y＝[y₁，y₂，...，y_j，...，y_N]

in the formula, Y is an N-dimensional vector, and N represents the number of labels of model classification; j ═ 1, 2., N, y_jRepresents the classification value of the model on the label j, and is more than or equal to 0 and less than or equal to y_j≤1；

Wherein, y_jTaking the label corresponding to the maximum value as the prediction classification result of the depth model, and taking the y corresponding to the label_jThe value is set as a reference value O, the label serial number is set as I, and the expression of the reference value O is as follows:

O＝max{y₁，y₂，...，y_j，...，y_N}

in the formula, y_jRepresents the classification value of the model on the label j, and is more than or equal to 0 and less than or equal to y_j≤1。

7. The method for visualizing the classification result of the depth model oriented to the electrocardiographic data according to claim 6, wherein the step 2 specifically comprises:

step 2.1, dynamically determining the length of an occlusion interval;

obtaining an R peak position label of each heartbeat from original electrocardiogram data, wherein an RR interval of one heartbeat is considered between two R peaks; setting the length of the kth shielding interval as follows:

Length_k＝x_k+1-x_k

in the formula, L ength_kIndicates the length, x, of the occlusion section provided in the k-th RR section_kX is 0. ltoreq. x representing the abscissa of the kth R peak position_kL en ≦ L en indicating the total length of the ECG sequence;

step 2.2, calculating influence factors of each heartbeat interval information on the output result of the depth model;

step 2.2.1, aligning the starting position of the shielding interval with the R peak position of the kth heartbeat, and setting the interval length as L ength_kEnabling the shielding section to cover the kth heartbeat section information;

step 2.2.2, uniformly assigning the vector values in the shielding intervals to be 0, keeping the vector values of the rest positions unchanged, and modifying the electrocardiogram sequence as follows:

S_k＝[s₁，s₂，...，0，...，0，...，s_n]

wherein s is_iData representing the ith point in the sequence, the region assigned a value of 0 starting from the R peak of the kth beat and having a length of L ength_k；

Step 2.2.3, adding the electrocardiogram sequence S with the shielded interval_kInputting the vector into the depth model to obtain a new depth model output result Y_k，Y_kIs an N-dimensional vector, and the expression is:

Y_k＝[y′₁，y′₂，...，y′_N]

of formula (II) to (III)'₁，y′₂，...，y′_NRepresent the output values on the 1, 2., N labels, respectively;

step 2.2.4, calculating to obtain a depth model result O for shielding the information of the kth heartbeat interval_kDifference from baseline result Δ O_k；ΔO_kFor the impact factor of the kth heartbeat interval, the expression is:

ΔO_k＝y_I-y′_I

in the formula, I representsLabel serial number, y of reference value O calculated in step 1_IAnd y'_IA depth model output value represented on the tag number; delta O_kRepresenting the influence factor of the kth heartbeat interval on the output result of the depth model; delta O_kIf the value is more than 0, the heartbeat interval has positive influence on the model classification result and is a support evidence of the model, and the larger the value is, the more the heartbeat interval is fit with the model classification result; delta O_kThe heartbeat interval is less than 0, which indicates that the heartbeat interval has negative influence on the final classification result and is the anti-evidence of the model, the value is a negative value, and the smaller the value is, the more deviation from the classification result of the model is indicated;

and through the numerical value of the delta O, the influence of different heartbeat intervals on the model classification result is distinguished, and the explanation of the model classification result is realized.

8. The method for visualizing the classification result of the depth model oriented to the electrocardiographic data according to claim 2, wherein the step 4 specifically comprises:

step 4.1, starting from the first data of the electrocardiogram sequence S vector, setting the data of the next L vector values as 0, keeping the vector values of the rest positions unchanged, and forming an occlusion interval;

an electrocardiogram sequence S with an occlusion interval added in the mth cycle_mThe vector data expression is:

S_m＝[s₁，s₂，...，s_m-1，0，0，...，0，s_m+L，...，s_n]

wherein s is₁，s₂，...，s_nRepresenting the individual data constituting the ECG sequence, as can be seen from the formula, s_m，s_m+1，...，s_m+L-1The shielding sections are added, and the data in the sections are all assigned to be 0;

step 4.2, calculating the difference value between the output result of the depth model and the reference result after the shielding interval is set point by point to obtain the value of an influence factor delta O of each point on the electrocardiogram;

the method comprises the following specific steps:

step 4.2.1, adding the electrocardiogram sequence S with the shielding section in the mth circulation_mInputting the vector into the depth model to obtain the output result Y of the model_mThe expression is:

Y_m＝[y′₁，y′₂，...，y′_N]

of formula (II) to (III)'₁，y′₂，...，y＇_NRepresent the output values on the 1, 2., N labels, respectively;

step 4.2.2, calculating the difference value delta O between the new model output result obtained in step 4.2.1 and the model reference result_mThis value reflects the influence of an individual point on the model output result, and the calculation formula is:

ΔO_m＝y_I-y′_I

wherein I represents the label number of the calculated reference value O, y_IAnd y'_IA depth model output value represented on the tag number; delta O_mRepresenting the influence factor of the mth data in the heartbeat sequence on the output result of the depth model; delta O_mThe value is more than 0, the point has positive influence on the final classification result and is the supporting evidence of the model, and the larger the value is, the more the point is fit with the final result of the model; delta O_m<0 indicates that the point has a negative effect on the final classification result and is the negative evidence of the model, and the value is negative, and the smaller the value is, the more deviation from the final result is indicated;

and obtaining the influence factor of each point on the electrocardiogram on the classification result of the model through the value of the delta O, and realizing the explanation of the detail information in the electrocardiogram data.

9. The method for visualizing the classification result of the depth model oriented to the electrocardiographic data according to claim 8, wherein the specific steps in the step 5 comprise:

step 5.1, encoding the delta O value of each point obtained in the step 4.2 into height, determining a point P on an electrocardiogram plane according to the position and the height of the point, wherein the delta O is more than 0, the point P is in the upper area of the electrocardiogram, and the corresponding point on the electrocardiogram is displayed as a preset color; Δ O ═ 0, which indicates that point P falls on the zero axis, and the corresponding point on the electrocardiogram is displayed as another preset color; Δ O <0, indicating that point P is in the lower region of the electrocardiogram, displaying the corresponding point on the electrocardiogram as another preset color; the preset colors are different;

step 5.2, connecting points formed by using the serial number as an abscissa and the delta O value as an ordinate by using a smooth curve, and enclosing a plurality of areas together with a zero axis; the height of the curve reflects the magnitude of the absolute value of delta O, and the peak and the valley of the curve reflect the key basis for supporting the model result and violating the model result;

and 5.3, filling the region surrounded by the curve of the step 5.2 and the zero axis by using preset different colors, and realizing the visualization of the classification result of the depth model.

10. The method as claimed in claim 8, wherein in step 4, the length L of the occlusion region is in the range of 10 ≦ L ≦ 20.

Images