Heart failure patient death risk assessment system based on deep learning

Abstract

The invention relates to the technical field of medical informatics, in particular to a heart failure patient death risk assessment system based on deep learning. The system comprises a memory and a processor, and the processor executes a computer program stored in the memory to realize the following steps: acquiring electrocardiogram data and respiratory rate data of a target person; according to the data difference and the time interval on the electrocardiogram data, obtaining the stability degree; performing mean filtering on the electrocardiogram data by using time windows of different scales to obtain data sequences under each scale; according to data changes on two sides of a maximum value point in the data sequence under each scale, a time interval of adjacent maximum value points and correlation between the data sequence and respiratory frequency data, obtaining a heart failure cycle specific value, and determining an error parameter by combining a sample entropy of the data sequence under each scale and a difference between the sample entropies under different scales; and thus, the attention to the target person is obtained. According to the invention, the accuracy and credibility of the attention determination result of the target person are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical informatics, and in particular to a deep learning-based mortality risk assessment system for heart failure patients. Background Art

[0002] Heart failure is a complex clinical syndrome characterized by an inability of cardiac output to meet the body's metabolic needs due to structural or functional abnormalities in the heart. Accurately assessing a patient's mortality risk is crucial for clinical decision-making, directly influencing treatment options (such as drug optimization, device implantation, or heart transplantation), resource allocation, and the timing of palliative care interventions.

[0003] In existing technologies, risk assessment mostly relies on static clinical characteristics (such as age, medical history) or a single dynamic physiological indicator, lacks an integrated analysis of real-time dynamic changes in cardiac function and basic health status, and is difficult to fully reflect the multidimensional risks of patients' disease progression; and based on a single time scale analysis of the RR interval (such as only focusing on short-term heart rate variability), it is unable to fully characterize the multidimensional abnormalities of the heart rhythm of heart failure patients (such as periodic oscillations at different time scales, rhythm disorders caused by sympathetic hyperactivity), and thus cannot accurately determine the level of attention paid to the person being tested, making subsequent risk assessments not timely enough. Summary of the Invention

[0004] In order to solve the problem that existing methods cannot accurately determine the attention level of the person to be tested, the purpose of the present invention is to provide a heart failure patient mortality risk assessment system based on deep learning. The technical solution adopted is as follows: The present invention provides a deep learning-based mortality risk assessment system for heart failure patients. The system includes a memory and a processor. The processor executes a computer program stored in the memory to implement the following steps: Obtain the target person's electrocardiogram data and respiratory rate data; The degree of stationarity is obtained based on the data differences and time intervals at different locations on the electrocardiogram data; mean filtering is performed on each data point on the electrocardiogram data using time windows of different scales to obtain data sequences at each scale; the heart failure cycle-specific value of the data sequence at each scale is obtained based on the data changes on both sides of the maximum point in the data sequence at each scale, the time interval between adjacent maximum points, the degree of stationarity, and the correlation between the data sequence at each scale and the respiratory rate data; determining an error parameter of the data sequence at each scale according to the sample entropy of the data sequence at each scale, the difference between the sample entropies at different scales, and the heart failure cycle-specific value; The degree of attention to the target person is determined using the error parameter and the sample entropy.

[0005] Preferably, obtaining the degree of smoothness based on data differences and time intervals at different locations on the electrocardiogram data includes: The data difference between adjacent data points on the electrocardiogram data is recorded as the first difference; The degree of smoothness is obtained according to the difference between each first difference and the average value of all first differences, the time interval between the maximum value and the minimum value on the electrocardiogram data, and the numerical difference between the maximum value and the minimum value on the electrocardiogram data.

[0006] Preferably, obtaining the degree of smoothness according to the difference between each first difference and the average of all first differences, the time interval between the maximum value and the minimum value on the electrocardiogram data, and the numerical difference between the maximum value and the minimum value on the electrocardiogram data includes: calculating a first ratio between the time interval and the numerical difference; The degree of smoothness is obtained based on the difference between each first difference and the average value of all first differences and the first ratio. The difference between each first difference and the average value of all first differences is negatively correlated with the degree of smoothness, and the first ratio is positively correlated with the degree of smoothness.

[0007] Preferably, obtaining the heart failure cycle-specific value of the data sequence at each scale based on the data changes on both sides of the maximum point in the data sequence at each scale, the time interval between adjacent maximum points, the degree of stability, and the correlation between the data sequence at each scale and the respiratory rate data includes: For any scale: Obtaining the oscillation performance value of the data sequence at any scale according to the difference between the data change degrees on the left and right sides of each maximum point at any scale, the number of maximum values, and the time interval between adjacent maximum points; Recording the correlation between the data sequence at any scale and the respiratory rate data sequence as a first correlation corresponding to the any scale, wherein the respiratory rate data sequence is obtained by sorting the respiratory rate data in chronological order; Obtaining a respiration correlation corresponding to any one of the scales according to the first correlation; The heart failure cycle-specific value of the data sequence at any scale is determined by combining the oscillation performance value, the respiratory correlation and the smoothness. The oscillation performance value is positively correlated with the heart failure cycle-specific value, and the respiratory correlation and the smoothness are both negatively correlated with the heart failure cycle-specific value.

[0008] Preferably, obtaining the oscillation performance value of the data sequence at any scale according to the difference between the data change degrees on the left and right sides of each maximum point at any scale, the number of maximum values, and the time interval between adjacent maximum points includes: Calculate the first average slope value of the data points on each side of each maximum point under any scale; Recording the difference between the first average slope values ​​corresponding to both sides of each maximum point at any scale as the second difference of each maximum point at any scale; Calculating a second ratio between the number of maximum values ​​of the data sequence at any scale and the average time interval between all adjacent maximum value points; According to the second difference and the second ratio of all maximum points at any scale, the oscillation performance value of the data sequence at any scale is obtained, the second difference is negatively correlated with the oscillation performance value, and the second ratio is positively correlated with the oscillation performance value.

[0009] Preferably, obtaining the respiratory correlation corresponding to any scale according to the first correlation includes: The difference between the length of the time window of any scale and the length of the time window of the scale corresponding to the largest first correlation is recorded as a third difference; A product of the third difference and the first correlation corresponding to any one scale is determined as the breathing correlation corresponding to any one scale.

[0010] Preferably, the determining of the heart failure cycle-specific value of the data sequence at any scale by combining the oscillatory performance value, the respiratory correlation and the smoothness comprises: Calculating a sum of the respiration correlation and an anti-zero parameter, wherein the anti-zero parameter is greater than 0; A third ratio of the oscillatory performance value to the sum value is calculated, and the product of the inverse of the smoothness level and the third ratio is determined as the heart failure cycle-specific value of the data sequence under any scale.

[0011] Preferably, determining the error parameter of the data sequence at each scale based on the sample entropy of the data sequence at each scale, the difference between the sample entropies at different scales, and the heart failure cycle-specific value includes: Sort the sample entropy of data sequences at all scales in ascending order to obtain a sample entropy sequence, and perform curve fitting on the data in the sample entropy sequence to obtain a sample entropy curve; For any scale: according to the slope value of the sample entropy curve of the any scale, the sample entropy of the data sequence of the any scale and the heart failure cycle-specific value, the error parameter of the data sequence at the any scale is obtained.

[0012] Preferably, obtaining the error parameter of the data sequence at any scale according to the slope value of the sample entropy curve at any scale, the sample entropy of the data sequence at any scale, and the heart failure cycle-specific value includes: The ratio of the absolute value of the slope value on the sample entropy curve of any scale to the sample entropy of the data sequence of any scale is recorded as the third ratio; the product of the third ratio and the heart failure cycle-specific value of the data sequence of any scale is determined as the error parameter of the data sequence at any scale.

[0013] Preferably, the determining the degree of attention to the target person by using the error parameter and the sample entropy includes: Sort all scales in descending order of sample entropy to obtain a scale sequence; use the error parameters of the data sequence at the first preset number of scales in the scale sequence as key parameters; Inputting the key parameters into the trained neural network to obtain predicted values; The degree of attention paid to the target person is determined based on the predicted value.

[0014] The present invention has at least the following beneficial effects: The present invention first evaluates the smoothness of the electrocardiogram data based on the data differences and time intervals at different positions on the electrocardiogram data, then uses time windows of different scales to perform mean filtering on each data point on the electrocardiogram data to obtain data sequences at each scale, and then accurately identifies the specificity of heart failure based on the data changes on both sides of the maximum point in the data sequence at each scale, the time interval between adjacent maximum points, and the correlation between the data sequence at each scale and the respiratory rate data, combined with the periodic oscillation characteristics unique to heart failure. Compared with single-scale analysis, the present invention can more comprehensively and meticulously reflect the dynamic changes in the cardiac function of the target person, and effectively improve the reference value of the feature; then, the heart failure specificity of the target person at each time scale is analyzed, and the error parameter is determined by combining the sample entropy of different data sequences, so that the parameters of the core data change characteristics are extracted, the interference of redundant information on the model is reduced, the computational complexity and the risk of overfitting are reduced, thereby realizing the determination of the attention level of the target person, and improving the accuracy and credibility of the results of the determination of the attention level of the target person. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0016] Figure 1 This is a flowchart of a method executed by a deep learning-based heart failure patient mortality risk assessment system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the death risk assessment system for heart failure patients based on deep learning proposed by the present invention is described in detail below in combination with the accompanying drawings and preferred embodiments.

[0018] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0019] The specific scheme of the heart failure patient death risk assessment system based on deep learning provided by the present invention is described in detail below with reference to the accompanying drawings.

[0020] Example of a heart failure patient mortality risk assessment system based on deep learning: The specific scenario targeted by this embodiment is: in the process of assessing the death risk of heart failure patients for target persons, the target persons need to be monitored. When all monitored indicators show large fluctuations or abnormalities, higher attention should be paid. Therefore, this embodiment will comprehensively analyze the data of different dimensions of the target persons to determine whether abnormalities have occurred, and determine the level of attention based on the results.

[0021] This embodiment proposes a deep learning-based mortality risk assessment system for heart failure patients. Figure 1 As shown, the method performed by the deep learning-based heart failure patient mortality risk assessment system of this embodiment includes the following steps: Step S1, obtaining the electrocardiogram data and respiratory rate data of the target person.

[0022] First, the electrocardiogram data and respiratory rate data of the target person in the current time period are obtained through the hospital information system. The current time period is a set of all historical moments whose time intervals with the current moment are less than or equal to the preset duration. The preset duration is set by the implementer according to the specific circumstances, and will not be elaborated here. In this embodiment, the sampling frequency of the electrocardiogram data and respiratory rate data are set to 125Hz. In specific applications, the implementer can set it according to the specific circumstances. It should be noted that: in this embodiment, the data are all obtained with full authorization. Then, a bandpass filter is used to filter the obtained electrocardiogram data to eliminate baseline drift. It should be noted that: the electrocardiogram data mentioned later are all filtered data.

[0023] So far, this embodiment has acquired the electrocardiogram data and respiratory rate data of the target person.

[0024] Step S2: Obtain the degree of smoothness based on the data differences and time intervals at different positions on the ECG data; perform mean filtering on each data point on the ECG data using time windows of different scales to obtain a data sequence at each scale; obtain the heart failure cycle-specific value of the data sequence at each scale based on the data changes on both sides of the maximum point in the data sequence at each scale, the time interval between adjacent maximum points, the degree of smoothness, and the correlation between the data sequence at each scale and the respiratory rate data.

[0025] In the electrocardiogram data of the target person, its unique long-term monitoring advantage is utilized to extract dynamic features, which are converted into multi-scale rhythm instability features, and then integrated with static clinical data to provide a dynamic and low-cost enhanced dimension for their mortality risk assessment.

[0026] The absolute value of the difference between adjacent data points in the ECG data is recorded as a first difference. Thus, multiple first differences are obtained, with one first difference existing between every two adjacent data points. The ratio between the time interval between the maximum and minimum values ​​in the ECG data and the numerical difference between the maximum and minimum values ​​in the ECG data is calculated and recorded as a first ratio.

[0027] The degree of smoothness is obtained based on the difference between each first difference and the average value of all first differences and the first ratio. The difference between each first difference and the average value of all first differences is negatively correlated with the degree of smoothness, and the first ratio is positively correlated with the degree of smoothness.

[0028] Among them, a positive correlation relationship indicates that the dependent variable will increase as the independent variable increases, and the dependent variable will decrease as the independent variable decreases. It can be an additive relationship, a multiplicative relationship, etc., which is determined by actual application; a negative correlation relationship indicates that the dependent variable will decrease as the independent variable increases, and the dependent variable will increase as the independent variable decreases. It can be a subtractive relationship, a division relationship, etc., which is determined by actual application.

[0029] In this embodiment, a calculation formula for the degree of stability is given, and the degree of stability can be specifically expressed as: in, Indicates the degree of stability, represents the number of data points on the ECG data, Represents the data difference between the i-th data point and the i+1-th data point on the electrocardiogram data, that is, the first difference; Represents the average value of the data differences between all adjacent data points on the electrocardiogram data, Represents the time interval between the maximum and minimum values ​​on the electrocardiogram data. Indicates the numerical difference between the maximum and minimum values ​​in the electrocardiogram data. Indicates the preset first adjustment parameter, Indicates the absolute value sign, Represents the normalization function.

[0030] In this embodiment, a preset first adjustment parameter is introduced into the calculation formula of the degree of smoothness in order to prevent the denominator from being 0. In this embodiment, the preset first adjustment parameter is 0.01. In specific applications, the implementer may set it according to specific circumstances. Represents the first ratio. When the difference between adjacent data is smaller, the short-term fluctuation of the ECG data during the monitoring process is smaller, that is, the more stable it is, and the larger the time scale division is, the more effective features can be captured. When the amplitude of the extreme value difference in the data sequence is larger, the data sequence is less stable, and if the time interval between the maximum and minimum values ​​is smaller, the data sequence should be divided into smaller time scales to capture features that change significantly in the short term.

[0031] Combined with the preset time windows of different scales, each data point on the electrocardiogram data is subjected to mean filtering processing. The duration range of the preset time window is [30, 500]. The difference between two adjacent data in this range is 10, that is, the duration of the time window of the first scale is 30, the duration of the time window of the second scale is 40, and the duration of the time window of the third scale is 50. By analogy, the difference between the durations of the time windows of two adjacent scales is 10, and data sequences at multiple different scales are obtained. In specific applications, the implementer can set the duration of time windows of different dimensions according to the specific situation. Mean filtering processing is a prior art and will not be described in detail here.

[0032] For data sequences at all scales, in order to accurately extract the rhythmic characteristics of the target individual, we quantify this rhythmic instability based on multi-scale entropy and integrate it as a dynamic feature with static clinical features to improve the performance of the deep learning-based mortality risk assessment model. It may exhibit rapid fluctuations on short time scales and slower trends on long time scales.

[0033] Sample entropy is calculated for each data sequence. Sample entropy represents a measure of the time series complexity of the data sequence. Lower sample entropy indicates lower complexity and higher self-similarity. Higher sample entropy reflects more complex and variable rhythms in the target individual, driven by healthy autonomic nervous system regulation. Higher entropy values ​​indicate more abnormal data, suggesting advanced decompensated heart failure. However, before calculating sample entropy, preliminary identification of the data within the data sequence is required, including periodic oscillations unique to heart failure patients.

[0034] The more evenly distributed the maximum points in a data series are, the more similar the rise and fall rates are, and the more similar the intervals between the maximum points are, the more stable the oscillation is. Based on these characteristics, the oscillation value of the data series at each scale can be determined.

[0035] For any scale: Calculate the average slope of the data points on each side of each maximum point at the scale, and record this average slope value as the first average slope value. The maximum point is obtained by fitting all data curves in the data sequence at the scale to obtain a corresponding fitted curve, and extracting the maximum point on the curve as the maximum point at the scale. It should be noted that when calculating the average slope of the data points on each side of each maximum point, for any maximum point, the data points on each side of the maximum point are all the data points between the maximum point and its adjacent extreme point.

[0036] The absolute value of the difference between the first average slope values ​​corresponding to each maximum point on the scale is recorded as the second difference for each maximum point on the scale. The ratio of the number of maxima in the data sequence on the scale to the average time interval between all adjacent maximum points is calculated and recorded as the second ratio. Based on the second differences of all maximum points on the scale and the second ratio, the oscillation performance value of the data sequence on the scale is obtained. The second difference is negatively correlated with the oscillation performance value, and the second ratio is positively correlated with the oscillation performance value.

[0037] In this embodiment, a specific calculation formula for the oscillation performance value is given. The oscillation performance value of the data series at each scale can be expressed as: in, Indicates the The oscillation performance value of the data series at each scale, Indicates the The number of maximum values ​​of the data sequence at each scale, represents the average time interval between all adjacent maximum points, Indicates the The average of the second differences of all maximum points under the scale, Indicates the preset second adjustment parameter, Indicates the absolute value sign.

[0038] No. The smaller the average value of the second difference of all maximum points under a scale, the more similar the rising and falling slope ratios of the data on the left and right sides of the maximum point are, and the data changes have an oscillation-related pattern; the more maximum points there are in the data sequence and the more similar the intervals between adjacent maximum points are, the more support it can provide as data support that the data sequence has periodic oscillations, and the higher its oscillation performance value is.

[0039] When the time scale of the data sequence is smaller, the target person's respiratory signal will affect the spacing between adjacent R waves on the electrocardiogram lead to form RSA. Therefore, it is necessary to determine the correlation of the respiratory signal at each scale. When the human body breathes normally, there is a coupling correlation between the respiratory signal and the heart rate. However, for patients with heart failure, the coherence between the data sequence and respiration is significantly reduced. In the data sequence with larger scales, the respiratory pattern changes caused by heart failure, and slow cycle coupling occurs. For example, multiple breaths correspond to one heart rate fluctuation, which is an abnormal manifestation of the RR interval. The smaller the time scale, the more significantly the correlation of the respiratory RR interval decreases. Therefore, the smaller the time scale, the less accurate the coupling obtained.

[0040] All collected respiratory rate data are sorted in chronological order to obtain a respiratory rate data sequence.

[0041] For any scale: The dynamic time warping (DTW) distance between the data sequence at that scale and the respiratory rate data sequence is calculated. The negative correlation normalization result of this DTW distance is used as the correlation between the two sequences, and this correlation is recorded as the first correlation corresponding to that scale. The negative correlation normalization result of the DTW distance is obtained by taking the value of an exponential function with a natural constant as the base and the negative of the DTW distance as the exponent. The scale with the highest correlation data is more effective in capturing the characteristics of heart failure patients. The respiratory correlation at each scale is thus determined. Each scale has a corresponding first correlation.

[0042] The absolute value of the difference between the length of the time window of the scale and the length of the time window of the scale corresponding to the largest first correlation is recorded as the third difference. The product of the third difference and the first correlation corresponding to the scale is determined as the respiratory correlation corresponding to the scale.

[0043] When the oscillation performance value of the target person at this scale is stronger and the respiratory correlation at the scale is weaker, it reflects that the data sequence data and the respiratory situation at this time are more disordered, the coupling is weaker, and the autonomic nervous system imbalance of the target person is worse.

[0044] Based on the above characteristics, this embodiment combines the oscillation performance value, respiratory correlation and smoothness to determine the heart failure cycle-specific value of the data sequence under this scale. The oscillation performance value is positively correlated with the heart failure cycle-specific value, and the respiratory correlation and smoothness are both negatively correlated with the heart failure cycle-specific value.

[0045] In this embodiment, a specific method for obtaining the heart failure cycle-specific value is provided. Specifically, the sum of the respiratory correlation corresponding to the scale and the anti-zero parameter is calculated, where the anti-zero parameter is greater than 0. In this embodiment, the anti-zero parameter is 0.01. In specific applications, the implementer can set it according to specific circumstances. The ratio of the target person's oscillation performance value at the scale to the sum is recorded as the third ratio. The product of the inverse of the degree of stability and the third ratio is determined as the heart failure cycle-specific value of the data sequence at the scale.

[0046] By using the above method, the heart failure cycle-specific value of the data sequence at each scale can be obtained.

[0047] Step S3, determining the error parameter of the data sequence at each scale according to the sample entropy of the data sequence at each scale, the difference between the sample entropies at different scales, and the heart failure cycle-specific value.

[0048] For the data sequence at each scale, its sample entropy is obtained; the sample entropy reflects the complex rhythmicity of the RR interval of the data sequence. The higher the entropy value, the richer the rhythmic changes of a healthy heart; the lower the entropy value, the more likely it is that the sympathetic nerve hyperactivity of the heart in heart failure is suppressing normal fluctuations.

[0049] The sample entropy of data sequences at all scales is sorted in order from small to large to obtain a sample entropy sequence, and curve fitting is performed on the data in the sample entropy sequence to obtain a sample entropy curve.

[0050] Next, based on the fluctuations of the sample entropy curve and combined with the specificity of the heart failure cycle, a reasonable parameter will be quantified for effective learning of the death risk assessment system for heart failure patients based on deep learning.

[0051] For each inflection point, the degree of scale deviation is determined by combining the specific evaluation values ​​at different time scales to determine the error parameter. The larger the scale difference between each segment between inflection points, the larger the change in sample entropy, which can reflect the unit scale change within the segment. Since the calculation of sample entropy only focuses on amplitude changes, the periodic oscillations unique to heart failure are difficult to capture.

[0052] For any scale, the ratio of the absolute value of the slope of the sample entropy curve for that scale to the sample entropy of the data sequence at that scale is recorded as the third ratio. The product of the third ratio and the heart failure cycle-specific value of the data sequence at that scale is determined as the error parameter of the data sequence at that scale. Using this method, the error parameter of the data sequence at each scale can be obtained.

[0053] Step S4: Determine the degree of attention to the target person using the error parameter and the sample entropy.

[0054] In multiscale entropy analysis, changes in sample entropy at different time scales contain crucial information about heart failure data. To accurately extract indicators that reflect the true characteristics of the target individual, systematic screening of error parameters across all time scales is required. The specific process is as follows: First, the decreasing value of sample entropy between two adjacent time scales is calculated. This decreasing value represents the magnitude of the temporal change in cardiac rhythm complexity. All decreasing values ​​are sorted and a preset number of values ​​with the largest decreasing values ​​are selected. In this embodiment, the preset number is 3; in specific applications, the user can set this value based on specific circumstances. This strategy is based on the following logic: a larger decreasing value indicates a more dramatic transition from relatively complex (high entropy) to significantly simpler (low entropy) rhythms within the corresponding time scale variation interval. This transition is highly consistent with the rhythm regulation abnormalities caused by autonomic dysfunction and myocardial electrical remodeling during the progression of heart failure, and therefore best reflects the true characteristics of the target individual. By selecting these three time scales with the largest decrements and extracting their corresponding error parameters as key parameters, the team not only preserved the trend characteristics of the cardiac rhythm at different time scales, but also significantly enhanced sensitivity to heart failure-specific patterns. Ultimately, these key parameters serve as input features for the deep learning model, improving the accuracy and reliability of the model's output.

[0055] To comprehensively measure the abnormality of target personnel, this embodiment uses a deep learning model that integrates static and dynamic features, and combines edge computing with a risk warning mechanism for processing. The specific steps are as follows: 1. Feature extraction and branch processing: A dual-branch structure is constructed. The static branch inputs clinical features (such as age, medical history, and laboratory indicators) and extracts high-level semantic information through a two-layer fully connected network (FCN). The dynamic branch uses three key features as input and captures multi-scale dynamic changes in cardiac rhythm through a one-layer FCN.

[0056] 2. Feature fusion and model training: The output vectors of the static and dynamic branches are concatenated and input into the fusion FCN layer to achieve feature interaction. Finally, the output layer outputs the anomaly indicators and uses historical data to train and optimize the model.

[0057] 3. Clinical Application and Risk Warning: Clinicians can use the neural network's output to determine whether to prioritize a target individual and assess their mortality risk. Higher abnormality indicators indicate a higher level of concern. When the neural network's output exceeds a preset threshold, the system immediately sends a warning to the physician, prompting them to promptly prioritize the individual and initiate appropriate intervention measures. Implementers can set these thresholds based on their specific circumstances, and we will not elaborate on them here.

[0058] At this point, the focus of the target person has been determined, and subsequent doctors can assess the target person's risk of death based on the observed situation.

[0059] This embodiment first evaluates the stationarity of the ECG data based on data differences and time intervals at different locations on the ECG data. Then, mean filtering is performed on each data point in the ECG data using time windows of different scales to obtain data sequences at each scale. Next, based on the data changes on both sides of the maximum point in the data sequence at each scale, the time intervals between adjacent maximum points, and the correlation between the data sequence at each scale and the respiratory rate data, the specificity of heart failure is accurately identified in combination with the periodic oscillation characteristics unique to heart failure. Compared with single-scale analysis, this method can more comprehensively and meticulously reflect the dynamic changes in the target person's cardiac function, effectively improving the reference value of the features. Furthermore, the specificity of heart failure for the target person at each time scale is analyzed, and error parameters are determined based on sample entropy of different data sequences. This allows the extraction of parameters that reflect the core data change features, reduces the interference of redundant information on the model, and reduces computational complexity and overfitting risk, thereby achieving determination of the target person's attention level and improving the accuracy and reliability of the results of the determination of the target person's attention level.

[0060] It should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A deep learning-based mortality risk assessment system for heart failure patients, characterized by: The system includes a memory and a processor, and the processor executes a computer program stored in the memory to implement the following steps: Obtain the target person's electrocardiogram data and respiratory rate data; The degree of stationarity is obtained based on the data differences and time intervals at different locations on the electrocardiogram data; mean filtering is performed on each data point on the electrocardiogram data using time windows of different scales to obtain data sequences at each scale; the heart failure cycle-specific value of the data sequence at each scale is obtained based on the data changes on both sides of the maximum point in the data sequence at each scale, the time interval between adjacent maximum points, the degree of stationarity, and the correlation between the data sequence at each scale and the respiratory rate data; determining an error parameter of the data sequence at each scale according to the sample entropy of the data sequence at each scale, the difference between the sample entropies at different scales, and the heart failure cycle-specific value; The degree of attention to the target person is determined using the error parameter and the sample entropy.

2. The heart failure patient death risk assessment system based on deep learning according to claim 1 is characterized in that: The method of obtaining the degree of smoothness based on the data differences and time intervals at different locations on the electrocardiogram data includes: The data difference between adjacent data points on the electrocardiogram data is recorded as the first difference; The degree of smoothness is obtained according to the difference between each first difference and the average value of all first differences, the time interval between the maximum value and the minimum value on the electrocardiogram data, and the numerical difference between the maximum value and the minimum value on the electrocardiogram data.

3. The deep learning-based heart failure patient mortality risk assessment system according to claim 2, characterized in that: Obtaining the degree of smoothness according to the difference between each first difference and the average of all first differences, the time interval between the maximum value and the minimum value on the electrocardiogram data, and the numerical difference between the maximum value and the minimum value on the electrocardiogram data includes: calculating a first ratio between the time interval and the numerical difference; The degree of smoothness is obtained based on the difference between each first difference and the average value of all first differences and the first ratio. The difference between each first difference and the average value of all first differences is negatively correlated with the degree of smoothness, and the first ratio is positively correlated with the degree of smoothness.

4. The heart failure patient death risk assessment system based on deep learning according to claim 1, characterized in that: The heart failure cycle-specific value of the data sequence at each scale is obtained based on the data changes on both sides of the maximum point in the data sequence at each scale, the time interval between adjacent maximum points, the degree of stability, and the correlation between the data sequence at each scale and the respiratory rate data, including: For any scale: Obtaining the oscillation performance value of the data sequence at any scale according to the difference between the data change degrees on the left and right sides of each maximum point at any scale, the number of maximum values, and the time interval between adjacent maximum points; Recording the correlation between the data sequence at any scale and the respiratory rate data sequence as a first correlation corresponding to the any scale, wherein the respiratory rate data sequence is obtained by sorting the respiratory rate data in chronological order; Obtaining a respiration correlation corresponding to any one of the scales according to the first correlation; The heart failure cycle-specific value of the data sequence at any scale is determined by combining the oscillation performance value, the respiratory correlation and the smoothness. The oscillation performance value is positively correlated with the heart failure cycle-specific value, and the respiratory correlation and the smoothness are both negatively correlated with the heart failure cycle-specific value.

5. The deep learning-based heart failure patient mortality risk assessment system according to claim 4, characterized in that: Obtaining the oscillation performance value of the data sequence at any scale according to the difference between the data change degrees on the left and right sides of each maximum point at any scale, the number of maximum values, and the time interval between adjacent maximum points includes: Calculate the first average slope value of the data points on each side of each maximum point under any scale; Recording the difference between the first average slope values ​​corresponding to both sides of each maximum point at any scale as the second difference of each maximum point at any scale; Calculating a second ratio between the number of maximum values ​​of the data sequence at any scale and the average time interval between all adjacent maximum value points; According to the second difference and the second ratio of all maximum points at any scale, the oscillation performance value of the data sequence at any scale is obtained, the second difference is negatively correlated with the oscillation performance value, and the second ratio is positively correlated with the oscillation performance value.

6. The heart failure patient death risk assessment system based on deep learning according to claim 4, characterized in that: Obtaining the respiratory correlation corresponding to any one scale according to the first correlation includes: The difference between the length of the time window of any scale and the length of the time window of the scale corresponding to the largest first correlation is recorded as a third difference; A product of the third difference and the first correlation corresponding to any one scale is determined as the breathing correlation corresponding to any one scale.

7. The heart failure patient death risk assessment system based on deep learning according to claim 4, characterized in that: The determining of the heart failure cycle-specific value of the data sequence at any scale by combining the oscillation performance value, the respiratory correlation and the smoothness includes: Calculating a sum of the respiration correlation and an anti-zero parameter, wherein the anti-zero parameter is greater than 0; A third ratio of the oscillatory performance value to the sum value is calculated, and the product of the inverse of the smoothness level and the third ratio is determined as the heart failure cycle-specific value of the data sequence under any scale.

8. The heart failure patient death risk assessment system based on deep learning according to claim 1, characterized in that: The step of determining the error parameter of the data sequence at each scale based on the sample entropy of the data sequence at each scale, the difference between the sample entropies at different scales, and the heart failure cycle-specific value includes: Sort the sample entropy of data sequences at all scales in ascending order to obtain a sample entropy sequence, and perform curve fitting on the data in the sample entropy sequence to obtain a sample entropy curve; For any scale: according to the slope value of the sample entropy curve of the any scale, the sample entropy of the data sequence of the any scale and the heart failure cycle-specific value, the error parameter of the data sequence at the any scale is obtained.

9. The deep learning-based mortality risk assessment system for heart failure patients according to claim 8, characterized in that: Obtaining an error parameter of the data sequence at any scale according to the slope value of the sample entropy curve at any scale, the sample entropy of the data sequence at any scale, and the heart failure cycle-specific value includes: The ratio of the absolute value of the slope value on the sample entropy curve of any scale to the sample entropy of the data sequence of any scale is recorded as the third ratio; the product of the third ratio and the heart failure cycle-specific value of the data sequence of any scale is determined as the error parameter of the data sequence at any scale.

10. The deep learning-based heart failure patient mortality risk assessment system according to claim 1, characterized in that: The determining of the degree of attention to the target person by using the error parameter and the sample entropy includes: Sort all scales in descending order of sample entropy to obtain a scale sequence; use the error parameters of the data sequence at the first preset number of scales in the scale sequence as key parameters; Inputting the key parameters into the trained neural network to obtain predicted values; The degree of attention paid to the target person is determined based on the predicted value.

Images