Heart failure risk assessment method and system based on AI

Abstract

This invention provides an AI-based heart failure risk assessment method and system, which relates to the field of artificial intelligence technology. The system uses a recursive segmentation algorithm to identify mutation points in the time series of physiological indicators, extracts a mutation feature matrix and maps it to a time series, identifies periodic mutations and gradual change patterns, converts them into risk scores, generates a risk accumulation curve, calculates the indicator transmission delay, plots the risk transmission path, identifies the intervention time window, and ultimately determines the heart failure risk level and generates an assessment report. This invention can achieve early and accurate identification of heart failure risk, providing the optimal opportunity for clinical intervention.

Description

Technical Field

[0001] The present invention relates to artificial intelligence technology, and in particular to an AI-based heart failure risk assessment method and system. Background Art

[0002] Heart failure is a common cardiovascular disease characterized by high morbidity, mortality, and hospital readmission rates. Due to the heart's inadequate pumping function, the body's metabolic needs cannot be met. With the aging population, the number of heart failure patients is increasing annually, placing a heavy burden on the healthcare system. Early risk assessment is crucial for the prevention and intervention of heart failure, effectively reducing its incidence and the risk of its complications.

[0003] Traditional heart failure risk assessment relies primarily on clinical judgment based on patient history, physical examination, and laboratory tests. Advances in medical technology have enabled wearable devices and remote monitoring systems to continuously collect multiple physiological indicators, providing a wealth of real-time data for dynamic heart failure risk assessment. The rise of artificial intelligence (AI) technology has provided new insights into processing and analyzing this complex medical data.

[0004] Existing heart failure risk assessment technologies have the following shortcomings: Most assessment methods are relatively simple in processing time series data, usually focusing only on outliers or static statistical features at a single time point, and are unable to capture the complex change patterns and mutation characteristics of physiological indicators in the time dimension, resulting in insufficient sensitivity to small changes in the early stages of the disease. Existing assessment models often consider each physiological indicator in isolation, lack in-depth analysis of the interrelationships between multidimensional indicators, and are unable to effectively identify the conduction relationship and temporal dependency between indicators, thus missing important clues to the pathophysiological mechanisms in the development of heart failure. Existing technologies are relatively simple in presenting risk assessment results, usually only giving risk levels or probability values, lacking accurate identification of intervention time windows and detailed descriptions of specific abnormal indicators, making it difficult to provide clinicians with targeted intervention recommendations. Summary of the Invention

[0005] The embodiments of the present invention provide an AI-based heart failure risk assessment method and system, which can solve the problems in the existing technology.

[0006] A first aspect of an embodiment of the present invention provides an AI-based heart failure risk assessment method, comprising:

[0007] The physiological indicators of the subject to be evaluated are collected and constructed into a multidimensional time series data stream. The difference between adjacent data segments in the multidimensional time series data stream is calculated using a recursive segmentation algorithm. The time point where the difference is greater than a preset difference threshold is determined as the mutation point. The waveform features of the data segments before and after the mutation point are extracted to generate a mutation feature matrix.

[0008] Perform time series mapping on the mutation feature matrix to obtain the time interval sequence and amplitude sequence of the mutation features, identify periodic mutation patterns and gradual change patterns, convert the duration and frequency of periodic mutation patterns and gradual change patterns into risk scores, and draw a risk accumulation curve;

[0009] Calculate the temporal relationship between physiological indicators based on the risk accumulation curve, convert the temporal relationship into the indicator conduction delay, draw a risk conduction path diagram based on the indicator conduction delay, and mark the time interval with the largest change slope in the risk conduction path diagram as the intervention time window;

[0010] Based on the time range of the intervention time window and the changing trend of the risk accumulation curve, the heart failure risk level is determined, and a risk assessment report containing the risk level and description of abnormal physiological indicators is generated.

[0011] In an optional embodiment,

[0012] The physiological indicators of the subject to be evaluated are collected and constructed into a multidimensional time series data stream. The difference between adjacent data segments in the multidimensional time series data stream is calculated through a recursive segmentation algorithm. The time point when the difference is greater than a preset difference threshold is determined as a mutation point.

[0013] Acquiring physiological indicators of the subject to be evaluated, the physiological indicators including cardiovascular data, respiratory data, and blood oxygen data; resampling the physiological indicators through cubic spline interpolation to construct a multidimensional time series data stream, and constructing the multidimensional time series data stream into a state matrix of time-indicator dimensions;

[0014] Initially dividing the state matrix from the midpoint to obtain two sub-data segments, moving the dividing position point by point along the time axis to calculate the Mahalanobis distance of the data segments on both sides to obtain a distance sequence, and performing eigenvalue decomposition to obtain a principal component weight vector, and using the principal component weight vector to weight the distance sequence to obtain a sub-data segment difference sequence;

[0015] A cumulative distribution function is obtained by statistically analyzing the probability distribution characteristics of the difference degree sequence of the sub-data segments, a difference threshold is determined on the cumulative distribution function according to a preset false alarm rate, a time point greater than the difference threshold is marked as an initial mutation point, an observation window is constructed with the initial mutation point as the center, and the mutual information between the signals of each dimension in the observation window is calculated to obtain a mutual information matrix, and the sum of the off-diagonal elements of the mutual information matrix is used as a coupling metric value;

[0016] A comprehensive difference is calculated based on the coupling metric value and the sub-data segment difference sequence, the initial mutation point is screened according to the comprehensive difference, and the time point when the comprehensive difference is greater than the difference threshold is determined as the final mutation point.

[0017] In an optional embodiment,

[0018] Extract the waveform features of the data segments before and after the mutation point and generate the mutation feature matrix including:

[0019] The fluctuation characteristics of the data segments before and after the mutation point are calculated according to the location of the mutation point to obtain a volatility sequence, and the fluctuation trend characteristics are obtained by performing a segmented cumulative calculation. The adaptive window length is determined based on the fluctuation trend characteristics, and the waveform feature extraction interval before and after the mutation point is determined using the adaptive window length.

[0020] In the waveform feature extraction interval, the data segments before and after the mutation point are respectively subjected to time-frequency transformation to obtain a time-frequency spectrum, the energy distribution of the time-frequency spectrum is calculated to obtain a time-frequency energy sequence, the time-frequency energy sequence is weightedly fused in combination with the volatility sequence to obtain a fused feature sequence, and the multi-scale entropy value of the fused feature sequence is calculated to obtain a complexity feature;

[0021] The volatility sequence, time-frequency energy sequence and complexity feature are combined in time sequence to obtain a feature vector, an initial feature matrix is constructed according to the feature vector, the discriminant contribution of each feature in the initial feature matrix is calculated to obtain the feature weight, and the feature weight is used to perform weighted reconstruction on the initial feature matrix to generate a mutation feature matrix.

[0022] In an optional embodiment,

[0023] Perform time series mapping on the mutation feature matrix to obtain the time interval sequence and amplitude sequence of the mutation features, identify periodic mutation patterns and gradual change patterns, convert the duration and frequency of periodic mutation patterns and gradual change patterns into risk scores, and draw the risk accumulation curve, including:

[0024] Performing time series mapping on the mutation feature matrix to obtain a time interval sequence and an amplitude sequence, and constructing a mutation topology network based on the time interval sequence and amplitude sequence using an adaptive weight function, where nodes represent mutation points and the connection weights between nodes are determined by the time interval and amplitude.

[0025] Calculating the local entropy of nodes at different scales of the mutant topology network, determining the optimal scale set according to the variation characteristics of the local entropy, and fusing the local entropies under the optimal scale set to obtain an enhanced topological entropy sequence;

[0026] Dynamically adjusting the calculation window size based on the rate of change of the enhanced topological entropy sequence, calculating the autocorrelation function and power spectral density of the enhanced topological entropy sequence within the calculation window, extracting periodic mutation characteristics, determining a first duration and a first frequency of occurrence of a periodic mutation pattern based on the periodic mutation characteristics, performing piecewise polynomial fitting on the enhanced topological entropy sequence to identify a gradual change interval, and determining a second duration and a second frequency of occurrence of the gradual change pattern based on the gradual change interval;

[0027] The first duration and the first frequency of occurrence are converted into a first risk score through a first mapping function, and the second duration and the second frequency of occurrence are converted into a second risk score through a second mapping function. A risk score is obtained based on a dynamic weighted combination of the first risk score and the second risk score, and the risk score is time-integrated to obtain a risk accumulation curve.

[0028] In an optional embodiment,

[0029] Calculating the local entropy of nodes at different scales of the mutant topology network, determining the optimal scale set based on the variation characteristics of the local entropy, and fusing the local entropies under the optimal scale set to obtain an enhanced topology entropy sequence includes:

[0030] Constructing a node neighborhood matrix in the mutant topology network, wherein the node neighborhood matrix represents the set of nodes that each node can reach within a specified number of steps, constructing a node feature vector including degree centrality and clustering coefficient based on the node neighborhood matrix, calculating node structural similarity, and using the node structural similarity to adaptively reconstruct network connections;

[0031] Calculating cumulative weights between nodes in the reconstructed network, where the cumulative weights are accumulated recursively at different numbers of steps, calculating transition probabilities between nodes based on the cumulative weights, and adaptively nonlinearly adjusting the transition probabilities according to the local density distribution of the network to obtain modified transition probabilities;

[0032] Based on the modified transition probability, the local entropy of the node is calculated at different scales to obtain a local entropy sequence, the multi-order differential characteristics and community evolution characteristics of the local entropy sequence are calculated, and a scale importance index is constructed;

[0033] An optimal scale set is determined according to the scale importance index, and the information redundancy of each scale in the optimal scale set is calculated based on the complementarity metric. The normalized inverse of the information redundancy is used as the fusion weight. The local entropy sequence under the optimal scale set is weightedly combined using the fusion weight to obtain an enhanced topological entropy sequence.

[0034] In an optional embodiment,

[0035] Calculate the temporal relationship between physiological indicators based on the risk accumulation curve, convert the temporal relationship into the indicator conduction delay, and draw a risk conduction path diagram based on the indicator conduction delay. Mark the time interval with the largest change slope in the risk conduction path diagram as the intervention time window, including:

[0036] Extracting local features of the risk accumulation curve at multiple time window scales, calculating a distance matrix between physiological indicators based on the local features, and calculating the temporal relationship of changes between the physiological indicators using a time warping algorithm with attention weights based on the distance matrix;

[0037] Constructing a hierarchical coupling network among the physiological indicators, recursively calculating direct coupling coefficients and indirect coupling coefficients in the hierarchical coupling network, and performing weighted correction on the temporal relationship of the changes to obtain the indicator conduction delay;

[0038] Constructing a risk transmission path diagram using the physiological indicators as nodes and the indicator transmission delay as edge weights, calculating the dynamic scores of the nodes in the risk transmission path diagram using a time-varying node importance algorithm, and determining the transmission path based on the dynamic scores and the indicator transmission delay;

[0039] An adaptive segmented planning algorithm is used on the conduction path to search for the maximum slope interval to obtain a candidate time window, and the information entropy of the slope sequence in the candidate time window is calculated to obtain the window confidence. When the window confidence is lower than a preset confidence threshold, the interval range is expanded along the conduction path to re-search the maximum slope interval, and the maximum slope interval that meets the window confidence is used as the final intervention time window.

[0040] In an optional embodiment,

[0041] Constructing a hierarchical coupling network among the physiological indicators, recursively calculating direct coupling coefficients and indirect coupling coefficients in the hierarchical coupling network, and weightedly correcting the temporal relationship of the changes to obtain the indicator conduction delay includes:

[0042] Sliding window segmentation is performed on the temporal relationship of changes between physiological indicators to obtain local temporal segments. The matching degree of the local temporal segments is calculated using a dynamic time warping algorithm to obtain a temporal similarity matrix. Multi-scale wavelet transform is performed on the local temporal segments to perform frequency domain decomposition. The phase synchronization of different frequency components is calculated to obtain a frequency coupling matrix. The temporal similarity matrix and the frequency coupling matrix are adaptively weighted and fused to obtain the coupling strength between indicators.

[0043] An initial hierarchical coupling network is constructed based on the coupling strength between indicators, conditional mutual information between network nodes is calculated, a causal entropy significance test is performed based on the conditional mutual information to screen stable coupling edges, the initial hierarchical coupling network is updated using the stable coupling edges to obtain an optimized hierarchical coupling network, and the coupling strength of network adjacent edges in the optimized hierarchical coupling network is extracted to obtain a direct coupling coefficient;

[0044] Performing a depth-first search on the optimized hierarchical coupling network to identify candidate conduction links between physiological indicators, calculating the information entropy of the candidate conduction links, setting an effective threshold based on the information entropy to screen effective conduction links, performing a recursive calculation on the effective conduction links, and taking the ratio of the cumulative effect of the coupling strength of each edge on the link to the path length as the indirect coupling coefficient;

[0045] A comprehensive coupling coefficient is calculated based on the direct coupling coefficient and the indirect coupling coefficient, and a nonlinear weighted correction is performed on the change timing relationship using the comprehensive coupling coefficient. The corrected change timing relationship is normalized and converted into a time unit to obtain an indicator conduction delay.

[0046] A second aspect of an embodiment of the present invention provides an AI-based heart failure risk assessment system, comprising:

[0047] The first unit is used to collect physiological indicators of the subject to be evaluated and construct them into a multidimensional time series data stream. The difference between adjacent data segments in the multidimensional time series data stream is calculated using a recursive segmentation algorithm. The time point where the difference is greater than a preset difference threshold is determined as a mutation point. The waveform features of the data segments before and after the mutation point are extracted to generate a mutation feature matrix.

[0048] The second unit is used to perform time series mapping on the mutation feature matrix, obtain the time interval sequence and amplitude sequence of the mutation features, identify periodic mutation patterns and gradual change patterns, convert the duration and frequency of periodic mutation patterns and gradual change patterns into risk scores, and draw the risk accumulation curve;

[0049] The third unit is used to calculate the temporal relationship between physiological indicators based on the risk accumulation curve, convert the temporal relationship into the indicator transmission delay, draw the risk transmission path diagram based on the indicator transmission delay, and mark the time interval with the largest change slope in the risk transmission path diagram as the intervention time window;

[0050] The fourth unit is used to determine the risk level of heart failure based on the time range of the intervention time window and the changing trend of the risk accumulation curve, and generate a risk assessment report containing the risk level and description of abnormal physiological indicators.

[0051] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0052] processor;

[0053] a memory for storing processor-executable instructions;

[0054] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0055] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0056] In this embodiment, the AI-based heart failure risk assessment method uses a recursive segmentation algorithm to identify mutation points in multidimensional time series data and extract a mutation feature matrix, which can accurately capture the key moments of changes in cardiac function and improve the accuracy and sensitivity of abnormal pattern recognition. By performing time-series mapping and pattern recognition on mutation features, this method can effectively distinguish between periodic mutations and gradual patterns, convert complex physiological signal changes into quantifiable risk scores, and realize dynamic assessment and visualization of heart failure risk, making it easier for medical personnel to intuitively understand the patient's condition. The innovative introduction of the concept of indicator conduction delay and the drawing of a risk conduction path diagram can not only reveal the intrinsic correlation between various physiological indicators, but also accurately mark the optimal intervention time window, providing a scientific basis for clinical intervention and effectively avoiding the problem of poor treatment effect caused by improper intervention timing in traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 Schematic diagram of the process of the AI-based heart failure risk assessment method according to an embodiment of the present invention;

[0058] Figure 2 This is a comparison chart of the performance of detecting mutation points in multi-dimensional physiological signals according to an embodiment of the present invention;

[0059] Figure 3 This is a performance comparison chart of the risk assessment and early warning system according to an embodiment of the present invention;

[0060] Figure 4 This is a comparison chart of delay calculation accuracy in embodiments of the present invention. DETAILED DESCRIPTION

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0062] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0063] Figure 1FIG. 1 is a flow chart of an AI-based heart failure risk assessment method according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0064] The physiological indicators of the subject to be evaluated are collected and constructed into a multidimensional time series data stream. The difference between adjacent data segments in the multidimensional time series data stream is calculated using a recursive segmentation algorithm. The time point where the difference is greater than a preset difference threshold is determined as the mutation point. The waveform features of the data segments before and after the mutation point are extracted to generate a mutation feature matrix.

[0065] Perform time series mapping on the mutation feature matrix to obtain the time interval sequence and amplitude sequence of the mutation features, identify periodic mutation patterns and gradual change patterns, convert the duration and frequency of periodic mutation patterns and gradual change patterns into risk scores, and draw a risk accumulation curve;

[0066] Calculate the temporal relationship between physiological indicators based on the risk accumulation curve, convert the temporal relationship into the indicator conduction delay, draw a risk conduction path diagram based on the indicator conduction delay, and mark the time interval with the largest change slope in the risk conduction path diagram as the intervention time window;

[0067] Based on the time range of the intervention time window and the changing trend of the risk accumulation curve, the heart failure risk level is determined, and a risk assessment report containing the risk level and description of abnormal physiological indicators is generated.

[0068] In an optional embodiment, collecting physiological indicators of the subject to be evaluated to construct a multidimensional time series data stream, calculating the difference between adjacent data segments in the multidimensional time series data stream using a recursive segmentation algorithm, and determining the time point when the difference is greater than a preset difference threshold as a mutation point includes:

[0069] Acquiring physiological indicators of the subject to be evaluated, the physiological indicators including cardiovascular data, respiratory data, and blood oxygen data; resampling the physiological indicators through cubic spline interpolation to construct a multidimensional time series data stream, and constructing the multidimensional time series data stream into a state matrix of time-indicator dimensions;

[0070] Initially dividing the state matrix from the midpoint to obtain two sub-data segments, moving the dividing position point by point along the time axis to calculate the Mahalanobis distance of the data segments on both sides to obtain a distance sequence, and performing eigenvalue decomposition to obtain a principal component weight vector, and using the principal component weight vector to weight the distance sequence to obtain a sub-data segment difference sequence;

[0071] A cumulative distribution function is obtained by statistically analyzing the probability distribution characteristics of the difference degree sequence of the sub-data segments, a difference threshold is determined on the cumulative distribution function according to a preset false alarm rate, a time point greater than the difference threshold is marked as an initial mutation point, an observation window is constructed with the initial mutation point as the center, and the mutual information between the signals of each dimension in the observation window is calculated to obtain a mutual information matrix, and the sum of the off-diagonal elements of the mutual information matrix is used as a coupling metric value;

[0072] A comprehensive difference is calculated based on the coupling metric value and the sub-data segment difference sequence, the initial mutation point is screened according to the comprehensive difference, and the time point when the comprehensive difference is greater than the difference threshold is determined as the final mutation point.

[0073] In order to accurately monitor the physiological state of the object to be evaluated, this embodiment provides a method for detecting mutation points of physiological indicators. This method calculates the difference between adjacent data segments in a multidimensional time series data stream through a recursive segmentation algorithm to achieve the identification and monitoring of physiological state mutations. First, the physiological indicators of the object to be evaluated are collected, including cardiovascular data, respiratory data and blood oxygen data. Cardiovascular data may include data such as heart rate and blood pressure; respiratory data may include data such as respiratory rate and respiratory depth; blood oxygen data may include data such as blood oxygen saturation and blood oxygen fluctuation range. For example, a patient is continuously monitored, and heart rate data (fluctuations of 60-80 times per minute), respiratory rate data (fluctuations of 14-18 times per minute) and blood oxygen saturation data (fluctuations within the range of 95%-99%) are collected.

[0074] The collected physiological indicator data is resampled using cubic spline interpolation to construct a multidimensional time series data stream. Because the sampling frequencies of various physiological indicators may be inconsistent, such as heart rate data being sampled every 1 second and respiration data every 2 seconds, to ensure temporal consistency, the original data is resampled using cubic spline interpolation so that all indicators have the same sampling time points. After resampling, the multidimensional time series data stream is constructed as a state matrix with a time-indicator dimension, where the rows represent time points and the columns represent different physiological indicators. The constructed state matrix is then initially partitioned at its midpoint to produce two sub-segments. For example, for a state matrix containing 1000 time points, the initial partition is performed at the 500th time point, resulting in two sub-segments, each containing 500 time points. Subsequently, the partition position is moved point by point along the time axis, and the Mahalanobis distance between the two data segments is calculated. The specific operation is as follows: Assuming time point t is the current partition point, calculate the Mahalanobis distance between the data segment from time point 1 to t and the data segment from time point t+1 to the end, obtaining a distance value. Then move the partition point to t+1 and repeat the above calculation process to obtain a distance sequence. For example, for a state matrix of length 1000, 998 Mahalanobis distance values will be obtained (from position 2 to position 999).

[0075] Perform eigenvalue decomposition on the resulting distance sequence to obtain the principal component weight vectors. Eigenvalue decomposition can reveal the main change patterns in the distance sequence and extract the principal component weight vectors. The distance sequence is weighted using the principal component weight vectors to obtain a sub-segment difference sequence. The weight vectors reflect the contribution of each time point to the overall difference. The weighted difference sequence more accurately represents the degree of state change.

[0076] The probability distribution characteristics of the difference sequence of the sub-data segments are statistically analyzed to obtain the cumulative distribution function. According to the preset false alarm rate (such as 0.05), the difference threshold is determined on the cumulative distribution function. For example, if the false alarm rate is set to 0.05, the difference threshold is the 95% quantile of the difference sequence. The time point greater than this threshold is marked as the initial mutation point. For each initial mutation point, an observation window is constructed with it as the center. The size of the observation window can be set according to the actual application scenario, such as a window of 10 time points before and after, for a total of 21 time points. The mutual information between the signals of each dimension in the observation window is calculated to obtain the mutual information matrix. Each element in the mutual information matrix represents the mutual information value between the signals of two dimensions, reflecting the correlation between the signals. For example, for data containing three dimensions of heart rate, respiratory rate and blood oxygen saturation, the mutual information matrix is a 3×3 matrix.

[0077] The sum of the off-diagonal elements of the mutual information matrix is used as the coupling metric. The coupling metric reflects the degree of coupling between signals in each dimension within the observation window; larger values indicate more coordinated changes in the signals in each dimension. The combined metric is combined with the sub-segment difference sequence to calculate the comprehensive difference. The comprehensive difference can be obtained using a weighted average. For example, setting the weight of the coupling metric to 0.4 and the weight of the sub-segment difference to 0.6 allows the weighted calculation to yield the comprehensive difference.

[0078] Initial mutation points are screened based on the comprehensive difference, and the time points where the comprehensive difference exceeds the difference threshold are determined as the final mutation points. For example, in a patient's data analysis, 15 mutation points were initially detected among 1,000 time points. After comprehensive difference screening, 8 time points were determined as the final mutation points. These time points correspond to significant changes in the patient's physiological state, such as rest after strenuous exercise, mood swings, or the onset of drug effects.

[0079] In this embodiment, through the construction of cubic spline interpolation resampling and multidimensional time series data streams, the scheme can effectively handle the temporal nature and complexity of physiological data, ensuring the smoothness and continuity of the data. On this basis, the method of combining Mahalanobis distance with principal component analysis can quantify the differences between time periods and accurately identify potential mutation points. In addition, by calculating the mutual information between signals and combining the coupling metric value, the assessment accuracy of the correlation between signals of different dimensions can be improved, ensuring that the identification of mutation points is more accurate and reliable. Finally, by screening mutation points based on comprehensive differences, the scheme can effectively reduce the false alarm rate and achieve efficient and accurate detection of physiological data anomalies.

[0080] Figure 2 This is a comparison chart of the performance of detecting mutation points in multi-dimensional physiological signals according to an embodiment of the present invention. Figure 2 As shown in the figure, the performance comparison of different mutation point detection methods is shown. The horizontal axis represents the false alarm rate, and the vertical axis represents the detection rate. The closer the curve is to the upper left corner, the better the performance. This technical solution shows the highest detection rate at all false alarm rate levels, especially when the false alarm rate is 0.2, the detection rate reaches 0.85, which is marked as the optimal working point. When the false alarm rate is 0.6, the detection rate of this technical solution is as high as 96.3%, while the detection rates of traditional spectral analysis method and traditional threshold method under the same conditions are only 82.7% and 89.1% respectively. The shaded area in the figure intuitively shows the performance improvement of this technical solution compared with the traditional method. This significant advantage mainly comes from the innovative algorithm of combining Mahalanobis distance with principal component weighting in this technical solution, and the method of introducing mutual information matrix to calculate coupling degree to further screen mutation points. Compared with traditional spectral analysis methods (such as Fourier transform spectrum analysis) and simple threshold methods (such as fixed threshold or adaptive threshold method), this scheme has achieved a qualitative leap in the balance between false alarm rate control and detection rate, especially the detection capability under low false alarm rate conditions is more outstanding, which has important clinical value for the accurate identification of key mutation points in physiological signals.

[0081] In an optional embodiment, extracting waveform features of data segments before and after a mutation point to generate a mutation feature matrix includes:

[0082] The fluctuation characteristics of the data segments before and after the mutation point are calculated according to the location of the mutation point to obtain a volatility sequence, and the fluctuation trend characteristics are obtained by performing a segmented cumulative calculation. The adaptive window length is determined based on the fluctuation trend characteristics, and the waveform feature extraction interval before and after the mutation point is determined using the adaptive window length.

[0083] In the waveform feature extraction interval, the data segments before and after the mutation point are respectively subjected to time-frequency transformation to obtain a time-frequency spectrum, the energy distribution of the time-frequency spectrum is calculated to obtain a time-frequency energy sequence, the time-frequency energy sequence is weightedly fused in combination with the volatility sequence to obtain a fused feature sequence, and the multi-scale entropy value of the fused feature sequence is calculated to obtain a complexity feature;

[0084] The volatility sequence, time-frequency energy sequence and complexity feature are combined in time sequence to obtain a feature vector, an initial feature matrix is constructed according to the feature vector, the discriminant contribution of each feature in the initial feature matrix is calculated to obtain the feature weight, and the feature weight is used to perform weighted reconstruction on the initial feature matrix to generate a mutation feature matrix.

[0085] This embodiment provides a method for extracting waveform features from data segments before and after a mutation point and generating a mutation feature matrix. This method primarily involves determining an adaptive window length, extracting waveform features, and constructing a feature matrix. Specifically, the original data sequence containing the mutation point is first acquired, assuming it is a sequence X of length N, with the mutation point at position p. For example, a voltage waveform in a power system can be acquired over 10 seconds at a sampling rate of 10 kHz, with a total length of 100,000 sampling points. The mutation point is located at the 45,000th sampling point.

[0086] Calculate the volatility sequence of the data before and after the mutation point. Take M data points before and after the mutation point, creating the sequences X(pM:p-1) and X(p:p+M-1), respectively, where M can be N / 10. For each subsequence, calculate the absolute value of the difference between adjacent data points and normalize them to generate volatility sequences V1 and V2. For example, for voltage data, set M=10,000 to obtain a volatility sequence of 10,000 points before and after the mutation point.

[0087] Perform a segmented cumulative calculation on the volatility series to obtain the volatility trend characteristics. Divide the volatility series V1 and V2 into K segments of equal length (e.g., K = 10), calculate the cumulative sum of the volatility within each segment, and form the volatility trend series T1 and T2. For example, divide a 10,000-point volatility series into 10 segments of 1,000 points each, calculate the cumulative sum of the volatility within each segment, and obtain a volatility trend series of length 10.

[0088] The adaptive window length is determined based on the fluctuation trend characteristics. The rate of change of the fluctuation trend sequences T1 and T2 is calculated. The L segments with the largest rate of change (e.g., L = 3) are selected and their lengths are added together to form the adaptive window length W. In a real-world example, if the rates of change in segments 2, 5, and 8 are the largest and each segment is 1,000 points long, the adaptive window length W is 3,000 points.

[0089] To extract waveform features within a defined adaptive window length, first determine the waveform feature extraction interval. Centered around the mutation point p, W points are taken forward and W points are taken backward, forming the feature extraction interval [pW, p+W]. For example, when W = 3,000, the feature extraction interval is [42,000, 48,000], containing a total of 6,000 data points. A time-frequency transform is then performed on the data segments before and after the mutation point. A short-time Fourier transform is performed on the data in the intervals [pW, p-1] and [p, p+W], respectively. The window length is set to W / 10, and the window overlap is 50%. This yields two time-frequency spectra, S1 and S2. For example, with a window length of 300 points, the resulting time-frequency spectrum has 19 time points in the time dimension and 150 frequency points in the frequency dimension.

[0090] Calculate the energy distribution of the time-frequency spectrum. For each time-frequency spectrum S1 and S2, calculate the sum of the energy of each frequency band along the frequency dimension to form time-frequency energy sequences E1 and E2. These two sequences reflect the frequency band energy distribution characteristics of the data at different time points. For example, a 19-point energy sequence is obtained, where each value represents the total energy of all frequency bands at the corresponding time point. Combined with the volatility sequence, the time-frequency energy sequence is weightedly fused. Resample the volatility sequences V1 and V2 to the same length as E1 and E2, and perform a point-to-point multiplication operation to obtain the fused feature sequences F1 and F2. For example, resample a 3,000-point volatility sequence to 19 points and multiply it with the energy sequence at the corresponding position to obtain the fused feature sequence.

[0091] Calculate the multi-scale entropy of the fused feature sequence. Calculate the sample entropy for F1 and F2 separately, set the scale factor to 1 to 5, the embedding dimension to 2, and the similarity tolerance to 0.15 multiplied by the sequence standard deviation. This yields the multi-scale entropy sequences H1 and H2 as complexity features. For example, calculate the entropy at five scales for each fused sequence, resulting in a complexity feature of length 5.

[0092] The volatility sequence, time-frequency energy sequence, and complexity features are combined in time sequence to form a feature vector. For the data segments before and after the mutation point, V1, E1, and H1 are combined with V2, E2, and H2, respectively, to obtain feature vectors FV1 and FV2. For example, concatenating a volatility sequence of length 19, an energy sequence of length 19, and a complexity feature of length 5 yields a feature vector of length 43. Feature vectors FV1 and FV2 are merged to construct the initial feature matrix FM. Each row of the matrix corresponds to a mutation sample, and each column corresponds to a feature dimension. For example, for 100 mutation samples, the resulting initial feature matrix size is 100 × 86 (43 × 2 feature dimensions per sample).

[0093] Calculate the discriminant contribution of each feature in the initial feature matrix. Use the information gain method to evaluate the contribution of each feature column to distinguishing different types of mutations and generate a feature weight vector W. For example, for an 86-dimensional feature, calculate 86 corresponding weight values, where a larger value indicates a greater contribution of the feature to mutation type discrimination. Use the feature weights to perform a weighted reconstruction of the initial feature matrix. Perform a dot product operation on the feature weight vector W and each column of the initial feature matrix FM to obtain the weighted mutation feature matrix WFM. For example, multiply the eigenvalue of column i by the corresponding weight W[i] to generate the reconstructed feature matrix.

[0094] In this embodiment, through analysis of fluctuation characteristics and adaptive window determination of fluctuation trend features, the dynamic changes of the waveform can be effectively identified and the extraction interval of waveform features can be precisely defined. The application of time-frequency transformation can capture the changes in data across different frequency bands. The time-frequency energy series is combined with the volatility series for weighted fusion, making feature extraction more refined, thereby more accurately reflecting the essential characteristics of the data. The calculation of multi-scale entropy values further enhances the ability to capture changes in complexity and makes the identification of mutation points more sensitive. Through the construction and weighted reconstruction of feature vectors, the solution can optimize the discriminative ability of features, improve the model's sensitivity and accuracy to mutation points, and thus effectively improve the accuracy and reliability of mutation detection.

[0095] In an optional embodiment, performing time series mapping on the mutation feature matrix, obtaining the time interval sequence and amplitude sequence of the mutation features, identifying the periodic mutation pattern and the gradual change pattern, converting the duration and occurrence frequency of the periodic mutation pattern and the gradual change pattern into risk scores, and drawing the risk accumulation curve includes:

[0096] Performing time series mapping on the mutation feature matrix to obtain a time interval sequence and an amplitude sequence, and constructing a mutation topology network based on the time interval sequence and amplitude sequence using an adaptive weight function, where nodes represent mutation points and the connection weights between nodes are determined by the time interval and amplitude.

[0097] Calculating the local entropy of nodes at different scales of the mutant topology network, determining the optimal scale set according to the variation characteristics of the local entropy, and fusing the local entropies under the optimal scale set to obtain an enhanced topological entropy sequence;

[0098] Dynamically adjusting the calculation window size based on the rate of change of the enhanced topological entropy sequence, calculating the autocorrelation function and power spectral density of the enhanced topological entropy sequence within the calculation window, extracting periodic mutation characteristics, determining a first duration and a first frequency of occurrence of a periodic mutation pattern based on the periodic mutation characteristics, performing piecewise polynomial fitting on the enhanced topological entropy sequence to identify a gradual change interval, and determining a second duration and a second frequency of occurrence of the gradual change pattern based on the gradual change interval;

[0099] The first duration and the first frequency of occurrence are converted into a first risk score through a first mapping function, and the second duration and the second frequency of occurrence are converted into a second risk score through a second mapping function. A risk score is obtained based on a dynamic weighted combination of the first risk score and the second risk score, and the risk score is time-integrated to obtain a risk accumulation curve.

[0100] This example analyzes mutation signatures during heart failure risk assessment, identifying mutation patterns and quantifying risk by constructing a mutation topology network. First, a time series mapping is performed on the mutation signature matrix to obtain the time interval sequence and amplitude sequence of the mutation signatures. The time interval sequence represents the time difference between adjacent mutation points, while the amplitude sequence represents the magnitude of change in the physiological indicator corresponding to each mutation point.

[0101] When constructing a mutation topology network based on a time interval sequence and an amplitude sequence, an adaptive weight function is used to determine the connection weights between nodes in the network. Specifically, the connection weight between any two mutation points is obtained by weightedly combining the normalized value of the time interval and the normalized value of the amplitude difference. For example, for a mutation point in the heart rate metric, when the time interval between two mutation points is ten minutes and the amplitude difference is twenty beats per minute, the time interval can be normalized to the range of zero to one, and the amplitude difference can also be normalized to the same range. The two normalized values are then combined using the adaptive weight coefficient to obtain the connection weight between the two mutation points.

[0102] In the constructed mutation topology network, the local entropy of nodes is calculated by setting different observation scales. Local entropy reflects the degree of clustering and correlation strength of mutation points at different time scales. By analyzing the trend of local entropy changes with scale, the optimal scale set that best reflects the mutation characteristics can be determined. For example, for a mutation point in the blood pressure indicator, the local entropy changes most significantly at the time scales of five minutes, thirty minutes, and two hours. These three scales constitute the optimal scale set.

[0103] The enhanced topological entropy sequence is obtained by weighted fusion of multiple local entropy sequences at the optimal scale. The weight coefficients in the fusion process can be dynamically adjusted based on the significance of the local entropy at each scale. For example, for the respiratory rate metric, if the local entropy fluctuations at the five-minute scale are the most dramatic, a larger weight coefficient will be assigned to that scale.

[0104] The computation window size is dynamically adjusted based on the changing characteristics of the enhanced topological entropy sequence. When the sequence changes dramatically, a smaller computation window is used to capture rapidly changing features; when the sequence is relatively stable, a larger computation window is used to extract long-term trends. Within the specified computation window, the autocorrelation function and power spectral density of the enhanced topological entropy sequence are calculated.

[0105] Periodic mutation features are extracted by analyzing the periodicity of the autocorrelation function and the peak characteristics of the power spectral density. For mutation sequences with significant periodicity, the period length can be determined by the position of the first non-zero peak in the autocorrelation function, and the mutation strength can be determined by the main frequency component of the power spectral density. For example, when analyzing blood oxygen saturation indicators, if a clear periodic mutation is found every four hours, and this periodic pattern repeats six times within 24 hours, the duration of the periodic mutation pattern can be recorded as 24 hours and the frequency of occurrence can be recorded as six times.

[0106] A piecewise polynomial fit is performed on the enhanced topological entropy sequence, and the gradient intervals are identified by analyzing the slope changes of the fitted curve. Within each gradient interval, the average slope and duration of the fitted curve are calculated, and the number of occurrences of the gradient interval is counted. For example, for the cardiac ejection fraction metric, if a gradient process is detected that lasts for eight hours and has an average rate of decline of 0.5 percent per hour, and a similar gradient process occurs three times during the observation period, the duration of the gradient pattern is recorded as 24 hours, and the frequency of occurrence is recorded as three times.

[0107] The duration and frequency of the periodic mutation pattern are input into the first mapping function to obtain the risk score corresponding to the periodic mutation. The first mapping function adopts a piecewise continuous nonlinear function form, with different mapping slopes in different duration and frequency intervals. For example, when the duration of the periodic mutation exceeds twenty-four hours or the frequency exceeds eight times a day, the slope of the mapping function will increase significantly, resulting in a higher risk score. Similarly, the duration and frequency of the gradual change pattern are input into the second mapping function to obtain the risk score corresponding to the gradual change pattern. The second mapping function also adopts a piecewise nonlinear design, but its slope change characteristics are different from those of the first mapping function to reflect the different ways in which the gradual change pattern and the periodic mutation pattern affect risk.

[0108] Based on the changing characteristics of physiological indicators within a time window, the weight coefficients of the periodic mutation risk score and the gradual change pattern risk score are dynamically adjusted. When dramatic periodic fluctuations are observed, the weight of the periodic mutation risk score is increased; when a sustained gradual change trend is observed, the weight of the gradual change pattern risk score is increased. The weighted sum of the two risk scores is used to obtain the comprehensive risk score.

[0109] The risk accumulation curve is generated by integrating the comprehensive risk score over time. The integration process uses an adaptive step size design, with smaller integration steps used in intervals with drastic risk score fluctuations and larger integration steps used in intervals with relatively stable risk scores. This resulting risk accumulation curve accurately reflects the accumulation of heart failure risk over time. For example, if frequent heart rate fluctuations and sustained blood pressure drops are observed during the nighttime monitoring of a patient with congestive heart failure, the risk accumulation rate during this period will be significantly higher than during other periods.

[0110] This technical solution comprehensively analyzes the sudden and gradual changes in physiological indicators, enabling precise quantification and dynamic tracking of heart failure risk. This solution is particularly suitable for heart failure patients requiring long-term continuous monitoring, enabling timely identification of abnormal patterns in the accumulation of risk and providing a basis for clinical intervention decisions.

[0111] Traditional techniques typically detect mutation points through simple thresholding or static feature extraction. However, these methods tend to overlook complex patterns of change in the data, resulting in insufficient sensitivity to mutations. This application introduces an adaptive weighting function to construct a mutation topology network. By combining time interval and amplitude information, it dynamically adjusts the weighting of mutation points, enabling more accurate quantification of the correlations between mutation points. Furthermore, the application of local entropy and enhanced topological entropy sequences captures detailed variations in mutation data at multiple scales, effectively identifying periodic and gradual mutation patterns. Existing techniques typically rely solely on fixed-scale or single-frequency domain analysis, which struggles to fully reveal the diversity and complexity of data. However, this application optimizes scale selection and dynamic window adjustment, enabling consistent and stable identification of periodic and gradual features in a changing environment, improving pattern recognition accuracy. Risk score calculation uses a multi-level mapping function to transform the duration and frequency of periodic and gradual patterns. Combined with dynamic weighting, this approach comprehensively reflects the severity and potential risk of the mutation. Compared to traditional risk assessment methods, this solution provides a more comprehensive and accurate risk assessment, avoiding overly simplistic approaches and improving the reliability and adaptability of mutation point detection and risk assessment.

[0112] Figure 3 This is a performance comparison chart of the risk assessment and early warning system according to an embodiment of the present invention. Figure 3The figure shows a performance comparison of risk assessment and early warning systems. The top half of the chart compares five key metrics: In terms of precision, this solution achieved 95.4%, significantly higher than the threshold detection method (65.0%) and the statistical clustering method (72.5%). In terms of recall, this solution achieved 93.2%, compared to 60.8% for the threshold detection method and 68.5% for the statistical clustering method. In terms of F1 score, this solution's advantage was even more pronounced, reaching 96.2%, surpassing the threshold detection method (62.8%) and the statistical clustering method (70.4%) by over 25 percentage points. In terms of the key metric of early warning time, this solution was able to issue warnings 186.5 seconds earlier, 3.2 times that of the threshold detection method (58.2 seconds) and twice that of the statistical clustering method (92.4 seconds). In terms of false alarm rate, this solution achieved only 3.8%, compared to 28.5% for the threshold detection method and 19.2% for the statistical clustering method.

[0113] In an optional embodiment, calculating the local entropy of nodes at different scales of the mutant topology network, determining an optimal scale set based on the variation characteristics of the local entropy, and fusing the local entropies at the optimal scale set to obtain an enhanced topology entropy sequence includes:

[0114] Constructing a node neighborhood matrix in the mutant topology network, wherein the node neighborhood matrix represents the set of nodes that each node can reach within a specified number of steps, constructing a node feature vector including degree centrality and clustering coefficient based on the node neighborhood matrix, calculating node structural similarity, and using the node structural similarity to adaptively reconstruct network connections;

[0115] Calculating cumulative weights between nodes in the reconstructed network, where the cumulative weights are accumulated recursively at different numbers of steps, calculating transition probabilities between nodes based on the cumulative weights, and adaptively nonlinearly adjusting the transition probabilities according to the local density distribution of the network to obtain modified transition probabilities;

[0116] Based on the modified transition probability, the local entropy of the node is calculated at different scales to obtain a local entropy sequence, the multi-order differential characteristics and community evolution characteristics of the local entropy sequence are calculated, and a scale importance index is constructed;

[0117] An optimal scale set is determined according to the scale importance index, and the information redundancy of each scale in the optimal scale set is calculated based on the complementarity metric. The normalized inverse of the information redundancy is used as the fusion weight. The local entropy sequence under the optimal scale set is weightedly combined using the fusion weight to obtain an enhanced topological entropy sequence.

[0118] The present invention provides a method for calculating node local entropy at different scales in a mutant topology network. First, a node neighborhood matrix is constructed in the mutant topology network. Taking a social network with 10 nodes as an example, the node neighborhood matrix is constructed by recording the set of nodes that each node can reach within different steps. For example, for node 1, the set of nodes reachable within 1 step is {2, 3, 5}, the set of nodes reachable within 2 steps is {2, 3, 4, 5, 7}, and so on to construct a complete node neighborhood matrix.

[0119] Node feature vectors are constructed based on the node neighborhood matrix, including degree centrality and clustering coefficient. Degree centrality is obtained by calculating the number of direct connections between nodes. For example, the degree centrality of node 1 is 3. The clustering coefficient is obtained by calculating the connection density between the node's neighbors. For example, the clustering coefficient of node 1 is 0.67, which represents the ratio of the actual number of connections between its neighboring nodes to the total number of possible connections. Node structural similarity is calculated, using cosine similarity to measure the degree of structural similarity between nodes. For nodes i and j, their feature vectors are extracted, and the cosine value of the angle between the vectors is calculated as the similarity. For example, the structural similarity between nodes 1 and 2 is 0.85, indicating that they have similar network positions and connection patterns.

[0120] The network connections are adaptively reconstructed using the node structure similarity. When the structural similarity of two nodes exceeds the threshold of 0.8, an additional weight of 0.2 is added to the original connection weight; when the similarity is lower than 0.3, the original connection weight is reduced by 10%. In this way, the connection strength between nodes with similar structures is enhanced, and the connection between nodes with significant differences is weakened, forming a network structure that is more in line with the actual interaction logic. The cumulative weights between nodes are calculated in the reconstructed network. First, the cumulative weight matrix is initialized. For directly connected nodes, the cumulative weight is equal to the connection weight between them; for indirectly connected nodes, it is accumulated recursively at different numbers of steps. For example, the 2-step cumulative weight from node 1 to node 4 is 0.35, which represents the cumulative connection strength through all possible 2-step paths.

[0121] The transition probabilities between nodes are calculated based on the cumulative weights. For node i, the transition probabilities from i to all other nodes j are calculated, that is, the cumulative weights starting from i are normalized. For example, the transition probabilities from node 1 to each node are distributed as follows: node 2 (0.25), node 3 (0.2), node 4 (0.15), and so on.

[0122] The modified transition probabilities are obtained by adaptively adjusting the transition probabilities nonlinearly based on the local density distribution of the network. First, the local density of each node is calculated, that is, the density of connections within the node's neighborhood. Then, an adjustment function is designed: for nodes in high-density areas, the probability of their transition to low-density areas is reduced; for nodes in low-density areas, the probability of their transition to high-density areas is increased. For example, if node 1 is located in a high-density area (density value 0.8), its transition probability to node 6 in a low-density area is adjusted from 0.1 to 0.08.

[0123] The local entropy of a node is calculated at different scales based on the modified transition probabilities. The scale parameter controls the extent to which the random walk explores the network, with 10 different scales ranging from 1 to 10. At each scale, the local entropy of the node is calculated, which represents the complexity of the structure surrounding the node. For example, the local entropy of node 1 is 1.58 at scale 2 and 2.36 at scale 5.

[0124] We calculated the multi-order differential characteristics of the local entropy sequence and the community evolution characteristics to construct a scale importance index. The multi-order differential characteristics were obtained by calculating the rate of change of local entropy at different scales, while the community evolution characteristics were obtained by monitoring the stability of the community structure at different scales. For example, when the scale changes from 3 to 4, the average local entropy of the nodes changes by 15%, and the stability of the community structure decreases by 12%. The overall importance index for this scale transition is 0.78.

[0125] The optimal scale set is determined based on the scale importance index. The scales corresponding to the scale transition points with the top 30% importance index are selected. For example, in this example, the optimal scale set selected is {2, 4, 7, 9}.

[0126] The information redundancy of each scale in the optimal scale set is calculated based on the complementarity metric. Information redundancy measures the degree of overlap between the local entropy information at one scale and the information at other scales. For example, the average mutual information between scale 2 and the other optimal scales is 0.4, so its information redundancy is 0.4. Similarly, the information redundancy of scale 4 is 0.35; scale 7 is 0.3; and scale 9 is 0.25. The normalized inverse of the information redundancy is used as the fusion weight. The calculation process is: first take the inverse to obtain {2.5, 2.86, 3.33, 4}, then normalize it to obtain the final fusion weight {0.2, 0.225, 0.26, 0.315}.

[0127] The enhanced topological entropy sequence is obtained by weightedly combining the local entropy sequences at the optimal scale set using the fusion weights. For example, for node 1, its local entropy values at the scale set {2, 4, 7, 9} are {1.58, 2.36, 2.85, 3.12}, respectively. After weighted combination, the enhanced topological entropy value is 2.58. This enhanced topological entropy sequence integrates structural information at multiple scales and can more accurately represent the importance and influence of nodes in the network.

[0128] Based on the above technical solution, it is possible to achieve refined analysis and enhanced feature extraction of mutation topology networks. Traditional network analysis methods mostly rely on single-scale or fixed node structural features for analysis, which easily ignores the multi-dimensional and complex relationships in the data, resulting in inaccurate capture of mutation features in the network. This application constructs a node neighborhood matrix and node feature vectors, combines degree centrality and clustering coefficient, and can more comprehensively reflect the structural relationship between nodes, and then optimize network connections through adaptive reconstruction, making the overall structure of the network more flexible and adaptable.

[0129] On this basis, the recursive accumulation of node weights and the calculation of modified transition probabilities can more accurately represent the strength of relationships and information flow between nodes, compensating for the shortcomings of traditional methods that ignore dynamic changes and nonlinear adjustments. In addition, the introduction of multi-order differential characteristics of local entropy sequences and community evolution characteristics enables the identification of mutation points to go beyond single-level analysis and comprehensively consider node evolution and information changes from multiple dimensions and levels.

[0130] Compared to existing technologies, this approach dynamically selects the optimal scale set and introduces weighted adjustments to information redundancy, enabling the fused local entropy sequence to more accurately reflect the core characteristics of the mutation while avoiding interference from redundant information. The enhanced topological entropy sequence, derived through weighted combination, comprehensively considers the diversity of mutational features at different scales, improving adaptability to complex data structures and enhancing the ability to recognize mutation patterns.

[0131] In an optional embodiment, the temporal relationship between physiological indicators is calculated based on the risk accumulation curve, the temporal relationship is converted into an indicator conduction delay, a risk conduction path diagram is drawn based on the indicator conduction delay, and the time interval with the largest change slope is marked in the risk conduction path diagram as the intervention time window, including:

[0132] Extracting local features of the risk accumulation curve at multiple time window scales, calculating a distance matrix between physiological indicators based on the local features, and calculating the temporal relationship of changes between the physiological indicators using a time warping algorithm with attention weights based on the distance matrix;

[0133] Constructing a hierarchical coupling network among the physiological indicators, recursively calculating direct coupling coefficients and indirect coupling coefficients in the hierarchical coupling network, and performing weighted correction on the temporal relationship of the changes to obtain the indicator conduction delay;

[0134] Constructing a risk transmission path diagram using the physiological indicators as nodes and the indicator transmission delay as edge weights, calculating the dynamic scores of the nodes in the risk transmission path diagram using a time-varying node importance algorithm, and determining the transmission path based on the dynamic scores and the indicator transmission delay;

[0135] An adaptive segmented planning algorithm is used on the conduction path to search for the maximum slope interval to obtain a candidate time window, and the information entropy of the slope sequence in the candidate time window is calculated to obtain the window confidence. When the window confidence is lower than a preset confidence threshold, the interval range is expanded along the conduction path to re-search the maximum slope interval, and the maximum slope interval that meets the window confidence is used as the final intervention time window.

[0136] Exemplarily, local features of the risk accumulation curve are extracted at multiple time window scales. Specifically, 5 minutes, 15 minutes, 30 minutes, 60 minutes, and 120 minutes can be selected as different time window scales. For the risk accumulation curve of each physiological indicator, statistical features are calculated separately at each time window scale, including mean, variance, kurtosis, skewness, maximum value, minimum value, and volatility. For example, for the blood pressure indicator, a mean of 120 mmHg, variance of 5.2, kurtosis of 2.1, skewness of 0.8, maximum value of 135 mmHg, minimum value of 105 mmHg, and volatility of 0.12 can be obtained at a 5-minute window; corresponding characteristic values such as a mean of 118 mmHg and a variance of 4.8 can be obtained at a 15-minute window. Similarly, the same feature extraction process is performed on other physiological indicators such as blood sugar and heart rate.

[0137] The distance matrix between physiological indicators is calculated based on local features. The Euclidean distance is used to calculate the distance between the characteristic vectors of each indicator at different time window scales to form the original distance matrix. For example, the distance between blood pressure and heart rate indicators in a 5-minute window is 2.35, and the distance in a 15-minute window is 2.89. Then, the attention weight mechanism is introduced to assign different weights to distances at different time window scales. Specifically, by calculating the variance of the distance at each window scale, a higher weight is assigned to the window with a larger variance. For example, the weight of a 5-minute window is 0.15, the weight of a 15-minute window is 0.20, the weight of a 30-minute window is 0.25, the weight of a 60-minute window is 0.30, and the weight of a 120-minute window is 0.10. Finally, the comprehensive distance matrix is obtained by weighted averaging.

[0138] Based on the distance matrix, a time warping algorithm with attention weights is used to calculate the temporal relationship between physiological indicators. This algorithm first constructs a time-aligned path, then uses an attention mechanism to identify key alignment points, ultimately determining the temporal relationship between indicators. For example, the calculation results show that blood pressure changes precede heart rate changes by 8.5 minutes, heart rate changes precede blood oxygen changes by 12.3 minutes, and blood oxygen changes precede body temperature changes by 5.2 minutes. Next, a hierarchical coupling network is constructed between physiological indicators. This network consists of three layers: a single-indicator internal fluctuation layer, a direct coupling layer between indicators, and an indirect coupling layer between indicators. During this construction process, network edges are first established based on the correlation between indicator changes, and then edge weights are assigned. For example, the direct coupling coefficient between blood pressure and heart rate is 0.78, and the direct coupling coefficient between heart rate and blood oxygen is 0.65.

[0139] Direct and indirect coupling coefficients are recursively calculated within the hierarchical coupling network. The direct coupling coefficient is calculated using the Pearson correlation coefficient. For example, the direct coupling coefficient between blood pressure and heart rate is 0.78. The indirect coupling coefficient is calculated by multiplying the paths. For example, the coupling coefficient for the indirect effect of blood pressure on blood oxygen through heart rate is 0.78 × 0.65 = 0.507. The time series relationship of the changes is then weighted and corrected to obtain the indicator conduction delay. After correction, the conduction delay from blood pressure change to heart rate change is 7.8 minutes, and the conduction delay from heart rate change to blood oxygen change is 11.5 minutes.

[0140] A risk transmission path diagram was constructed using physiological indicators as nodes and indicator transmission delays as edge weights. In this diagram, edge direction represents the direction of risk transmission, and edge weights represent the transmission delay. For example, the edge weight from the blood pressure node to the heart rate node is 7.8 minutes, indicating that it takes 7.8 minutes for risk to be transmitted from blood pressure to heart rate. A time-varying node importance algorithm was used to calculate the dynamic scores of nodes in the risk transmission path diagram. This algorithm comprehensively considers the node's connectivity, betweenness centrality, and closeness centrality, and introduces a time decay factor to calculate the importance scores of nodes at different time points. For example, 30 minutes before onset, the blood pressure node scored 0.85, and the heart rate node scored 0.72. However, 15 minutes before onset, the blood pressure node scored 0.79, and the heart rate node scored 0.83, indicating that over time, the heart rate node has become more important than the blood pressure node.

[0141] The transmission path is determined based on the dynamic score and indicator transmission delay. The path with the highest total score is selected as the main risk transmission path, such as: blood pressure → heart rate → blood oxygen → body temperature, with a total score of 3.19.

[0142] An adaptive segmented planning algorithm was used to search for the interval with the maximum slope along the identified conduction pathway, identifying candidate time windows. The algorithm first divided the cumulative risk curve along the conduction pathway into segments, calculated the slope for each segment, and then used dynamic programming to identify the continuous interval with the maximum slope. For example, in the blood pressure to heart rate conduction pathway, the slope was found to be the highest between 25 and 20 minutes before onset, at 0.045 / minute.

[0143] The information entropy of the slope sequence within the candidate time window is calculated to obtain the window confidence. The lower the information entropy, the more stable the slope change and the higher the window confidence. For example, the information entropy of the slope sequence in the interval 25-20 minutes before the onset of the disease is 0.28, and the corresponding window confidence is 0.72. The window confidence threshold is set to 0.7. When the window confidence is lower than the threshold, the interval range is expanded along the conduction path to re-search the maximum slope interval. For example, if the confidence of a candidate window is 0.65, which is lower than the threshold of 0.7, the search range is expanded from the original 5 minutes to 7 minutes, and the maximum slope interval is recalculated to be 27-20 minutes before the onset of the disease, with a corresponding confidence of 0.75, which meets the threshold requirement. The maximum slope interval that meets the window confidence is used as the intervention time window. In this embodiment, the intervention time window determined is 27-20 minutes before the onset of the disease, which indicates that medical intervention within this time window may be most effective.

[0144] Based on the above technical solution, dynamic monitoring of changes in physiological indicators and precise identification of intervention opportunities can be achieved. Traditional technologies usually use static methods to analyze the relationship between physiological indicators, which easily ignores the temporal nature and dynamic conduction effects of changes between indicators, resulting in inaccurate judgment of intervention timing. This application extracts local features of the risk accumulation curve and combines it with a time warping algorithm with attention weights to more accurately capture the temporal relationship between changes in physiological indicators, thereby revealing the time delay and mutual influence of indicator changes.

[0145] Furthermore, the constructed hierarchical coupling network enhances the modeling capability of complex interactions between indicators by recursively calculating direct and indirect coupling coefficients, making the calculation of indicator transmission delay more comprehensive and accurate. Compared with traditional simple distance calculations, this method can better reflect the multi-level coupling relationships between physiological indicators. By constructing a risk transmission pathway diagram and applying a time-varying node importance algorithm, the score of each node can be dynamically calculated to help identify the main pathways of risk transmission, avoiding the shortcomings of previous reliance on fixed models or static structures.

[0146] To identify intervention time windows, this solution incorporates an adaptive segmented planning algorithm and information entropy calculation, enabling the identification of critical moments in risk transmission based on changes in slope. This process overcomes the limitations of traditional methods that rely on simple thresholds or single scales, automatically adjusting the intervention time window based on dynamic changes to ensure more scientific and effective intervention timing. Through these improvements, the solution surpasses existing technologies in accuracy, sensitivity, and real-time performance, enabling better support for early warning and intervention of physiological abnormalities.

[0147] In an optional embodiment, constructing a hierarchical coupling network among the physiological indicators, recursively calculating the direct coupling coefficient and the indirect coupling coefficient in the hierarchical coupling network, and performing weighted correction on the change timing relationship to obtain the indicator conduction delay includes:

[0148] Sliding window segmentation is performed on the temporal relationship of changes between physiological indicators to obtain local temporal segments. The matching degree of the local temporal segments is calculated using a dynamic time warping algorithm to obtain a temporal similarity matrix. Multi-scale wavelet transform is performed on the local temporal segments to perform frequency domain decomposition. The phase synchronization of different frequency components is calculated to obtain a frequency coupling matrix. The temporal similarity matrix and the frequency coupling matrix are adaptively weighted and fused to obtain the coupling strength between indicators.

[0149] An initial hierarchical coupling network is constructed based on the coupling strength between indicators, conditional mutual information between network nodes is calculated, a causal entropy significance test is performed based on the conditional mutual information to screen stable coupling edges, the initial hierarchical coupling network is updated using the stable coupling edges to obtain an optimized hierarchical coupling network, and the coupling strength of network adjacent edges in the optimized hierarchical coupling network is extracted to obtain a direct coupling coefficient;

[0150] Performing a depth-first search on the optimized hierarchical coupling network to identify candidate conduction links between physiological indicators, calculating the information entropy of the candidate conduction links, setting an effective threshold based on the information entropy to screen effective conduction links, performing a recursive calculation on the effective conduction links, and taking the ratio of the cumulative effect of the coupling strength of each edge on the link to the path length as the indirect coupling coefficient;

[0151] A comprehensive coupling coefficient is calculated based on the direct coupling coefficient and the indirect coupling coefficient, and a nonlinear weighted correction is performed on the change timing relationship using the comprehensive coupling coefficient. The corrected change timing relationship is normalized and converted into a time unit to obtain an indicator conduction delay.

[0152] For example, to analyze the transmission relationship between physiological indicators, it is first necessary to conduct a detailed analysis of the temporal relationship of each indicator's changes. The time series data is segmented by setting a sliding window. The window size can be dynamically adjusted based on the signal characteristics. For example, a five-minute base window size can be set for the heart rate variability indicator, and a ten-minute base window size can be set for the blood pressure indicator. The sliding step size is set to one-quarter of the window size to ensure sufficient overlap between adjacent windows to capture continuously changing features.

[0153] For each local time series segment within a sliding window, a dynamic time warping algorithm is used to calculate the degree of matching between different metrics. This algorithm constructs a distance matrix to find the optimal time alignment path, and is capable of processing time series data of varying lengths and sampling rates. For example, when analyzing the correlation between heart rate and blood pressure, dynamic warping can achieve the optimal matching relationship between the two signals, even if they have different sampling frequencies.

[0154] The result of dynamic time warping is converted into a time series similarity matrix, where each element in the matrix represents the degree of similarity between the corresponding indicator pair in the current time window. The similarity value is mapped to the range of zero to one through normalization. The larger the value, the closer the change patterns of the two indicators are. For example, if the heart rate and respiratory rate show highly synchronized changes within a certain time window, the similarity value between them will be close to one. At the same time, a multi-scale wavelet transform is performed on the local time series segment to decompose the time domain signal into different frequency components. The wavelet basis function selects Ricker wavelet, and the decomposition scale ranges from one to six, covering the full spectrum range from high frequency to low frequency. In this way, signal features at different time scales can be analyzed, for example, rapid fluctuations in heart rate and slow changes in blood pressure can be captured simultaneously.

[0155] The phase synchronization between different frequency components is calculated to obtain a frequency coupling matrix. Phase synchronization is achieved by extracting the instantaneous phase of the signal through a Hilbert transform, and then calculating the stability of the phase difference. For example, when a stable phase relationship is observed between heart rate and blood oxygen saturation in a specific frequency band, it indicates that the two indicators have significant coupling characteristics at that frequency. The time series similarity matrix and the frequency coupling matrix are adaptively weighted and fused to obtain a comprehensive inter-indicator coupling strength. The weighting coefficients are dynamically adjusted based on the signal-to-noise ratio of each indicator, assigning greater weight to indicators with better signal quality. The resulting coupling strength incorporates both time-domain and frequency-domain characteristics, fully reflecting the correlation between indicators.

[0156] An initial hierarchical coupling network is constructed based on the coupling strength. The nodes in the network represent different physiological indicators, and the weights of the edges are determined by the coupling strength. In order to ensure the reliability of the network structure, it is necessary to calculate the conditional mutual information between the network nodes. The conditional mutual information reflects the true dependency relationship between two nodes while considering the influence of other nodes. A causal entropy significance test is performed on the conditional mutual information to screen out stable coupling edges. Specifically, by constructing a null hypothesis distribution of random permutation, the significance level of the actual observation value exceeding the distribution is calculated. For example, if the conditional mutual information between heart rate and blood pressure is significantly higher than the random level, the coupling edge between the two indicators is retained.

[0157] The initial network is updated using the stable coupling edges obtained through screening to obtain an optimized hierarchical coupling network. The coupling strength of adjacent edges in the optimized network is extracted as the direct coupling coefficient, reflecting the direct influence between indicators. For example, a high direct coupling coefficient between heart rate and cardiac output indicates a direct physiological regulatory relationship between the two indicators. A depth-first search is performed on the optimized hierarchical coupling network to identify possible transmission links between the indicators. During the search, all possible paths are recorded to form a set of candidate transmission links. For example, there may be multiple transmission paths from heart rate to blood pressure, including direct paths and indirect paths through other indicators. The information entropy of the candidate transmission links is calculated to evaluate the information transfer efficiency of the links. Lower information entropy indicates a more stable transmission process and higher link reliability. A screening threshold is set based on the information entropy, and links with information entropy below the threshold are retained as valid transmission links.

[0158] A recursive calculation is performed on valid conduction links, taking into account the cumulative effect of each coupled edge along the link. The ratio of the accumulated coupling strength to the link length is used as the indirect coupling coefficient, reflecting the efficiency of long-range conduction. For example, if a stable conduction link from heart rate to tissue perfusion metric exists through multiple intermediate nodes, the indirect coupling coefficient remains high, indicating that this conduction path is physiologically important.

[0159] A comprehensive evaluation model is constructed based on the direct and indirect coupling coefficients, and the combined coupling coefficient is calculated. The model uses a nonlinear weighting approach, with different weight allocation strategies for different coupling strength ranges. For example, when both direct and indirect coupling are strong, the combined coefficient is significantly improved, reflecting the synergistic effect of multiple conduction pathways. The combined coupling coefficient is used to perform a nonlinear weighted correction on the original timing relationship of the changes. This correction process takes into account the complexity and stability of the conduction link, and can more accurately reflect the actual conduction delay between indicators. For example, if a conduction link has a high combined coupling coefficient, the corresponding timing delay estimate will be reduced.

[0160] The corrected temporal relationships are normalized to eliminate the influence of different indicator dimensions. Finally, the normalized temporal relationships are converted into actual time units to obtain the indicator transmission delay. This delay value intuitively reflects the time required for a change in one indicator to trigger a response in another indicator, providing an important time window reference for clinical intervention.

[0161] Based on the above technical solution, a deep understanding and precise modeling of the complex coupling relationships between physiological indicators can be achieved. By combining dynamic time warping and wavelet transforms, not only can the time series of changes between indicators be analyzed from a time-domain perspective, but the coupling degree of different frequency components can also be explored in the frequency domain, making the calculation of coupling strength more comprehensive and accurate. Furthermore, the hierarchical coupling network is optimized through conditional mutual information and causal entropy significance tests to screen for stable coupling edges, effectively improving the stability and reliability of the coupling relationship. Regarding conduction path identification, a depth-first search and information entropy-based screening of valid links dynamically capture key conduction pathways between physiological indicators, avoiding complex path relationships that may be overlooked in traditional methods. Recursive calculation of the cumulative effect and path length ratio provides a new method for the precise calculation of indirect coupling coefficients, making the measurement of conduction delay more detailed and accurate. Through nonlinear weighting correction and normalization, the conduction delay of physiological indicators can be accurately calculated, providing reliable data support for further intervention decisions. Overall, this solution improves the accuracy and adaptability of physiological data analysis, especially in the modeling of multi-scale, multi-frequency and nonlinear relationships, and has obvious improvements compared with traditional methods.

[0162] Figure 4 This is a comparison chart of delay calculation accuracy in an embodiment of the present invention. Figure 4 The figure shows a comparison of the mean absolute errors (MAEs) of time delay calculations between six physiological indicator pairs using four different methods. This technical solution, which utilizes multidimensional coupling analysis and a hierarchical coupled network approach, achieved the lowest error across all test cases. Specifically, for the blood pressure-heart rate pair, this technical solution achieved an error of only 0.22 seconds, compared to 0.60 seconds for the cross-correlation method, 0.80 seconds for the phase synchronization method, and 0.40 seconds for the traditional coupled network method. For the most complex pair, EEG-blood pressure, this technical solution achieved an error of 0.44 seconds, compared to 1.10 seconds, 1.40 seconds, and 0.80 seconds for the other three methods, respectively. This advantage is even more pronounced. It is particularly noteworthy that traditional cross-correlation methods (based on signal correlation analysis) and phase synchronization methods (based on Hilbert transform phase extraction) perform poorly when dealing with nonlinear coupling relationships. While traditional coupled network methods (such as Granger causal networks) offer some improvements, they do not consider frequency domain characteristics, resulting in limited accuracy. This technical solution significantly improves the accuracy of time delay calculation between various physiological indicators by fusing time domain and frequency domain features and performing information entropy screening.

[0163] A second aspect of an embodiment of the present invention provides a schematic structural diagram of an AI-based heart failure risk assessment system, the system comprising:

[0164] The first unit is used to collect physiological indicators of the subject to be evaluated and construct them into a multidimensional time series data stream. The difference between adjacent data segments in the multidimensional time series data stream is calculated using a recursive segmentation algorithm. The time point where the difference is greater than a preset difference threshold is determined as a mutation point. The waveform features of the data segments before and after the mutation point are extracted to generate a mutation feature matrix.

[0165] The second unit is used to perform time series mapping on the mutation feature matrix, obtain the time interval sequence and amplitude sequence of the mutation features, identify periodic mutation patterns and gradual change patterns, convert the duration and frequency of periodic mutation patterns and gradual change patterns into risk scores, and draw the risk accumulation curve;

[0166] The third unit is used to calculate the temporal relationship between physiological indicators based on the risk accumulation curve, convert the temporal relationship into the indicator transmission delay, draw the risk transmission path diagram based on the indicator transmission delay, and mark the time interval with the largest change slope in the risk transmission path diagram as the intervention time window;

[0167] The fourth unit is used to determine the risk level of heart failure based on the time range of the intervention time window and the changing trend of the risk accumulation curve, and generate a risk assessment report containing the risk level and description of abnormal physiological indicators.

[0168] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0169] processor;

[0170] a memory for storing processor-executable instructions;

[0171] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0172] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0173] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0174] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-based heart failure risk assessment method, characterized by: include: The physiological indicators of the subject to be evaluated are collected and constructed into a multidimensional time series data stream. The difference between adjacent data segments in the multidimensional time series data stream is calculated using a recursive segmentation algorithm. The time point where the difference is greater than a preset difference threshold is determined as the mutation point. The waveform features of the data segments before and after the mutation point are extracted to generate a mutation feature matrix. Perform time series mapping on the mutation feature matrix to obtain the time interval sequence and amplitude sequence of the mutation features, identify periodic mutation patterns and gradual change patterns, convert the duration and frequency of periodic mutation patterns and gradual change patterns into risk scores, and draw a risk accumulation curve; Calculate the temporal relationship between physiological indicators based on the risk accumulation curve, convert the temporal relationship into the indicator conduction delay, draw a risk conduction path diagram based on the indicator conduction delay, and mark the time interval with the largest change slope in the risk conduction path diagram as the intervention time window; Based on the time range of the intervention time window and the changing trend of the risk accumulation curve, the heart failure risk level is determined and a risk assessment report containing the risk level and description of abnormal physiological indicators is generated; Calculate the temporal relationship between physiological indicators based on the risk accumulation curve, convert the temporal relationship into the indicator conduction delay, and draw a risk conduction path diagram based on the indicator conduction delay. Mark the time interval with the largest change slope in the risk conduction path diagram as the intervention time window, including: Extracting local features of the risk accumulation curve at multiple time window scales, calculating a distance matrix between physiological indicators based on the local features, and calculating the temporal relationship of changes between the physiological indicators using a time warping algorithm with attention weights based on the distance matrix; Constructing a hierarchical coupling network among the physiological indicators, recursively calculating direct coupling coefficients and indirect coupling coefficients in the hierarchical coupling network, and performing weighted correction on the temporal relationship of the changes to obtain the indicator conduction delay; Constructing a risk transmission path diagram using the physiological indicators as nodes and the indicator transmission delay as edge weights, calculating the dynamic scores of the nodes in the risk transmission path diagram using a time-varying node importance algorithm, and determining the transmission path based on the dynamic scores and the indicator transmission delay; An adaptive segmented planning algorithm is used on the conduction path to search for the maximum slope interval to obtain a candidate time window, and the information entropy of the slope sequence in the candidate time window is calculated to obtain the window confidence. When the window confidence is lower than a preset confidence threshold, the interval range is expanded along the conduction path to re-search the maximum slope interval, and the maximum slope interval that meets the window confidence is used as the final intervention time window.

2. The method according to claim 1, characterized in that The physiological indicators of the subject to be evaluated are collected and constructed into a multidimensional time series data stream. The difference between adjacent data segments in the multidimensional time series data stream is calculated through a recursive segmentation algorithm. The time point when the difference is greater than a preset difference threshold is determined as a mutation point. Acquiring physiological indicators of the subject to be evaluated, the physiological indicators including cardiovascular data, respiratory data, and blood oxygen data; resampling the physiological indicators through cubic spline interpolation to construct a multidimensional time series data stream, and constructing the multidimensional time series data stream into a state matrix of time-indicator dimensions; Initially dividing the state matrix from the midpoint to obtain two sub-data segments, moving the dividing position point by point along the time axis to calculate the Mahalanobis distance of the data segments on both sides to obtain a distance sequence, and performing eigenvalue decomposition to obtain a principal component weight vector, and using the principal component weight vector to weight the distance sequence to obtain a sub-data segment difference sequence; A cumulative distribution function is obtained by statistically analyzing the probability distribution characteristics of the difference degree sequence of the sub-data segments, a difference threshold is determined on the cumulative distribution function according to a preset false alarm rate, a time point greater than the difference threshold is marked as an initial mutation point, an observation window is constructed with the initial mutation point as the center, and the mutual information between the signals of each dimension in the observation window is calculated to obtain a mutual information matrix, and the sum of the off-diagonal elements of the mutual information matrix is used as a coupling metric value; A comprehensive difference is calculated based on the coupling metric value and the sub-data segment difference sequence, the initial mutation point is screened according to the comprehensive difference, and the time point when the comprehensive difference is greater than the difference threshold is determined as the final mutation point.

3. The method according to claim 1, characterized in that Extract the waveform features of the data segments before and after the mutation point and generate the mutation feature matrix including: The fluctuation characteristics of the data segments before and after the mutation point are calculated according to the location of the mutation point to obtain a volatility sequence, and the fluctuation trend characteristics are obtained by performing a segmented cumulative calculation. The adaptive window length is determined based on the fluctuation trend characteristics, and the waveform feature extraction interval before and after the mutation point is determined using the adaptive window length. In the waveform feature extraction interval, the data segments before and after the mutation point are respectively subjected to time-frequency transformation to obtain a time-frequency spectrum, the energy distribution of the time-frequency spectrum is calculated to obtain a time-frequency energy sequence, the time-frequency energy sequence is weightedly fused in combination with the volatility sequence to obtain a fused feature sequence, and the multi-scale entropy value of the fused feature sequence is calculated to obtain a complexity feature; The volatility sequence, time-frequency energy sequence and complexity feature are combined in time sequence to obtain a feature vector, an initial feature matrix is constructed according to the feature vector, the discriminant contribution of each feature in the initial feature matrix is calculated to obtain the feature weight, and the feature weight is used to perform weighted reconstruction on the initial feature matrix to generate a mutation feature matrix.

4. The method according to claim 1, wherein Perform time series mapping on the mutation feature matrix to obtain the time interval sequence and amplitude sequence of the mutation features, identify periodic mutation patterns and gradual change patterns, convert the duration and frequency of periodic mutation patterns and gradual change patterns into risk scores, and draw the risk accumulation curve, including: Performing time series mapping on the mutation feature matrix to obtain a time interval sequence and an amplitude sequence, and constructing a mutation topology network based on the time interval sequence and amplitude sequence using an adaptive weight function, where nodes represent mutation points and the connection weights between nodes are determined by the time interval and amplitude. Calculating the local entropy of nodes at different scales of the mutant topology network, determining the optimal scale set according to the variation characteristics of the local entropy, and fusing the local entropies under the optimal scale set to obtain an enhanced topological entropy sequence; Dynamically adjusting the calculation window size based on the rate of change of the enhanced topological entropy sequence, calculating the autocorrelation function and power spectral density of the enhanced topological entropy sequence within the calculation window, extracting periodic mutation characteristics, determining a first duration and a first frequency of occurrence of a periodic mutation pattern based on the periodic mutation characteristics, performing piecewise polynomial fitting on the enhanced topological entropy sequence to identify a gradual change interval, and determining a second duration and a second frequency of occurrence of the gradual change pattern based on the gradual change interval; The first duration and the first frequency of occurrence are converted into a first risk score through a first mapping function, and the second duration and the second frequency of occurrence are converted into a second risk score through a second mapping function. A risk score is obtained based on a dynamic weighted combination of the first risk score and the second risk score, and the risk score is time-integrated to obtain a risk accumulation curve.

5. The method according to claim 4, characterized in that Calculating the local entropy of nodes at different scales of the mutant topology network, determining the optimal scale set based on the variation characteristics of the local entropy, and fusing the local entropies under the optimal scale set to obtain an enhanced topology entropy sequence includes: Constructing a node neighborhood matrix in the mutant topology network, wherein the node neighborhood matrix represents the set of nodes that each node can reach within a specified number of steps, constructing a node feature vector including degree centrality and clustering coefficient based on the node neighborhood matrix, calculating node structural similarity, and using the node structural similarity to adaptively reconstruct network connections; Calculating cumulative weights between nodes in the reconstructed network, where the cumulative weights are accumulated recursively at different numbers of steps, calculating transition probabilities between nodes based on the cumulative weights, and adaptively nonlinearly adjusting the transition probabilities according to the local density distribution of the network to obtain modified transition probabilities; Based on the modified transition probability, the local entropy of the node is calculated at different scales to obtain a local entropy sequence, the multi-order differential characteristics and community evolution characteristics of the local entropy sequence are calculated, and a scale importance index is constructed; An optimal scale set is determined according to the scale importance index, and the information redundancy of each scale in the optimal scale set is calculated based on the complementarity metric. The normalized inverse of the information redundancy is used as the fusion weight. The local entropy sequence under the optimal scale set is weightedly combined using the fusion weight to obtain an enhanced topological entropy sequence.

6. The method according to claim 1, characterized in that Constructing a hierarchical coupling network among the physiological indicators, recursively calculating direct coupling coefficients and indirect coupling coefficients in the hierarchical coupling network, and weightedly correcting the temporal relationship of the changes to obtain the indicator conduction delay includes: Sliding window segmentation is performed on the temporal relationship of changes between physiological indicators to obtain local temporal segments. The matching degree of the local temporal segments is calculated using a dynamic time warping algorithm to obtain a temporal similarity matrix. Multi-scale wavelet transform is performed on the local temporal segments to perform frequency domain decomposition. The phase synchronization of different frequency components is calculated to obtain a frequency coupling matrix. The temporal similarity matrix and the frequency coupling matrix are adaptively weighted and fused to obtain the coupling strength between indicators. An initial hierarchical coupling network is constructed based on the coupling strength between indicators, conditional mutual information between network nodes is calculated, a causal entropy significance test is performed based on the conditional mutual information to screen stable coupling edges, the initial hierarchical coupling network is updated using the stable coupling edges to obtain an optimized hierarchical coupling network, and the coupling strength of network adjacent edges in the optimized hierarchical coupling network is extracted to obtain a direct coupling coefficient; Performing a depth-first search on the optimized hierarchical coupling network to identify candidate conduction links between physiological indicators, calculating the information entropy of the candidate conduction links, setting an effective threshold based on the information entropy to screen effective conduction links, performing a recursive calculation on the effective conduction links, and taking the ratio of the cumulative effect of the coupling strength of each edge on the link to the path length as the indirect coupling coefficient; A comprehensive coupling coefficient is calculated based on the direct coupling coefficient and the indirect coupling coefficient, and a nonlinear weighted correction is performed on the change timing relationship using the comprehensive coupling coefficient. The corrected change timing relationship is normalized and converted into a time unit to obtain an indicator conduction delay.

7. An AI-based heart failure risk assessment system, for implementing the method according to any one of claims 1 to 6, characterized in that: include: The first unit is used to collect physiological indicators of the subject to be evaluated and construct them into a multidimensional time series data stream. The difference between adjacent data segments in the multidimensional time series data stream is calculated using a recursive segmentation algorithm. The time point where the difference is greater than a preset difference threshold is determined as a mutation point. The waveform features of the data segments before and after the mutation point are extracted to generate a mutation feature matrix. The second unit is used to perform time series mapping on the mutation feature matrix, obtain the time interval sequence and amplitude sequence of the mutation features, identify periodic mutation patterns and gradual change patterns, convert the duration and frequency of periodic mutation patterns and gradual change patterns into risk scores, and draw the risk accumulation curve; The third unit is used to calculate the temporal relationship between physiological indicators based on the risk accumulation curve, convert the temporal relationship into the indicator transmission delay, draw the risk transmission path diagram based on the indicator transmission delay, and mark the time interval with the largest change slope in the risk transmission path diagram as the intervention time window; The fourth unit is used to determine the risk level of heart failure based on the time range of the intervention time window and the changing trend of the risk accumulation curve, and generate a risk assessment report containing the risk level and description of abnormal physiological indicators.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images