Detection method for combined variability index of cardiovascular system

Abstract

The invention discloses a cardiovascular system joint variability index detection method which comprises the following steps: synchronously acquiring an electrocardiosignal, an arterial blood pressure signal and a respiratory signal through a multi-channel physiological signal acquisition device, and carrying out denoising and baseline correction on each signal by adopting a self-adaptive filtering algorithm to obtain a preprocessed multi-modal signal sequence; respectively extracting a heart rate variability time domain index, a blood pressure variability frequency domain index and a breath variability nonlinear index according to the multi-modal signal sequence, mapping the three indexes into a unified sliding window by adopting a timestamp alignment mechanism, and generating an aligned variability feature set; performing Z-score-based standardization processing on the variability feature set, and constructing a dynamic weight distribution matrix in combination with individual physiological parameters; the objective of the invention is to solve the problems of time sequence mismatch, dimension difference, insufficient individual adaptation and historical data continuity deficiency in the process of acquisition, synchronization, feature extraction and fusion of multi-source physiological signals in the prior art.

Description

Technical Field

[0001] This invention relates to the technical field of biomedical signal processing and cardiovascular health assessment, specifically to a method for detecting combined variability indicators of the cardiovascular system. Background Technology

[0002] In the detection of combined variability indicators of the cardiovascular system, the core technical challenge lies in how to integrate multi-source physiological signals such as heart rate variability, blood pressure variability, and respiratory variability to construct a comprehensive assessment indicator that reflects the state of autonomic nervous system regulation, while ensuring the temporal consistency and physiological interpretability of the indicator calculation process. This problem involves extracting effective features from physiological signals with different sampling frequencies and time scales and synchronizing them. However, multimodal signals are easily affected by motion artifacts, baseline drift, or equipment sampling errors during the acquisition process, leading to feature extraction deviations and thus affecting the stability of the combined indicator.

[0003] Furthermore, when mapping various variability indicators to a unified analytical framework, the differences in the dimensions, dynamic range, and physiological significance of each indicator may lead to an imbalance in feature weights if standardization or normalization is not performed, thus weakening the discriminative ability of the joint indicators.

[0004] Furthermore, in real-time or near-real-time monitoring scenarios, the calculation of joint variability indicators depends on the reasonable setting of the sliding window length and update frequency. If the window is too short, it may introduce statistical fluctuations, while if the window is too long, it may reduce the sensitivity to rapid physiological changes and affect the timeliness of clinical decision-making.

[0005] Ultimately, when outputting comprehensive evaluation results through algorithmic models, if the impact of individual differences (such as age, underlying diseases, or medication status) on variability is not fully considered, the generalization ability of the evaluation results may be limited. Furthermore, in long-term monitoring, if the data storage structure is not optimized or lacks a timestamp verification mechanism, the continuity and reliability of historical trend analysis may be affected.

[0006] This complex issue spans the entire process from multi-source signal acquisition and feature fusion to dynamic assessment, involving complex interactions of signal processing, physiological modeling, individual adaptation, and data management, and directly affecting the accuracy and clinical applicability of cardiovascular risk assessment. Summary of the Invention

[0007] This invention provides a method for detecting combined variability indicators of the cardiovascular system, aiming to solve the problems of temporal mismatch, dimensional differences, insufficient individual adaptation, and lack of continuity of historical data in the process of acquisition, synchronization, feature extraction and fusion of multi-source physiological signals in the prior art.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for detecting joint variability indices of the cardiovascular system includes: simultaneously acquiring electrocardiogram (ECG), arterial blood pressure (APS), and respiratory signals using a multi-channel physiological signal acquisition device; performing noise reduction and baseline correction on each signal using an adaptive filtering algorithm to obtain a preprocessed multimodal signal sequence; extracting time-domain indices of heart rate variability, frequency-domain indices of blood pressure variability, and nonlinear indices of respiratory variability from the multimodal signal sequence; mapping the three types of indices to a unified sliding window using a timestamp alignment mechanism to generate an aligned variability feature set; performing Z-score-based standardization on the variability feature set; constructing a dynamic weight allocation matrix based on individual physiological parameters; generating an initial joint variability vector using a weighted fusion strategy; and using an adaptive adjustment module for the sliding window length. The window length is dynamically adjusted based on the fluctuation range of the current physiological state. If the fluctuation range exceeds a preset threshold, the window length is shortened; otherwise, the window length is extended to obtain an optimized joint variability vector. The optimized joint variability vector is input into a multilayer perceptron model trained individually, and a comprehensive assessment score reflecting the state of autonomic nervous regulation is output. If the comprehensive assessment score falls into the high-risk range, a local early warning logic is triggered. The original signal segment and assessment result are encapsulated using the AES-256 encryption algorithm and written to a local database with a timestamp index. Based on the timestamp index in the local database, a time-series hash chain is constructed, and the hash value is calculated segment by segment using the SHA3-256 algorithm to form an immutable physiological record chain for subsequent trend analysis and clinical retrospection.

[0009] In one aspect of this disclosure, the step of simultaneously acquiring electrocardiogram signals, arterial blood pressure signals, and respiratory signals through a multi-channel physiological signal acquisition device, and using an adaptive filtering algorithm to denoise and correct the baseline of each signal to obtain a preprocessed multimodal signal sequence includes: Synchronous acquisition is initiated by a hardware-level synchronous trigger signal using ECG electrodes with a sampling frequency of 1000Hz, an arterial pressure sensor with a sampling frequency of 200Hz, and a respiratory impedance sensor with a sampling frequency of 50Hz, to obtain the original multi-channel signal. Wavelet threshold denoising combined with morphological filtering was used to remove electromyographic interference and power frequency noise from the electrocardiogram signal while preserving the QRS complex characteristics. Savitzky-Golay smoothing filter was used on the arterial blood pressure signal to eliminate pulse wave baseline drift and retain the peak points of systolic and diastolic blood pressure. Empirical Mode Decomposition (EMD) is used to extract intrinsic mode functions from respiratory signals, and the IMF components corresponding to the main frequency band are selected as valid respiratory waveforms. The three preprocessed signals are aligned according to the sampling points, and interpolation is used to resample them to a uniform sampling rate of 200Hz to generate a time-aligned multimodal signal sequence. The multimodal signal sequence is segmented with a sliding window of 300 seconds and a step size of 30 seconds, and each segment is independently marked with a start timestamp; Each signal segment is stored in a circular buffer. When the buffer is full, the earliest segment is automatically overwritten, and the feature extraction process is triggered simultaneously.

[0010] In one aspect of this disclosure, the step of extracting time-domain indices of heart rate variability, frequency-domain indices of blood pressure variability, and nonlinear indices of respiratory variability from the multimodal signal sequence, and mapping the three types of indices to a unified sliding window using a timestamp alignment mechanism to generate an aligned variability feature set includes: The peak value of the R wave is detected from the electrocardiogram signal, the adjacent RR interval is calculated, and SDNN, RMSSD, and pNN50 are extracted as time-domain indicators. A continuous systolic blood pressure sequence was extracted from the arterial blood pressure signal, and a fast Fourier transform was performed. The power ratio of low frequency (0.04–0.15 Hz) to high frequency (0.15–0.4 Hz) was calculated as a frequency domain index. The standard deviation of the respiratory cycle and the sample entropy are calculated from the respiratory signals as indicators of nonlinear variability. Add the same start timestamp as its corresponding signal segment to each type of indicator; Establish a timestamp index table to precisely match the three types of indicators to the same sliding window based on their timestamps; If a certain type of indicator is missing due to signal quality, forward padding combined with linear interpolation will be used to complete it, and the number of completions will not exceed two consecutive times. Generate an aligned variable feature set containing 9-dimensional features, with each feature labeled with its source type and timestamp.

[0011] In one aspect of this disclosure, the steps of performing Z-score-based standardization on the variability feature set, constructing a dynamic weight allocation matrix in conjunction with individual physiological parameters, and generating an initial joint variability vector using a weighted fusion strategy include: Obtain the age, gender, medical history, and list of commonly used medications entered by the user during registration; Based on age and history of underlying diseases, the corresponding feature weight bias vector is retrieved from the preset rule base; For each feature in the variability feature set, calculate its mean and standard deviation in the most recent 7 days of historical data, and perform Z-score standardization; The standardized feature vector is multiplied element-wise with the weight bias vector to generate individualized weighted coefficients. The weighted coefficients are Softmax normalized to ensure that the sum is 1, thus forming a dynamic weight allocation matrix. The standardized feature vector is multiplied by the dynamic weight assignment matrix to generate an 8-dimensional initial joint variability vector. Bind the initial joint variability vector to the current timestamp and store it in the feature queue.

[0012] In one aspect of this disclosure, the step of dynamically adjusting the window length according to the fluctuation amplitude of the current physiological state through the sliding window length adaptive adjustment module, shortening the window length if the fluctuation amplitude exceeds a preset threshold, and extending the window length otherwise, to obtain the optimized joint variability vector, includes: Calculate the Euclidean distance between the joint variability vectors of the three most recent windows, and take the maximum value as the current fluctuation amplitude; Set the fluctuation threshold to 0.35. If the fluctuation amplitude is greater than 0.35, adjust the sliding window length from 300 seconds to 180 seconds. If the fluctuation amplitude is less than 0.20 three times in a row, the window length will be restored to 300 seconds. After adjusting the window length, the feature extraction and fusion process is re-executed to generate a joint variability vector under the new window. Perform a time continuity check between the new vector and the old vector. If the time interval exceeds 1.5 times the window step size, then insert an interpolated vector. Attach window length metadata to the optimized joint variability vector for use as subsequent model input; Replace the oldest record in the feature queue with the optimized vector, while maintaining the queue length at 10.

[0013] In one aspect of this disclosure, the step of inputting the optimized joint variability vector into a multilayer perceptron model trained individually, and outputting a comprehensive evaluation score reflecting the state of autonomic neural regulation, includes: Load a user-specific multilayer perceptron model, which has been fine-tuned based on the user's historical data and contains 3 hidden layers with 64 neurons in each layer; The optimized joint variability vector is used as input and propagated layer by layer through the ReLU activation function; The output layer uses the Sigmoid function to generate a comprehensive evaluation score between 0 and 1; If the score is greater than 0.75, it is considered as excessive activation of the sympathetic nervous system; if it is less than 0.25, it is considered as inhibition of parasympathetic function. Write the evaluation score and the judgment category to the structured output buffer; If the score changes by more than 0.2 in two consecutive evaluations, the model retraining process is triggered, and the weights are updated using data from the most recent 24 hours. The updated model parameters are encrypted and stored in a local secure chip.

[0014] In one aspect of this disclosure, the step of triggering a local early warning logic if the comprehensive evaluation score falls into a high-risk range, and encapsulating the original signal segment and evaluation result using the AES-256 encryption algorithm and writing them into a local database with a timestamp index, includes: When the comprehensive assessment score is greater than 0.8 or less than 0.2, it is determined to be in the high-risk range; Extract the raw multi-channel signals from the circular buffer within 5 minutes before and after the corresponding timestamp; The signal data and evaluation results are concatenated into a JSON format payload. The payload is encrypted using the device's unique key and the AES-256 algorithm. The metadata for generating encrypted data packets includes the start timestamp, duration, and risk level. Write the encrypted data packet and metadata into the SQLite database, with the primary key being an ISO8601 format timestamp; After the database write is successful, an encrypted warning notification is pushed to the user terminal.

[0015] In one aspect of this disclosure, the step of constructing a time-series hash chain based on the timestamp index in the local database, calculating hash values ​​segment by segment using the SHA3-256 algorithm to form an immutable physiological record chain for subsequent trend analysis and clinical retrospection includes: Iterate through all records in the local database in ascending order of timestamp; Calculate the SHA3-256 hash value for the encrypted data packet of each record, and use it as the hash of the current block; The current block hash is concatenated with the previous block hash and then hashed again to generate a chained hash. The hash of the first record is set to an all-zero string. Append the chained hash value to the validation field of the corresponding database record; Perform integrity verification periodically, recursively tracing the hash chain backwards from the latest record, and marking the data as abnormal if any verification fails. During clinical retrospective analysis, a continuous hash chain is returned by querying the time range, allowing third parties to verify that the data has not been tampered with.

[0016] Compared with the prior art, the present invention has the following beneficial effects: First, this invention addresses the temporal mismatch and noise interference issues of multi-source physiological signals during the acquisition phase through multi-channel synchronous acquisition and adaptive filtering technology, providing a high signal-to-noise ratio and strictly time-aligned signal foundation for subsequent feature extraction. Second, by employing timestamp alignment and a unified sliding window mapping mechanism, it achieves spatiotemporal synchronous fusion of three heterogeneous variability indicators—heart rate, blood pressure, and respiration—effectively overcoming the feature mismatch problem caused by the asynchronous signal acquisition frequency and physiological events in traditional methods. Third, through dynamic weight allocation based on individual physiological parameters and Z-score standardization, it balances group universality with individual differences, improving the physiological interpretability and clinical suitability of the combined indicators. Furthermore, the sliding window length can be dynamically and adaptively adjusted according to physiological fluctuations, avoiding the sluggish response of long windows to rapid changes and suppressing the random fluctuations introduced by short windows, thus achieving a balance between stability and sensitivity in dynamic monitoring. Ultimately, by outputting a comprehensive evaluation score through a multilayer perceptron model and combining it with local encrypted storage and hash chain construction, not only was real-time and objective evaluation and risk warning of autonomic nervous function achieved, but also the integrity, immutability and reliability of physiological data and clinical backtesting were ensured, thus improving the accuracy, security and practicality of cardiovascular system status monitoring as a whole. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is one of the flowcharts for a method for detecting combined variability indicators of the cardiovascular system according to the present invention.

[0019] Figure 2 This is the second flowchart of a method for detecting combined variability indicators of the cardiovascular system according to the present invention.

[0020] Figure 3 This is the third flowchart of a method for detecting combined variability indicators of the cardiovascular system according to the present invention. Detailed Implementation

[0021] The present invention will be further described below with reference to embodiments. These embodiments are merely some, not all, of the embodiments of the present invention. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the protection scope of the present invention.

[0022] Please see Figures 1-3 As shown in the figure, this embodiment discloses a method for detecting combined variability indicators of the cardiovascular system according to the present invention. The specific implementation process is as follows: First, a multi-channel physiological signal acquisition device is deployed. This device includes electrocardiogram (ECG) electrodes, an arterial pressure sensor, and a respiratory impedance sensor, all three connected to a unified data acquisition card. The data acquisition card has a built-in hardware synchronization trigger controller. The ECG electrodes are attached to the subject's chest wall in a standard lead II configuration, with a sampling frequency set to 1000Hz. The arterial pressure sensor acquires continuous arterial blood pressure waveforms through radial artery puncture or a non-invasive cuff-type pressure probe, with a sampling frequency set to 200Hz. The respiratory impedance sensor is integrated into a chest band, surrounding the subject's chest at the 4th to 6th intercostal spaces, with a sampling frequency set to 50Hz. The three sensor signal cables are all connected to different analog input channels of the same data acquisition card. The data acquisition card is equipped with a high-precision clock source, and simultaneous sampling of the three signals is initiated through a hardware-level synchronization trigger signal, ensuring strict alignment of the original multi-channel signals at the time starting point.

[0023] The acquired raw multi-channel signals are sequentially sent to the signal preprocessing module for processing. For ECG signals, five-level wavelet decomposition is first performed using the db6 wavelet basis function, and soft thresholding is applied to the detail coefficients for noise reduction. Then, morphological opening and closing operations are used to filter out residual baseline drift and high-frequency electromyographic interference, while preserving the typical morphological characteristics of the QRS complex. For arterial blood pressure signals, a Savitzky-Golay smoothing filter with a window length of 51 points and a polynomial order of 3 is used to process the raw waveform, eliminating slow baseline drift caused by breathing or changes in body position, while preserving the precise positions of the systolic blood pressure peak and diastolic blood pressure trough. For respiratory signals, the Empirical Mode Decomposition (EMD) algorithm is executed to decompose the raw signal into several intrinsic mode functions (IMFs). The center frequency of each IMF component is calculated, and the main IMF component with a center frequency in the range of 0.1–0.4 Hz is selected as the effective respiratory waveform, removing high-frequency noise and low-frequency trend terms. After completing the aforementioned independent preprocessing steps, the three signals are aligned according to the timestamps of the original sampling points, and then resampled to a uniform sampling rate of 200Hz using a cubic spline interpolation algorithm to generate a time-strictly aligned multimodal signal sequence. This multimodal signal sequence is fed into a circular buffer with a capacity set to store 10 signal segments, each lasting 300 seconds, with a step size of 30 seconds. That is, a 300-second signal segment is extracted every 30 seconds and its starting timestamp is marked. When the buffer is full, the oldest stored signal segment is automatically overwritten, and a trigger signal is sent to the feature alignment module.

[0024] After receiving the trigger signal, the feature alignment module reads the latest multimodal signal sequence from the circular buffer and extracts three types of variability indicators respectively. For ECG signals, the Pan-Tompkins algorithm was used to detect the R-wave peak position, calculate the adjacent RR interval sequence, and then calculate the time-domain indices SDNN (standard deviation of all normal sinus intervals), RMSSD (root mean square of the difference between adjacent normal sinus intervals), and pNN50 (percentage of adjacent intervals with a difference greater than 50ms). For arterial blood pressure signals, systolic blood pressure values ​​were extracted beat-by-beat from the preprocessed waveform to form a continuous systolic blood pressure time series. A Fast Fourier Transform (FFT) was performed on this series to calculate the power integral values ​​of the low-frequency band (0.04–0.15Hz) and high-frequency band (0.15–0.4Hz) in the spectrum, and the ratio LF / HF was obtained as a frequency-domain index of blood pressure variability. For respiratory signals, the length of continuous respiratory cycles was calculated based on the effective IMF component, and its standard deviation was obtained as an index of respiratory cycle variability. Furthermore, sample entropy was calculated to characterize the complexity of the respiratory rhythm. After each type of index was calculated, the same start timestamp as the corresponding signal segment was appended and written into a timestamp index table. The timestamp index table uses a hash mapping structure, with timestamps as keys and three types of indicators as values, achieving precise matching of the three types of indicators by timestamp. If an indicator cannot be extracted due to poor signal quality, the system checks if a valid value exists in the previous window. If so, a forward padding strategy is used to complete the value. If more than two consecutive windows are missing, linear interpolation is used to fill in the gaps between valid values, but the padding operation can only be performed a maximum of two consecutive times. Finally, an aligned variability feature set containing nine dimensions is generated. Dimensions 1–3 are time-domain indicators of heart rate variability, dimensions 4–6 are frequency-domain indicators of blood pressure variability (including LF power, HF power, and LF / HF ratio), and dimensions 7–9 are nonlinear indicators of respiratory variability (respiratory cycle standard deviation, sample entropy, and respiratory rate). Each feature dimension is labeled with its source type and corresponding timestamp.

[0025] After receiving the aligned set of variability features, the weight fusion module first retrieves user registration information, including age, gender, history of underlying diseases (such as hypertension, diabetes, heart failure, etc.), and a list of commonly used medications (such as beta-blockers, calcium channel blockers, etc.). Based on the combination of age range (e.g., <40 years, 40–65 years, >65 years) and underlying diseases, it retrieves the corresponding feature weight bias vector from a preset rule base. This vector is 9-dimensional, with initial values ​​set by clinical experts based on pathophysiological mechanisms, and can be dynamically updated through subsequent model training. Subsequently, for each feature in the current set of variability features, it queries the locally stored historical data of the last 7 days, calculates the mean μ and standard deviation σ of the feature within the historical window, and performs Z-score standardization: (x−μ) / σ. The standardized 9-dimensional feature vector is multiplied element-wise with the weight bias vector to obtain preliminary weighting coefficients. Then, Softmax normalization is performed on this coefficient vector to ensure that the sum of all weights is 1, forming a dynamic weight allocation matrix. The standardized feature vector is multiplied by the weight matrix to generate an 8-dimensional initial joint variability vector (one dimension is removed due to redundancy or excessive correlation, with the specific removal rule based on historical covariance analysis). This initial vector is then bound to the current timestamp and stored in a feature queue. The queue adopts a first-in, first-out structure with a maximum capacity of 10 vectors.

[0026] The window adjustment module continuously monitors the initial joint variability vectors of the three most recent windows in the feature queue, calculates the Euclidean distance between each pair, and takes the maximum value as the fluctuation amplitude of the current physiological state. The system presets the upper limit of the fluctuation threshold to be 0.35 and the lower limit to be 0.20. If the fluctuation amplitude is greater than 0.35, it is determined that the current physiological state is in a stage of violent fluctuation, and the sliding window length is adjusted from the default 300 seconds to 180 seconds to improve the temporal resolution; if the fluctuation amplitude is less than 0.20 three times in a row, it is determined that the physiological state tends to be stable, and the window length is restored to 300 seconds. After the window length is adjusted, the signal preprocessing module and the feature alignment module re-execute the signal segmentation, feature extraction, and alignment process with the new window length to generate the variability feature set under the new window, and generate a new joint variability vector through the weight fusion module. After the new vector is generated, a time continuity check is performed with the oldest vector in the feature queue: the difference between the two timestamps is calculated, and if it exceeds 1.5 times the current window step size (default 30 seconds) (i.e., 45 seconds), an interpolation vector is inserted between the two, and the interpolation method is linear interpolation. The optimized joint variability vector is appended with window length metadata (e.g., "window_length=180"), and the oldest record in the feature queue is replaced, maintaining the queue length at a constant 10.

[0027] The evaluation output module loads a user-specific Multilayer Perceptron (MLP) model. This model structure includes an input layer (8 dimensions), three hidden layers (64 neurons per layer, with ReLU activation function), and an output layer (1 neuron, with Sigmoid activation function). Model parameters are initialized based on general population data during the user's first use, and subsequently fine-tuned using evaluation results collected during each user session and subsequent clinical feedback. The optimized joint variability vector is fed into the model as input, and after layer-by-layer forward propagation, it outputs a comprehensive evaluation score between 0 and 1. The system sets a high-risk interval as a score >0.8 or <0.2: if the score is >0.75 and ≤0.8, it is marked as "sympathetic overactivation warning"; if the score is >0.8, it is marked as "high-risk sympathetic hyperactivity"; if the score is <0.25 and ≥0.2, it is marked as "parasympathetic function suppression warning"; if the score is <0.2, it is marked as "high-risk parasympathetic exhaustion". The evaluation results and scores are written together into a structured output buffer. If the absolute change in scores between two consecutive assessments exceeds 0.2, the model retraining process is triggered: extract all optimized joint variability vectors and their corresponding clinical annotations (if any) from the local database within the last 24 hours, update the model weights using stochastic gradient descent, and store the updated parameters in the device's built-in security chip after AES-256 encryption, which can only be accessed by this device.

[0028] When the comprehensive assessment score falls into the high-risk range (>0.8 or <0.2), the early warning storage module is activated. This module first backtracks through the circular buffer based on the timestamp of the current assessment result, extracting the raw multi-channel signals from 5 minutes before and after that timestamp (i.e., a total of 10 minutes of raw data). The raw ECG, arterial blood pressure, and respiratory signals are then concatenated with metadata such as the assessment score, risk category, and window length into a JSON-formatted payload. A unique key (burned into the secure element at the factory) is used to initialize the key scheduler for the AES-256 encryption algorithm, encrypting the JSON payload and generating a ciphertext data packet. Simultaneously, metadata records are generated, including the event start timestamp (ISO8601 format), duration (600 seconds), risk level (high / medium / low), and encryption algorithm identifier. This ciphertext data packet and metadata are written to a local SQLite database. The database table structure includes the fields: timestamp (primary key), encrypted_payload, risk_level, window_length, and model_version. After successful writing, an encrypted warning notification is pushed to the user's bound mobile terminal via Bluetooth or Wi-Fi module. The notification content only includes the risk level and time, and does not include the original physiological data.

[0029] The hash chain building module runs periodically (once per hour by default), traversing all records in the SQLite database in ascending order of timestamp. For each record, it calculates the SHA3-256 hash value of the encrypted_payload field, which is then used as the hash H of the current block. n H n Chain hash H with the previous record n ₋1 Concatenate (set the H0 predecessor of the first record to a 64-bit all-zero string), calculate SHA3-256 again, and generate a chained hash C. n =SHA3-256(H n ||C n ₋1). The C n The value is written to the chain_hash field of the corresponding database record. The system supports an integrity verification mode: starting from the latest record, the expected chain hash is calculated recursively in reverse. If the chain_hash field value of any record is inconsistent with the calculated value, that record and all subsequent records are marked as "data abnormal" and prohibited from use in clinical retrospection. In a clinical retrospection request, the doctor's terminal specifies a time range, and the system returns all records within that range and their chain_hash fields. Third parties can verify the integrity and immutability of the entire physiological record using a publicly available hash algorithm.

[0030] The entire system's modules are connected via an internal message bus. The signal preprocessing module outputs a multimodal signal sequence to a circular buffer. The feature alignment module reads data from the buffer and outputs a set of variability features to the weight fusion module. The weight fusion module generates an initial vector and sends it to the window adjustment module. The window adjustment module outputs an optimized vector to the evaluation output module. The evaluation output module triggers an early warning storage module, which writes the data to the database and is periodically processed by the hash chain construction module. All data streams carry timestamps to ensure consistency between processing order and timing. The system runs on an embedded Linux platform, with each module deployed as an independent process or thread. Synchronization is achieved through shared memory and semaphores, ensuring real-time performance and reliability.

[0031] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principles of this invention are further supplemented below with a specific application scenario.

[0032] In the intensive care unit, a 68-year-old male patient was admitted due to acute heart failure. He had a history of hypertension and type 2 diabetes and had been taking metoprolol and amlodipine long-term. Medical staff deployed the multi-channel physiological signal acquisition device described in this invention: a standard lead II ECG electrode was attached to his chest wall and connected to an analog input channel with a sampling rate of 1000Hz; a non-invasive cuff-type arterial pressure sensor was wrapped around his upper arm, continuously recording arterial blood pressure waveforms at a sampling rate of 200Hz; a respiratory impedance chest strap was fixed to the 4th to 6th intercostal spaces, acquiring respiratory impedance signals at a sampling rate of 50Hz. The three sensor cables were connected to the same data acquisition card, which has a built-in high-precision temperature-compensated crystal oscillator clock source and uses a hardware synchronous trigger controller to simultaneously start the three-channel ADC conversion at t=0, thereby eliminating millisecond-level time offsets caused by asynchronous sampling. This ensures that the physiological events (such as R waves, peak systolic blood pressure, and inspiratory initiation points) upon which subsequent variability analysis depends are strictly aligned on the original time axis. This is a key technical means to solve the problem of timing mismatch in multi-source signals.

[0033] After the raw signal enters the signal preprocessing module, the ECG signal is decomposed into five layers of db6 wavelet. High-frequency detail coefficients are soft-thresholded using a general threshold to suppress electromyographic noise. Then, morphological opening and closing operations are performed to eliminate baseline drift caused by respiratory motion, and the steep rising edge of the QRS complex is preserved to ensure the accuracy of R-wave detection. The arterial blood pressure signal is fitted with a Savitzky-Golay filter (window length 51, order 3) to smooth the slow baseline fluctuations without distorting the pulse wave morphology, so that the systolic blood pressure peak positioning error is controlled within ±2 mmHg. After the respiratory signal is decomposed by EMD, the instantaneous frequency of each IMF component is calculated by Hilbert transform. Only the main IMF with a center frequency falling within 0.1–0.4 Hz (corresponding to a respiratory rate of 6–24 breaths / minute) is retained, effectively eliminating non-physiological high-frequency components introduced by body movement or equipment vibration. The three preprocessed signals are then aligned according to the timestamps of the original sampling points, and uniformly resampled to 200Hz through cubic spline interpolation to generate a multimodal sequence with strict time synchronization. This process solves the phase misalignment problem caused by inconsistent time granularity in the feature extraction stage of signals with different sampling rates.

[0034] The multimodal sequence is written into a circular buffer and stored in segments with 300-second windows and 30-second steps. When the buffer is full of 10 segments, the newest segment overwrites the oldest segment and triggers the feature alignment module. This module uses the Pan-Tompkins algorithm to accurately locate the R wave from the latest 300-second segment and calculates SDNN, RMSSD, and pNN50. It extracts the systolic blood pressure value from the blood pressure waveform beat by beat, and obtains the low-frequency power (0.04–0.15Hz, reflecting sympathetic-adrenal medullary activity) and high-frequency power (0.15–0.4Hz, reflecting vagal modulation) through FFT, and calculates the LF / HF ratio. It calculates the standard deviation of the continuous respiratory cycle length and the sample entropy (embedding dimension m=2, similarity tolerance r=0.2×standard deviation) from the effective respiratory IMF to quantify the regularity and complexity of the respiratory rhythm. All three indicators are appended with the same start timestamp and written into a hash index table. If a window is temporarily lost due to a patient's cough, the valid value of the previous window is retrieved and filled forward. If two consecutive windows are missing, linear interpolation is performed based on the valid windows before and after, but the consecutive interpolation is limited to no more than two times, so as to ensure data continuity while avoiding excessive fabrication of physiological states.

[0035] The weighted fusion module reads patient registration information, identifies patients over 65 years old with heart failure and diabetes, and retrieves the corresponding 9-dimensional weight bias vector from the rule base (e.g., RMSSD and respiratory sample entropy weights are increased, LF / HF weights are decreased to adapt to the pathological characteristics of generally reduced parasympathetic function in elderly heart failure patients). Simultaneously, the system queries locally stored historical data from the past 7 days, calculates Z-score standardized values ​​for each of the current 9-dimensional features, and eliminates inter-individual differences in dimensions and dynamic range. The standardized vector and bias vector are element-wise multiplied and then normalized using Softmax to form a dynamic weight matrix. This matrix is ​​then dot-producted with the standardized features to generate an 8-dimensional initial joint vector. Respiratory frequency is automatically removed because it is highly correlated with the standard deviation of the respiratory cycle (historical covariance > 0.85). This mechanism reduces the interference of redundant features on the fusion result through a data-driven approach.

[0036] The window adjustment module monitors the joint vector of the three most recent windows in the feature queue in real time and calculates their pairwise Euclidean distances. Monitoring revealed that a patient experienced a sudden change in heart rate at 3:15 AM due to paroxysmal atrial fibrillation, with the maximum distance between the three windows reaching 0.41, exceeding the threshold of 0.35. The system immediately switched the sliding window length from 300 seconds to 180 seconds. The signal preprocessing and feature alignment module then re-segmented the data using the new window to improve temporal resolution and capture rapid autonomic nervous system imbalance events. The newly generated optimized vector was compared with the oldest vector in the queue to verify temporal continuity. If a timestamp interval reached 50 seconds (>45 seconds threshold) due to system scheduling delays, an intermediate vector was generated through linear interpolation to ensure no jumps in the time series within the queue and maintain the temporal integrity of subsequent model inputs.

[0037] The assessment output module loads the patient-specific MLP model (fine-tuned based on 24-hour data on the first day of admission), and outputs a comprehensive assessment score of 0.83 after inputting an 8-dimensional optimized vector. Because this value > 0.8, the system determines it as "high-risk sympathetic hyperactivity," triggering the early warning storage module. This module backtracks to the circular buffer, extracting 10 minutes of raw ECG, blood pressure, and respiratory signals from 03:10 to 03:20, and encapsulates them along with metadata such as the score, risk label, and window_length=180 into a JSON payload. The unique key in the device's secure element initializes the AES-256 key scheduler, encrypts the payload, and writes it to the SQLite database with the ISO8601 timestamp "2024-06-15T03:15:00+08:00" as the primary key.

[0038] Subsequently, the hash chain building module runs at the top of the hour, traversing the database records. It calculates the SHA3-256 hash of the aforementioned encrypted payload to obtain H. n Combine it with the chain hash C of the previous record. n ₋1 is concatenated and then hashed again to generate C. n The data is then stored in the `chain_hash` field. Several days later, the doctor initiates a clinical retrospective request, specifying the time period from 03:00 to 04:00. The system returns all records for that time period and their `chain_hash` values. A third party verifies C by recursively tracing from the first record using the publicly available SHA3-256 algorithm. n Consistency ensures that the data has not been tampered with, thereby guaranteeing the legal validity and scientific credibility of long-term monitoring data.

[0039] Throughout the process, each module transmits timestamped messages in the shared memory of the embedded Linux platform. A semaphore mechanism ensures that the feature alignment module is only awakened after a complete 300-second segment has been written to the circular buffer, avoiding the reading of incomplete data. The evaluation output module receives the optimized vector only after the window adjustment module completes the time continuity check, ensuring that the input model data is reliable both temporally and numerically. These mechanisms work synergistically to fundamentally solve the challenges of temporal consistency, individual adaptability, dynamic responsiveness, and data integrity in multi-source physiological signal fusion.

[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting combined variability indicators of the cardiovascular system, characterized in that, include: Electrocardiogram (ECG), arterial blood pressure (AP) and respiratory signals are acquired simultaneously using a multi-channel physiological signal acquisition device. An adaptive filtering algorithm is then used to denoise and correct the baseline of each signal to obtain a preprocessed multimodal signal sequence. Based on the multimodal signal sequence, time-domain indicators of heart rate variability, frequency-domain indicators of blood pressure variability, and nonlinear indicators of respiratory variability are extracted respectively. A timestamp alignment mechanism is used to map the three types of indicators into a unified sliding window to generate an aligned variability feature set. The variability feature set is subjected to Z-score-based standardization, and a dynamic weight allocation matrix is ​​constructed by combining individual physiological parameters. An initial joint variability vector is generated by using a weighted fusion strategy. The sliding window length adaptive adjustment module dynamically adjusts the window length according to the fluctuation range of the current physiological state. If the fluctuation range exceeds the preset threshold, the window length is shortened; otherwise, the window length is extended to obtain the optimized joint variability vector. The optimized joint variability vector is input into a multilayer perceptron model trained individually, and a comprehensive evaluation score reflecting the state of autonomic nervous regulation is output. If the comprehensive evaluation score falls into the high-risk range, the local early warning logic is triggered, and the original signal fragment and evaluation result are encapsulated using the AES-256 encryption algorithm and written to the local database with a timestamp index. Based on the timestamp index in the local database, a time series hash chain is constructed, and the hash value is calculated segment by segment using the SHA3-256 algorithm to form an immutable physiological record chain for subsequent trend analysis and clinical retrospection.

2. The method for detecting combined variability indicators of the cardiovascular system according to claim 1, characterized in that, The steps of simultaneously acquiring electrocardiogram (ECG), arterial blood pressure, and respiratory signals using a multi-channel physiological signal acquisition device, and then using an adaptive filtering algorithm to denoise and correct the baseline of each signal to obtain a preprocessed multimodal signal sequence include: Synchronous acquisition is initiated by a hardware-level synchronous trigger signal using ECG electrodes with a sampling frequency of 1000Hz, an arterial pressure sensor with a sampling frequency of 200Hz, and a respiratory impedance sensor with a sampling frequency of 50Hz, to obtain the original multi-channel signal. Wavelet threshold denoising combined with morphological filtering was used to remove electromyographic interference and power frequency noise from the electrocardiogram signal while preserving the QRS complex characteristics. Savitzky-Golay smoothing filter was used on the arterial blood pressure signal to eliminate pulse wave baseline drift and retain the peak points of systolic and diastolic blood pressure. Empirical mode decomposition was used to extract intrinsic mode functions from the respiratory signal, and the IMF components corresponding to the main frequency band were selected as the effective respiratory waveforms. The three preprocessed signals are aligned according to the sampling points, and interpolation is used to resample them to a uniform sampling rate of 200Hz to generate a time-aligned multimodal signal sequence. The multimodal signal sequence is segmented with a sliding window of 300 seconds and a step size of 30 seconds, and each segment is independently marked with a start timestamp; Each signal segment is stored in a circular buffer. When the buffer is full, the earliest segment is automatically overwritten, and the feature extraction process is triggered simultaneously.

3. The method for detecting combined variability indicators of the cardiovascular system according to claim 1, characterized in that: The step of extracting time-domain indices of heart rate variability, frequency-domain indices of blood pressure variability, and nonlinear indices of respiratory variability from the multimodal signal sequence, and mapping the three types of indices to a unified sliding window using a timestamp alignment mechanism to generate an aligned variability feature set includes: The peak value of the R wave is detected from the electrocardiogram signal, the adjacent RR interval is calculated, and SDNN, RMSSD, and pNN50 are extracted as time-domain indicators. A continuous systolic blood pressure sequence was extracted from the arterial blood pressure signal, and a fast Fourier transform was performed. The power ratio of the low frequency (0.04–0.15 Hz) to the high frequency (0.15–0.4 Hz) was calculated as a frequency domain index. The standard deviation of the respiratory cycle and the sample entropy are calculated from the respiratory signals as indicators of nonlinear variability. Add the same start timestamp as its corresponding signal segment to each type of indicator; Establish a timestamp index table to precisely match the three types of indicators to the same sliding window based on their timestamps; If a certain type of indicator is missing due to signal quality, forward padding combined with linear interpolation will be used to complete it, and the number of completions will not exceed two consecutive times. Generate an aligned variable feature set containing 9-dimensional features, with each feature labeled with its source type and timestamp.

4. The method for detecting combined variability indicators of the cardiovascular system according to claim 1, characterized in that: The steps of performing Z-score-based standardization on the variability feature set, constructing a dynamic weight allocation matrix in conjunction with individual physiological parameters, and generating an initial joint variability vector using a weighted fusion strategy include: Obtain the age, gender, medical history, and list of commonly used medications entered by the user during registration; Based on age and history of underlying diseases, the corresponding feature weight bias vector is retrieved from the preset rule base; For each feature in the variability feature set, calculate its mean and standard deviation in the most recent 7 days of historical data, and perform Z-score standardization; The standardized feature vector is multiplied element-wise with the weight bias vector to generate individualized weighted coefficients. The weighted coefficients are Softmax normalized to ensure that the sum is 1, thus forming a dynamic weight allocation matrix. The standardized feature vector is multiplied by the dynamic weight assignment matrix to generate an 8-dimensional initial joint variability vector. Bind the initial joint variability vector to the current timestamp and store it in the feature queue.

5. The method for detecting combined variability indicators of the cardiovascular system according to claim 1, characterized in that: The step of dynamically adjusting the window length according to the fluctuation amplitude of the current physiological state through the sliding window length adaptive adjustment module, shortening the window length if the fluctuation amplitude exceeds a preset threshold, and extending the window length otherwise, to obtain the optimized joint variability vector, includes: Calculate the Euclidean distance between the joint variability vectors of the three most recent windows, and take the maximum value as the current fluctuation amplitude; Set the fluctuation threshold to 0.

35. If the fluctuation amplitude is greater than 0.35, adjust the sliding window length from 300 seconds to 180 seconds. If the fluctuation amplitude is less than 0.20 three times in a row, the window length will be restored to 300 seconds. After adjusting the window length, the feature extraction and fusion process is re-executed to generate a joint variability vector under the new window. Perform a time continuity check between the new vector and the old vector. If the time interval exceeds 1.5 times the window step size, then insert an interpolated vector. Attach window length metadata to the optimized joint variability vector for use as subsequent model input; Replace the oldest record in the feature queue with the optimized vector, while maintaining the queue length at 10.

6. The method for detecting combined variability indicators of the cardiovascular system according to claim 1, characterized in that: The step of inputting the optimized joint variability vector into an individually trained multilayer perceptron model and outputting a comprehensive evaluation score reflecting the state of autonomic neural regulation includes: Load a user-specific multilayer perceptron model, which has been fine-tuned based on the user's historical data and contains 3 hidden layers with 64 neurons in each layer; The optimized joint variability vector is used as input and propagated layer by layer through the ReLU activation function; The output layer uses the Sigmoid function to generate a comprehensive evaluation score between 0 and 1; If the score is greater than 0.75, it is considered as excessive activation of the sympathetic nervous system; if it is less than 0.25, it is considered as inhibition of parasympathetic function. Write the evaluation score and the judgment category to the structured output buffer; If the score changes by more than 0.2 in two consecutive evaluations, the model retraining process is triggered, and the weights are updated using data from the most recent 24 hours. The updated model parameters are encrypted and stored in a local secure chip.

7. The method for detecting combined variability indicators of the cardiovascular system according to claim 1, characterized in that: The step of triggering a local early warning logic if the comprehensive evaluation score falls into the high-risk range, and encapsulating the original signal segment and evaluation result using the AES-256 encryption algorithm and writing it into a local database with a timestamp index, includes: When the comprehensive assessment score is greater than 0.8 or less than 0.2, it is determined to be in the high-risk range; Extract the raw multi-channel signals from the circular buffer within 5 minutes before and after the corresponding timestamp; The signal data and evaluation results are concatenated into a JSON format payload. The payload is encrypted using the device's unique key and the AES-256 algorithm. The metadata for generating encrypted data packets includes the start timestamp, duration, and risk level. Write the encrypted data packet and metadata into the SQLite database, with the primary key being an ISO8601 format timestamp; After the database write is successful, an encrypted warning notification is pushed to the user terminal.

8. The method for detecting combined variability indicators of the cardiovascular system according to claim 1, characterized in that: The steps of constructing a time-series hash chain based on the timestamp index in the local database, calculating hash values ​​segment by segment using the SHA3-256 algorithm to form an immutable physiological record chain for subsequent trend analysis and clinical retrospection include: Iterate through all records in the local database in ascending order of timestamp; Calculate the SHA3-256 hash value for the encrypted data packet of each record, and use it as the hash of the current block; The current block hash is concatenated with the previous block hash and then hashed again to generate a chained hash. The hash of the first record is set to an all-zero string. Append the chained hash value to the validation field of the corresponding database record; Perform integrity verification periodically, recursively tracing the hash chain backwards from the latest record, and marking the data as abnormal if any verification fails. During clinical retrospective analysis, a continuous hash chain is returned by querying the time range, allowing third parties to verify that the data has not been tampered with.

Images