Cardiovascular trend prediction method and system based on time sequence medical health data

Abstract

The invention provides a cardiovascular trend prediction method and system based on time sequence medical health data, and relates to the technical field of medical health data analysis. The method comprises the following steps: collecting time sequence medical health data of a patient through a multi-source sensor, wherein the time sequence medical health data comprises dynamic physiological indexes such as heart rate, blood pressure and oxyhemoglobin saturation; performing time calibration and feature extraction on the acquired multi-source data; constructing a time sequence feature model based on the time sequence features, and fusing the time sequence feature model with the static information of the patient to generate a comprehensive feature vector; using a deep learning model to train historical data, and learning a dynamic change rule of cardiovascular health indexes; predicting the change trend of the cardiovascular health indexes in real time, and evaluating the risk level of cardiovascular diseases; the model parameters are optimized through online learning, the prediction precision is improved, the change trend of cardiovascular health indexes can be predicted in real time, and support is provided for early discovery and personalized medical treatment of cardiovascular diseases.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical health data analysis, and more specifically, to a cardiovascular trend prediction method and system based on time series medical health data. BACKGROUND

[0002] Early detection and intervention are key to reducing the risk of cardiovascular disease. Traditional medical monitoring methods mainly rely on static data (such as physical examination reports, medical records, etc.), which can only reflect the health status of patients at a certain time point and cannot capture the dynamic change trend of cardiovascular health indicators. In addition, although some existing dynamic monitoring devices can collect real-time data such as heart rate and blood pressure, they lack deep analysis capabilities for time series data and cannot accurately predict the potential risk of cardiovascular disease.

[0003] With the development of wearable devices and telemedicine technology, more and more patients can continuously collect time series data such as heart rate, blood pressure, and blood oxygen saturation through smart bands, smart watches, and other devices. However, the large amount of time series data generated by these devices is often not fully utilized, and existing technologies fail to effectively integrate multi-source time series data and perform trend analysis, resulting in insufficient prediction accuracy and inability to provide reliable decision support for doctors and patients. SUMMARY

[0004] To solve the above technical problems, the present application provides a cardiovascular trend prediction method and system based on time series medical health data, which can to some extent solve the problem that existing monitoring devices lack deep analysis capabilities for time series data and cannot accurately predict the potential risk of cardiovascular disease.

[0005] According to one aspect of the present application, a cardiovascular trend prediction method based on time series medical health data is provided, comprising:

[0006] S1: Collecting time series medical health data of patients, including dynamic physiological indicators of patients, including heart rate, blood pressure, and blood oxygen saturation, and static information of patients, including age, gender, and medical history;

[0007] S2: Preprocessing the time series medical health data to obtain time series features: for sampling frequency difference, using a cubic spline interpolation function for time alignment; for heart rate data, calculating sliding window statistics, determining and removing outliers, and using a linear interpolation formula to fill in missing values; for blood pressure data, using a weighted moving average filter to smooth noise;

[0008] S3: Extracting time domain, frequency domain, and clinical medical features from the time series features to obtain multi-dimensional features;

[0009] S4: fuse the multi-dimensional features with the static information of the patient; wherein the time sequence channel extracts dynamic features, and the static channel converts the age and gender structured data into high-dimensional vectors through an embedding layer;

[0010] S5: construct and train a deep learning model; the deep learning model adopts a bidirectional long short-term memory network to process the time sequence channel data, and introduces a residual connection between network layers, and finally splices and fuses the hidden state and the high-dimensional vector of the static channel as the input of the full connection layer;

[0011] S6: use the deep learning model to predict the changes of the index in real time, match with the static features, and evaluate the cardiovascular disease risk level.

[0012] Preferably, the dynamic physiological indexes collected in step S1 also include activity amount, body temperature and electrocardiogram data, and the multi-source sensor includes a wearable device and a home medical monitoring device.

[0013] Preferably, in step S2, the cubic spline interpolation function is specifically:

[0014] Let the low-frequency data point sequence be , wherein is the timestamp, is the measured value. On the interval , the spline function is a cubic polynomial:

[0015] ;

[0016] wherein represents the function value at , represents the starting value of the interval, represents the slope at , represents the change rate, is the curvature at , represents the change acceleration, is the cubic term coefficient, and controls the bending degree of the curve.

[0017] Preferably, the heart rate data is recorded once per second, including the resting heart rate and the instantaneous fluctuation during exercise; the blood pressure data is automatically measured once every 5 minutes, recording the systolic pressure, diastolic pressure and pulse frequency; the blood oxygen saturation is detected by a red light sensor once per minute; the electrocardiogram data is continuously collected at a sampling rate of 250Hz, capturing the morphological changes of P wave and QRS complex.

[0018] Preferably, the time-domain features include statistics of each index in the sliding window, including mean, standard deviation, skewness, and zero-crossing rate; the frequency-domain features include: decomposing the signal into different frequency components by fast Fourier transform, 0.04-0.15Hz as low-frequency components, 0.15-0.4Hz as high-frequency components, and calculating the power ratio of low-frequency and high-frequency components.

[0019] Preferably, the clinical medical features include pulse pressure difference between systolic pressure and diastolic pressure, morning peak blood pressure, and nighttime blood pressure drop rate.

[0020] Preferably, in the step S4 of fusing the time-domain features and the frequency-domain features with the static information, a weighted fusion strategy is adopted, wherein the weights of the time-domain features and the frequency-domain features are dynamically adjusted according to their contribution to the risk of cardiovascular disease, and the weight of the static information is modified according to the patient's medical history and family history.

[0021] Preferably, in the step S6 of evaluating the cardiovascular disease risk level, a preset medical threshold and a risk assessment matrix are combined, wherein the risk assessment matrix comprehensively considers the patient's age, gender, medical history, and real-time physiological indicators.

[0022] Another aspect of the present application provides a cardiovascular trend prediction system based on time-series medical health data, comprising:

[0023] The acquisition module acquires time-series medical health data of the patient, including dynamic physiological indicators of the patient, including heart rate, blood pressure, and blood oxygen saturation, and static information of the patient, including age, gender, and medical history;

[0024] The preprocessing module pre-processes the time-series medical health data to obtain time-series features: for the difference in sampling frequency, a cubic spline interpolation function is used for time alignment; for heart rate data, sliding window statistics are calculated, abnormal values are determined and removed, and a linear interpolation formula is used to fill in the missing values; for blood pressure data, a weighted moving average filter is used to smooth the noise;

[0025] The extraction module extracts time-domain, frequency-domain, and clinical medical features from the time-series features to obtain multi-dimensional features.

[0026] The vector generation module fuses the multi-dimensional features with the static information of the patient; wherein the time-series channel extracts dynamic features, and the static channel converts the age and gender structured data into high-dimensional vectors through an embedding layer.

[0027] The model construction module constructs and trains a deep learning model; the deep learning model uses a bidirectional long short-term memory network to process time-series channel data, and introduces residual connections between network layers, and finally concatenates and fuses the hidden state and the high-dimensional vector of the static channel as the input of the full connection layer.

[0028] A prediction module uses a deep learning model to predict changes in indicators in real time and matches them with static features to assess cardiovascular disease risk levels.

[0029] In another aspect, the application provides an electronic device, comprising: one or more processors; a memory storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method.

[0030] Compared with the prior art, the application has the following beneficial effects:

[0031] Through the collection and fusion of multi-source time series data and the combination of deep learning models, the application can predict the trend of cardiovascular health indicators in real time and provide accurate risk assessment. The system not only improves the early detection ability of cardiovascular diseases, but also provides strong support for personalized medicine and remote monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor. In the drawings:

[0033] Figure 1 A block diagram of a cardiovascular trend prediction system based on time series medical health data according to an embodiment of the application. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solutions and advantages of the application more clear, the application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and not to limit the application. In addition, the technical features involved in each embodiment of the application described below can be combined with each other as long as they do not conflict with each other. In order to achieve the above purpose, the application adopts the following technical solutions.

[0035] A cardiovascular trend prediction method based on time series medical health data, comprising:

[0036] S1: Collecting time series medical health data of a patient, including dynamic physiological indicators of the patient, including heart rate, blood pressure, blood oxygen saturation, and static information of the patient, including age, gender, medical history;

[0037] S2: Preprocessing of time series medical health data to obtain time series features: For the difference in sampling frequency, a cubic spline interpolation function is used for time alignment; for heart rate data, sliding window statistics are calculated, abnormal values are determined and removed, and linear interpolation formula is used to fill in the missing; for blood pressure data, a weighted moving average filter is used to smooth the noise;

[0038] S3: Time domain, frequency domain and clinical medical feature extraction of time series features to obtain multi-dimensional features;

[0039] S4: Fusion of multi-dimensional features and patient's static information; wherein, the time series channel extracts dynamic features, and the static channel converts the age, gender structured data into a high-dimensional vector through an embedding layer;

[0040] S5: Construct and train deep learning model; the deep learning model uses a bidirectional long short-term memory network to process time series channel data, and introduces residual connection between network layers, and finally splices and fuses the hidden state and high-dimensional vector of the static channel as the input of the full connection layer;

[0041] S6: Real-time prediction of index changes using a deep learning model, and matching with static features to assess cardiovascular disease risk levels.

[0042] Step S1 includes: multi-source data acquisition and dynamic management, multi-source sensors including wearable devices and home medical monitoring devices; real-time acquisition of dynamic physiological index data of patients through wearable devices (such as smart watches with medical certification, heart rate chest straps) and home health monitoring devices (such as Bluetooth connected electronic sphygmomanometers, blood oxygen rings). Dynamic physiological indicators are continuously uploaded to the cloud server at different frequencies: dynamic physiological indicators include heart rate, blood pressure, blood oxygen saturation, etc., as well as activity, body temperature and electrocardiogram data, heart rate data is recorded once per second, including resting heart rate and instantaneous fluctuations during exercise; blood pressure data is measured automatically every 5 minutes, recording systolic pressure, diastolic pressure and pulse frequency; blood oxygen saturation is detected by a red light sensor every minute; electrocardiogram data is continuously collected at a high sampling rate of 250Hz, capturing the shape changes of P wave and QRS complex. Static information is imported through medical records when the user first registers, including age, gender, medical history (such as hypertension duration), surgical records, family genetic history (whether immediate relatives have had heart attack or stroke), etc., and is associated and bound with dynamic data through a unique encrypted patient ID.

[0043] Meanwhile, the state of each sensor is monitored in real time: when it is detected that a device has not uploaded data for 30 seconds in a row (such as a smart watch with less than 10% battery), the system automatically triggers a three-level response mechanism - first, it attempts to reconnect the device through Bluetooth / WiFi; if it fails after 3 attempts, it pushes a device anomaly reminder to the patient's mobile phone APP; for critical indicators (such as interruption of electrocardiogram monitoring), a short message is sent to the emergency contact person. All transmission channels use TLS 1.3 encryption protocol, and patient sensitive data is desensitized on the edge computing node (such as replacing the ID card number with a hash value).

[0044] In step S2, the multi-source data is time calibrated, and a spline interpolation method is used for synchronous processing of data with different sampling frequencies. For the difference in sampling frequencies of multi-source sensors (such as heart rate per second, blood pressure every 5 minutes), the application uses a cubic spline interpolation function for time synchronization, as follows:

[0045] Let the low-frequency data point sequence be , where is the timestamp, and is the measured value. On the interval , the spline function is a cubic polynomial:

[0046]

[0047] where represents the function value at , represents the starting value of the interval, represents the slope at , represents the rate of change, is the curvature at , represents the acceleration of change, is the coefficient of the cubic term, which controls the degree of curvature of the curve.

[0048] The coefficients are solved by boundary conditions (such as natural boundary conditions ) and continuity conditions ( , , continuity). This interpolation ensures smooth data and continuous derivatives, which is better than simple linear interpolation. Spline interpolation reduces information loss due to different sampling rates.

[0049] Noise filtering and outlier processing are performed, including: for heart rate data, 3σ principle is used to detect and eliminate outliers; for blood pressure data, a moving average filter is used for smoothing. Specifically:

[0050] Time alignment and data cleaning: To address the timestamp difference between different devices (e.g., wristwatch clock and sphygmomanometer have a second-level deviation), the network time protocol (NTP) is used to synchronize all devices at the millisecond level, ensuring the comparability of data collected at the same time. For the problem of sampling frequency difference, such as blood pressure measured every 5 minutes and heart rate recorded every second, the system uses an interpolation algorithm for data alignment: the blood pressure data measured at low frequency is generated every second by time interpolation algorithm to estimate the value, so that all indicators have a unified time reference.

[0051] Among them, the data cleaning stage implements a hierarchical processing strategy including heart rate data: calculate the average value and standard deviation in the past 60 seconds sliding window, automatically remove the abnormal value (such as instantaneous 300 times / heartbeat due to poor device contact) outside the range of 3 times standard deviation, and use linear interpolation to fill in the missing value with the previous and next 5 seconds.

[0052] Specifically, based on the sliding window (e.g., 60 seconds), the mean μ and standard deviation σ are calculated, and the 3σ principle is used:

[0053] Abnormal conditions are .

[0054] After removing abnormal values, linear interpolation is used to fill in:

[0055] ;

[0056] is the heart rate measurement value at the current time point t, μ is the arithmetic mean of the heart rate data in the sliding window, σ is the standard deviation of the heart rate data in the sliding window, and are the heart rate values of the nearest valid data points before and after the abnormal point, and are the timestamps of the corresponding valid data points, t is the time of the abnormal point, is the filled heart rate value.

[0057] Blood pressure data smoothing: a weighted moving average filter is used, with window size k=5 and weight decaying with time:

[0058] where λ is the decay factor, is the smoothed blood pressure value at time t, is the original blood pressure measurement value at time t-i, is the normalization denominator, ensuring that the sum of all weights is 1. Abnormal value processing avoids the interference of noise on trend prediction.

[0059] Blood pressure data: a weighted moving average filter based on exponential decay weight is applied. According to the formula Compute weights, set decay factor The weight of the most recent measurement is higher, and the weight decreases exponentially over time, thus suppressing accidental fluctuations caused by patient arm movement while preserving the latest blood pressure trend.

[0060] ECG data is filtered by a digital filter bank to remove power line interference (50 Hz noise) and myoelectric noise, and an R-wave detection algorithm is used to identify valid heartbeat cycles, and flat segments caused by electrode detachment are removed.

[0061] The cleaned data is stored in a time series database, and each record contains metadata such as timestamp, device ID, index type, value, data quality label (such as whether it has been annotated with interpolation repair), etc.

[0062] In step S3, the time series decomposition method is used to extract the trend, periodic and random terms, and the sliding window technique is used to capture short-term and long-term dependencies, where the size of the sliding window is dynamically adjusted according to the historical fluctuation rules of the specific physiological indicators.

[0063] Specifically, time series feature extraction is divided into three dimensions: time domain, frequency domain and clinical medical features:

[0064] Among them, the time domain feature: calculate the statistics of each index in the sliding window (the default window is 30 minutes), including mean (reflecting the baseline level), standard deviation (characterizing the fluctuation intensity), skewness (distribution asymmetry), zero-crossing rate (change frequency). For example, for heart rate variability (HRV) analysis, the standard deviation (SDNN) of adjacent R-wave intervals (RR intervals) is extracted. Frequency domain feature: the signal is decomposed into different frequency components by fast Fourier transform, and the low frequency (0.04-0.15Hz) and high frequency (0.15-0.4Hz) power ratio (LF / HF) is calculated to evaluate the balance of the autonomic nervous system. Clinical features: based on medical guidelines, key indicator combinations such as pulse pressure difference between systolic and diastolic blood pressure, morning peak blood pressure (blood pressure rising within 2 hours after getting up), and night blood pressure drop rate (judging dipper / non-dipper blood pressure pattern) are extracted. For trend analysis, time series decomposition technology is used to decompose the data into long-term trend (such as continuous weekly blood pressure slowly rising), periodic change (such as natural heart rate reduction at night) and random fluctuation (occasional premature beat event). The sliding window mechanism dynamically adjusts according to the characteristics of the indicators: when the heart rate fluctuates more than 20 times per minute within 10 minutes (such as sudden atrial fibrillation), the window is automatically reduced to 5 minutes to capture the details; when the blood pressure is in a stable period, the window is expanded to 2 hours to smooth the random noise.

[0065] In step S4, when fusing multi-dimensional features and static information, a weighted fusion strategy is used, where the weight of multi-dimensional features is dynamically adjusted according to its contribution to cardiovascular disease risk, and the weight of static information is modified according to the patient's medical history and family history.

[0066] Specifically, the feature fusion and weight distribution includes a differentiated fusion strategy for time-domain and frequency-domain features and static information when constructing the comprehensive feature vector: dynamic feature weight: analyze historical data through a machine learning model to evaluate the contribution of time-domain and frequency-domain features to the prediction of cardiovascular events. For example, if it is found that for every 10% decrease in nighttime blood pressure drop rate, cardiovascular risk increases by 15%, the weight coefficient of this feature is automatically increased to 0.9 (full score is 1). Static feature correction: adjust the weight according to the individual differences of patients. For example, for diabetic patients over 65 years old, the age weight is increased from the baseline value of 0.6 to 0.8, and the diabetes history weight is increased from 0.4 to 0.7; while for young patients with exercise habits, the static disease history weight is appropriately adjusted downward. The fusion process introduces an attention mechanism: the model automatically identifies the most relevant feature combination. For example, when the system detects that the patient's activity level has been zero for 3 consecutive hours and the heart rate has been consistently high, it will increase the attention weight of "resting heart rate" and "heart rate variability", and reduce the influence of irrelevant features such as "blood pressure during exercise". In addition, in order to better fuse heterogeneous data, when constructing the comprehensive feature vector, the static channel maps discrete structured data such as age and gender into dense high-dimensional vectors through an embedding layer, and then performs deep concatenation fusion with time-series features to fully exploit the implicit influence of static information on dynamic trends.

[0067] The deep learning model used in step S5 uses a bidirectional long short-term memory network (Bi-LSTM), and residual connections are introduced between network layers to prevent gradient vanishing. Finally, the hidden state of the LSTM is concatenated with the high-dimensional vector of the static channel and used as input to the fully connected layer. Specifically, the deep learning model training and optimization includes:

[0068] Specifically, the deep learning model training and optimization includes: first, implementing a data balancing strategy: to solve the class imbalance problem in cardiovascular disease data, where the number of diseased samples is much less than that of healthy samples, a combination of SMOTE and ADASYN oversampling strategy is adopted. The SMOTE algorithm is used to perform linear interpolation between minority class samples and their neighbors to increase sample diversity; at the same time, the ADASYN algorithm is used to generate more synthetic samples in the difficult sample area near the decision boundary according to the density distribution of the samples, thereby balancing the positive and negative sample ratio. On this basis, the model architecture adopts a dual-channel design: the time sequence channel: a bidirectional LSTM network is used to process dynamic physiological indicators, each LSTM unit contains 128 neurons and can remember the data pattern of the previous 72 hours. Residual connections are introduced between network layers to prevent gradient vanishing. The static channel: structured data such as age and gender are converted to high-dimensional vectors through an embedding layer and concatenated with the final hidden state of the LSTM. The training strategy is implemented in three stages: pre-training stage: train the base model on a million-level public medical dataset (such as MIMIC-III) to learn general physiological patterns. Transfer learning stage: fine-tune the model using three months of historical data from the target patient, focusing on adjusting the parameters of the fully connected layer to adapt to individual differences. Incremental learning stage: daily new data is input in small batches (256 samples), and the elastic weight consolidation technology (EWC) is used to update the parameters to avoid forgetting the learned common patterns. The model output is a multi-dimensional prediction result, including 6-hour heart rate interval prediction, 24-hour hypertension crisis probability, 7-day heart attack risk score, etc. The input of the deep learning model includes a comprehensive feature vector, which plays a key role in model training: on the one hand, it enables the model to consider both dynamic physiological characteristics and static information of the patient, thereby accurately capturing the impact of individualized features on cardiovascular health; on the other hand, the comprehensive feature vector is fused with the model through an attention mechanism, making the weight distribution of dynamic changes and static risk factors more flexible, thereby significantly improving the training effect.

[0069] In step S6, when evaluating the cardiovascular disease risk level, a preset medical threshold and a risk assessment matrix are combined, wherein the risk assessment matrix comprehensively considers the patient's age, gender, medical history, and real-time physiological indicators.

[0070] Specifically, the hierarchical early warning and clinical decision support includes risk level assessment using a dynamic threshold mechanism: low risk (level 1): when the predicted systolic blood pressure will rise to 130-139 mmHg within 6 hours and there are no other abnormalities, the APP pushes health recommendations such as "suggest reducing salt intake and taking a 30-minute walk". Medium risk (level 2): if the systolic blood pressure is detected to be > 140 mmHg for more than 1 hour, or the heart rate variability SDNN < 50 ms, the system automatically generates an electronic report to the contracted family doctor, prompting to arrange a remote consultation. High risk (level 3): when the ST segment depression is > 0.1 mV with chest pain self-recorded, a red alert is triggered, an emergency call is made immediately, and the patient's real-time positioning and the latest 3 hours of vital signs are sent to the emergency center dispatch system. The early warning logic introduces a time decay factor: for example, a one-time short-term increase in heart rate only produces a temporary early warning, but if the same abnormality occurs 3 times within 24 hours, the risk level is automatically upgraded. All early warning records generate a traceable timeline map for the doctor to review when needed. For example, a heart rate exceeding 100 beats per minute for more than 5 minutes increases the risk level by one level; a blood pressure exceeding 140 / 90 mmHg for more than 10 minutes increases the risk level by one level. Among them, when the risk is low, a prompt message is generated, suggesting that the patient pay attention to his health; when the risk is medium, a warning message is generated, suggesting that the patient consult a doctor; when the risk is high, an emergency early warning message is generated, directly notifying the patient's emergency contact and medical rescue agencies.

[0071] The model parameters are continuously optimized through online learning, using the Adam optimization algorithm, and a learning rate decay strategy is set to prevent overfitting.

[0072] Specifically, continuous learning and system iteration includes a double-check mechanism to ensure the safety of updates in the online learning module: data verification: new data must pass reasonableness checks (such as blood pressure values not exceeding 300 mmHg) and time series continuity detection (rejecting sudden jumps in abnormal data injection). Model verification: the updated model needs to run in shadow mode for 24 hours, predicting in parallel with the old model but not yet effective. When the F1 score of the new model on the test set improves by more than 2% and the false positive rate decreases by 1%, the old version is replaced. Model degradation monitoring runs in real time: when the average absolute error (MAE) between the predicted results and the true measured values increases by more than 15% for 3 consecutive days, it is automatically rolled back to the previous stable version and triggers a data quality review process. Multiple historical version models are retained to support querying the prediction logic for a specific time period by timestamp. New time series data is collected at fixed intervals (such as every 24 hours); the model is fine-tuned using the new data to update the model parameters; the learning rate is adjusted based on the performance of the validation set, and when the validation loss no longer decreases, the learning rate is reduced to 1 / 10 of the original.

[0073] Specifically, the whole-process collaborative operation mechanism adopts a micro-service architecture to realize modular deployment: the data access layer receives device data through the Internet of Things protocol (MQTT) and uses a Kafka message queue to buffer peak traffic to prevent server overload. The stream processing engine uses Apache Flink to perform real-time cleaning, alignment, and feature calculation on the data, with a processing delay controlled within 500 milliseconds. The model service layer deploys multiple versions of models through TensorFlow Serving, supporting thousands of concurrent prediction requests per second. The early warning execution layer is based on a rule engine (such as Drools) to configure 300+ medical rules, and when model prediction and rule base trigger at the same time, the more stringent early warning action is executed first. The operation and maintenance monitoring system includes full-link tracking: each link from sensor data collection to early warning information issuance records indicators such as latency and success rate, and abnormal situations (such as a group of users in a certain area experiencing data delay) automatically trigger geofence alerts to prompt checking of local network infrastructure. The system generates a health report every week, summarizing key indicators such as overall risk trends of patients, device availability statistics, and model prediction accuracy, providing data support for operational decision-making. This process, through closely linked modular design, realizes a closed loop from data perception to intelligent decision-making, meeting the stringent requirements of real-time in the medical scenario, and maintaining the prediction accuracy of the model through a continuous learning mechanism, providing all-weather dynamic monitoring for cardiovascular disease management.

[0074] For example, an old man who takes antihypertensive drugs regularly but occasionally misses a dose, because he wears a smart watch supporting ECG monitoring, and his home is equipped with a networked electronic sphygmomanometer and a blood oxygen meter. One night, the system monitors abnormal fluctuations in his blood pressure and triggers a high-risk warning, successfully avoiding a potential acute myocardial infarction event. At 19:30 that night, the patient records through the APP: chest tightness after dinner, at 22:30: the sphygmomanometer measures a systolic pressure of 162 mmHg (exceeding the patient's baseline of 135 ± 10 mmHg), the system starts a high-frequency monitoring mode, and the watch heart rate monitoring frequency is increased to 2 times per second. At 23:15: the electrocardiogram detects a ST segment depression of 0.12 mV (normal range <-0.05 mV), while the heart rate suddenly increases from 72 beats per minute to 98 beats per minute. At 23:20: the oxygen saturation decreases from 97% to 93%, and the activity sensor confirms that the patient is in a static state. Among them, when the blood pressure is detected to be excessive, the system automatically increases the monitoring frequency to avoid missing short-term abnormalities. By simultaneously analyzing the ST segment changes of electrocardiogram, oxygen saturation decrease and static symptoms (chest tightness), false positives of single indicators (such as short-term decrease of oxygen saturation due to device displacement) are excluded, and the systolic pressure rise slope in the past 30 minutes is calculated (from 140 mmHg to 162 mmHg, slope > 0.7 mmHg / minute). Periodic comparison: comparing the patient's historical night blood pressure regularity (usually decreases by 5-10%), it is found that the current blood pressure has increased by 12%. Symptom relevance: the "chest tightness" symptom is associated with the current ST segment depression and increased heart rate (weighting coefficient from 0.4 to 0.8), and a dynamic baseline is established based on the patient's historical data, rather than using a fixed medical threshold (such as 140 / 90 mmHg), reducing false positives caused by individual differences. Cross-modal enhancement: symptom text data is extracted by NLP model ( "chest tightness", "last for 20 minutes"), converted into feature vector for fusion, improving the model's quantitative ability for subjective symptoms. First-level detection: when the systolic pressure > 160 mmHg and lasts for 10 minutes, trigger a medium-level warning, notify the on-duty doctor. Secondary confirmation: the doctor remotely views the real-time electrocardiogram and 72-hour trend chart, confirms the dynamic change of ST segment, and manually upgrades to a high-risk warning. Tertiary response: the system automatically calls 120 emergency center, sends patient address, medication history (sensitive to nifedipine) and real-time vital signs. Through early warning, the time from symptom occurrence to medical intervention for such high-risk patients is shortened from an average of 4.2 hours to 1.1 hour. This example embodies a complete closed loop from data perception to clinical intervention, and through the organic cooperation of multiple technical modules, it realizes the core goals of precise monitoring, rapid response and continuous optimization, providing a reusable intelligent solution for chronic disease management.

[0075] According to another aspect of the present application, a cardiovascular trend prediction system based on time-series medical health data is provided, such as Figure 1As shown, it comprises: a data acquisition module: for collecting the time series medical health data of the patient; a feature processing module: for feature extraction and preprocessing of the collected time series data; a model training module: for training a deep learning model to learn the dynamic change rule of cardiovascular health indicators; a prediction and evaluation module: for real-time prediction of the change trend of cardiovascular health indicators and evaluation of the risk level of cardiovascular diseases; a feedback optimization module: for optimizing model parameters through online learning to improve prediction accuracy.

Claims

1. A cardiovascular trend prediction method based on time-series medical health data, characterized in that, Comprising: S1: collecting time-series medical health data of a patient, including dynamic physiological indicators of the patient, including heart rate, blood pressure, blood oxygen saturation, and static information of the patient, including age, gender, medical history; S2: pre-processing the time-series medical health data to obtain time-series features: for the difference in sampling frequency, a cubic spline interpolation function is used for time alignment; for heart rate data, sliding window statistics are calculated, abnormal values are determined and removed, and linear interpolation formula is used to fill in the missing; for blood pressure data, a weighted moving average filter is used to smooth the noise; S3: extracting time-domain, frequency-domain and clinical medical features from the time-series features to obtain multi-dimensional features; S4: fusing the multi-dimensional features with the static information of the patient; wherein the time-series channel extracts dynamic features, and the static channel converts the age and gender structured data into a high-dimensional vector through an embedding layer; S5: constructing and training a deep learning model; the deep learning model uses a bidirectional long short-term memory network to process time-series channel data, and introduces residual connection between network layers, and finally splices and fuses the hidden state and the high-dimensional vector of the static channel as the input of the full connection layer; S6: using the deep learning model to predict the changes of the indicators in real time, and matching with the static features to evaluate the cardiovascular disease risk level.

2. The cardiovascular trend prediction method based on time-series medical health data according to claim 1, wherein, The dynamic physiological indicators collected in step S1 also include activity amount, body temperature and electrocardiogram data.

3. The cardiovascular trend prediction method based on time-series medical health data of claim 1, wherein, In step S2, the cubic spline interpolation function is specifically: Let the sequence of low frequency data points be where the time stamp, is the measured value; on the interval the spline function is a cubic polynomial: ； wherein denotes the function value at the start value of the interval, denotes the slope at the change rate, is the curvature at the change acceleration, is the cubic term coefficient, which controls the degree of bending of the curve.

4. The cardiovascular trend prediction method based on time-series medical health data of claim 2, wherein, The heart rate data is recorded once per second, including resting heart rate and instantaneous fluctuation during exercise; the blood pressure data is automatically measured every 5 minutes, recording systolic pressure, diastolic pressure and pulse frequency; the blood oxygen saturation is detected by a red light sensor every minute; the electrocardiogram data is continuously collected at a sampling rate of 250Hz, capturing the morphological changes of P wave and QRS complex.

5. The cardiovascular trend prediction method based on time-series medical health data as claimed in claim 1, wherein, The time-domain features include statistics of each indicator in the sliding window, including mean, standard deviation, skewness and zero-crossing rate; The frequency domain features include: the signal is decomposed into different frequency components by fast Fourier transform, 0.04-0.15Hz is low frequency component, 0.15-0.4Hz is high frequency component, and the low frequency and high frequency component power ratio is calculated.

6. The cardiovascular trend prediction method based on time-series medical health data of claim 5, wherein, The clinical medical features include: the pulse pressure difference of systolic pressure and diastolic pressure, morning peak blood pressure, and night blood pressure drop rate.

7. The cardiovascular trend prediction method based on time-series medical health data of claim 5, wherein, In step S4, when fusing the time-domain features and frequency-domain features with the static information, a weighted fusion strategy is used, wherein the weights of the time-domain features and frequency-domain features are dynamically adjusted according to their contribution to the cardiovascular disease risk, and the weight of the static information is modified according to the patient's medical history and family history.

8. The cardiovascular trend prediction method based on time-series medical health data of claim 1, wherein, In step S6, when evaluating the cardiovascular disease risk level, a preset medical threshold and a risk assessment matrix are combined, wherein the risk assessment matrix considers the patient's age, gender, medical history and real-time physiological indicators.

9. A cardiovascular trend prediction system based on time-series medical health data, characterized by, Comprising: a collection module for collecting time-series medical health data of a patient, including dynamic physiological indicators of the patient, including heart rate, blood pressure, blood oxygen saturation, and static information of the patient, including age, gender, medical history; The preprocessing module pre-processes the time-series medical health data to obtain time-series features: for the difference in sampling frequency, a cubic spline interpolation function is used for time alignment; for heart rate data, sliding window statistics are calculated, abnormal values are determined and removed, and a linear interpolation formula is used to fill in the missing values; for blood pressure data, a weighted moving average filter is used to smooth the noise; The extraction module extracts time-domain, frequency-domain and clinical medical features from the time-series features to obtain multi-dimensional features; The vector generation module fuses the multi-dimensional features with the static information of the patient; wherein the time-series channel extracts dynamic features, and the static channel converts the age and gender structured data into high-dimensional vectors through an embedding layer; The model construction module constructs and trains a deep learning model; the deep learning model uses a bidirectional long short-term memory network to process the time-series channel data, and introduces a residual connection between network layers, and finally concatenates and fuses the hidden state and the high-dimensional vector of the static channel as the input of the full connection layer; The prediction module uses the deep learning model to predict the changes of the index in real time, and matches with the static features to evaluate the risk level of cardiovascular diseases.

10. An electronic device, comprising: The method comprises the following steps: one or more processors; a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1 to 8.

Images