Chronic disease dynamic monitoring and intervention system based on multi-modal data fusion and deep learning

Abstract

The invention discloses a chronic disease dynamic monitoring and intervention system based on multi-modal data fusion and deep learning, and the system integrates a wearable sensing technology, a multi-modal data fusion algorithm, a deep learning model (DeepSeek) and a dynamic intervention strategy, and achieves the real-time monitoring, risk prediction and personalized management of chronic diseases. The system has the following functions: performing filing by inputting user information, performing reasoning processing by using a deep model after data acquisition, outputting short-term prediction and risk assessment results, and performing a dynamic personalized intervention scheme (including emergency event processing, drug dose optimization and behavior intervention strategies).

Description

Technical Field

[0001] The present invention belongs to the cross - technical field of intelligent healthcare and artificial intelligence, and particularly relates to a full - cycle intelligent management system for chronic diseases (such as diabetes, hypertension, chronic respiratory diseases). Background Art

[0002] Chronic diseases (such as diabetes, hypertension, chronic respiratory diseases, etc.) have become a major challenge in the global public health field. The traditional chronic disease management model mainly relies on regular outpatient follow - up and patient self - reporting, and there are problems such as fragmented data collection, insufficient real - time performance, and lagged intervention. In recent years, with the development of wearable devices, mobile healthcare, and Internet of Things technologies, the acquisition of multi - modal health data (such as physiological indicators, behavioral data, environmental information, electronic medical records, etc.) has become possible. However, there are some problems in existing chronic disease management technologies: single - modal data (such as single - point measurement of blood glucose or blood pressure) cannot reflect the overall physiological state of patients; traditional prediction models (such as logistic regression) have insufficient ability to capture complex non - linear relationships; there is a lack of real - time dynamic intervention mechanisms, and patient compliance is low.

[0003] In recent years, multi - modal data fusion and deep reinforcement learning technologies have provided new ideas for chronic disease management.

[0004] The present invention belongs to the cross - technical field of intelligent healthcare and artificial intelligence, and particularly relates to a full - cycle intelligent management system for chronic diseases (such as diabetes, hypertension, chronic respiratory diseases). This system integrates wearable sensing technology, multi - modal data fusion algorithms, deep learning models (DeepSeek), and dynamic intervention strategies to achieve real - time monitoring, risk prediction, and personalized management of chronic diseases. The present invention is particularly applicable to the family / community chronic disease management scenario, and solves the three core problems of data fragmentation, insufficient generalization ability of prediction models, and lagged intervention in the existing technology through technological innovation, providing a standardized and scalable technical solution for chronic disease prevention and control. Summary of the Invention

[0005] In view of the problems existing in the existing technical solutions, it is desired to provide a chronic disease dynamic monitoring and intervention system based on multi - modal data fusion and deep learning.

[0006] This system integrates wearable sensing technology, multi - modal data fusion algorithms, deep learning models (DeepSeek), and dynamic intervention strategies to achieve real - time monitoring, risk prediction, and personalized management of chronic diseases.

[0007] At the technical implementation level, the present invention covers the following core directions:

[0008] Biomedical Sensing Technology: Through high-precision PPG (Photoplethysmography), ECG (Electrocardiogram), and bioimpedance sensor arrays, continuous non-invasive monitoring of key physiological parameters such as blood glucose, blood pressure, and blood oxygen saturation is achieved, breaking through the limitations of traditional single-point measurements.

[0009] Edge Computing and Cloud Computing Collaborative Architecture: Based on an improved Transformer-UNet hybrid neural network, a two-layer inference framework of a lightweight model at the edge side (parameter count < 1MB) and a high-precision model in the cloud is constructed, reducing device power consumption while ensuring prediction accuracy.

[0010] Multimodal Time-Series Data Analysis: For multi-source heterogeneous data such as physiological signals (sampling rate 100 - 500Hz), environmental data (temperature, humidity, air quality), and user behavior (diet, exercise), a spatio-temporal attention mechanism and a Kalman filter fusion algorithm are designed to improve the efficiency of feature extraction.

[0011] Closed-Loop Intervention Mechanism: Combining reinforcement learning and a medical knowledge graph, personalized intervention plans (such as drug dose adjustment, exercise recommendations) are dynamically generated, and a closed-loop control of warning - feedback - evaluation is achieved through a low-latency communication protocol (BLE 5.0 + LoRa).

[0012] This system adopts a three-level hierarchical architecture (data acquisition → intelligent analysis → dynamic intervention). The core innovation lies in the efficient fusion of multimodal data and a real-time decision-making mechanism based on the DeepSeek model. The specific implementation is as follows:

[0013] In the multimodal data acquisition layer:

[0014] The optical sensing module of the wearable sensing unit uses a dual-wavelength LED (660nm + 940nm) and a photodiode array.

[0015] Achieved Metrics: Blood Glucose Trend Monitoring: Analyze the change in vascular elasticity through the second derivative of PPG (error ±0.8mmol / L vs. fingertip blood).

[0016] Blood Pressure Estimation: Based on the pulse wave transit time (PWTT) and a deep learning correction model (MAP error ±5mmHg).

[0017] Electrophysiological Module: 4-electrode bioimpedance measurement (frequency 50kHz - 100kHz) detects changes in body fluid distribution (accuracy 0.1Ω) for early warning of pulmonary edema in heart failure patients.

[0018] In the DeepSeek Model Analysis Layer:

[0019] Data Preprocessing: Dynamic time warping (DTW) algorithm is used for time series alignment to solve the sampling rate differences of multiple sensors.

[0020] Noise suppression: Combining wavelet transform (db4 basis function) with adaptive Kalman filtering.

[0021] Hybrid neural network architecture: Spatiotemporal feature extraction branch: 1D-CNN layer (kernel_size=7, stride=2) extracts local physiological features, and the Transformer encoder (4 layers, 4-head attention) captures long-range dependencies.

[0022] Multimodal fusion branch: Design the weight assignment formula for the cross-modal attention gate mechanism: α = σ(W_g · [h_sensor; h_env; h_behavior] + b_g), where σ is the Sigmoid function and W_g is the trainable parameter matrix.

[0023] Joint optimization objective: The main loss function is the time-weighted mean square error (TW-MSE).

[0024] Auxiliary task: Physiological signal reconstruction (enhancing feature interpretability through the decoder).

[0025] In the dynamic intervention module:

[0026] Hierarchical warning strategy:

[0027]

[0028] Personalized plan generation: Dose optimization based on reinforcement learning, state space: recent physiological indicators of the patient + medication records.

[0029] Reward function: R = w1*(indicator stability) - w2*(side effect probability).

[0030] Policy network: Twin Delayed DDPG algorithm.

[0031] Behavior intervention simulation: Using conditional generative adversarial network (cGAN) to simulate the effects of different exercise / diet plans, generator input: patient baseline data + intervention type (such as "walk fast for 30 minutes every day").

[0032] Discriminator: 3-layer CNN, judging the distribution consistency between the generated data and real clinical cases.

[0033] In the edge-cloud collaborative computing:

[0034] Edge device side: Runs a lightweight model: Compresses the original model by 1 / 50 through knowledge distillation technology.

[0035] Technical parameters: Memory occupancy: <512KB.

[0036] Inference speed: <15 ms / time (based on Cortex-M7 processor).

[0037] Cloud training platform: Federated learning framework: Allows hospitals to share model updates instead of raw data.

[0038] Differential privacy protection: Adds Laplace noise (ε = 0.5) to protect patient privacy.

[0039] Comparison of technical advantages:

[0040]

[0041] Prediction performance (taking diabetes as an example): Blood glucose prediction MAE (Mean Absolute Error): 0.6 mmol / L (1.2 mmol / L for traditional LSTM model).

[0042] AUC-ROC for hypoglycemia warning: 0.93 (compared with 0.85 of XGBoost).

[0043] Hardware efficiency: Inference power consumption at the edge side: ≤15 mW (using model quantization + hardware accelerator).

[0044] Data transmission compression ratio: 12% of the raw data (based on improved Huffman coding). Description of the drawings

[0045] Figure 1 It is a schematic flow chart of a chronic disease dynamic monitoring and intervention system based on multi-modal data fusion and deep learning described in this application. Specific implementation manners

[0046] The following further elaborates on this application in conjunction with the drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the related invention and are not intended to limit the invention.

[0047] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments can be combined with each other. The following will elaborate on this application in detail with reference to the drawings and embodiments.

[0048] A chronic disease dynamic monitoring and intervention system based on multi-modal data fusion and deep learning includes a system deployment process, a data process and processing, an intervention execution specification, and an embodiment description.

[0049] System deployment process:

[0050] Hardware installation pre-configuration:

[0051] Initialization of wearable devices:

[0052] Step 1: Activate the device by NFC touch and automatically pair with the home gateway (BLE MAC address binding).

[0053] Step 2: Calibrate the sensors:

[0054] Optical module: The user sits still for 5 minutes to collect the reference PPG signal (sampling rate 100Hz).

[0055] ECG module: Dry electrode impedance detection (requirement <200kΩ, otherwise prompt to adjust the wearing position).

[0056] Deployment of environmental sensors:

[0057] Positioning: Install 1 environmental monitoring node in each of the bedroom and living room (height 1.2 - 1.5 meters, away from direct air blowing of the air conditioner).

[0058] Networking: Self - organize the network through ZigBee 3.0 (network ID encryption, AES - 128 key).

[0059] Software configuration:

[0060] Establishment of user profiles:

[0061] Input: Age, gender, medical history (diabetes / hypertension, etc.), medication regimen (drug name + dose + time).

[0062] Generate: Personalized monitoring parameter thresholds (such as automatically setting the blood glucose warning line according to the ADA standard).

[0063] Model initialization:

[0064] Load the basic model: A pre - trained DeepSeek model based on 100,000 cases of general chronic disease data.

[0065] Incremental training: The data collected in the first 3 days is used for transfer learning (learning rate 1e - 4, batch size 32).

[0066] Data flow and processing:

[0067] Real - time monitoring stage and data collection specifications:

[0068]

[0069] Motion artifact removal in edge pre - processing:

[0070] def motion_artifact_removal(signal, accel_data):

[0071] # Build a regression model using accelerometer data.

[0072] X = wavelet_transform(accel_data);

[0073] clean_signal = signal - ridge_regression(X, signal);

[0074] return butterworth_filter(clean_signal, 0.5 - 40Hz);

[0075] Data compression: Improved Delta encoding (compression ratio 5:1, fidelity > 99%).

[0076] Cloud analysis phase:

[0077] Deep model inference:

[0078] Input feature dimension:

[0079] Time series data: 60 - minute window, including 20 - dimensional features such as heart rate, HRV, SpO2 trend, etc.

[0080] Environmental data: 6 - dimensional features such as temperature, humidity, PM2.5, etc.

[0081] Behavioral data: 8 - dimensional features such as medication time, exercise amount, etc.

[0082] Short - term prediction: Blood glucose change curve in the next 2 hours (at 5 - minute intervals).

[0083] Risk assessment: Probability of cardiovascular and cerebrovascular events (graded 0 - 1).

[0084] Model update mechanism:

[0085] Weekly incremental training: Federated learning framework aggregates data from 10 medical nodes (differential privacy ε = 0.7).

[0086] Anomaly detection: Trigger model recalibration when the distribution shift of user data exceeds the KL divergence threshold of 0.3.

[0087] Intervention execution specification:

[0088] Early warning trigger logic:

[0089] Multi - condition composite judgment IF (blood glucose > 13.9mmol / L for 3 consecutive times) AND (recent insulin injection amount < preset value 80%), THEN trigger level II early warning (push insulin adjustment advice).

[0090] Emergency event handling:

[0091] Step 1: The local device vibrates + the LED flashes red (frequency 2 Hz).

[0092] Step 2: Send an alarm message (including GPS location) to the family member APP via the MQTT protocol.

[0093] Step 3: If there is no response within 10 minutes, automatically call the contracted medical institution (connected to the IVR system).

[0094] Personalized intervention plan:

[0095] Drug dose optimization:

[0096] Input: Current blood glucose value + diet record + exercise amount.

[0097] Algorithm: Basal dose = patient's historical average dose, adjustment amount = Kp×Δblood glucose + Ki×∫Δblood glucose dt + Kd×d(Δblood glucose) / dt (PID coefficients are dynamically adjusted by reinforcement learning), Output: Insulin recommended dose (accurate to 0.5 unit).

[0098] Behavioral intervention strategy:

[0099] Exercise advice: Generate a personalized exercise plan (duration / intensity / frequency) based on GAN.

[0100] Diet planning: Image recognition of meal photos → Estimation of carbohydrate content → Push adjustment suggestions.

[0101] Exemplification:

[0102] Example 1: Diabetes management.

[0103] Morning high blood glucose treatment:

[0104] At 06:00, the system detects that the blood glucose continuously > 7.8 mmol / L (fasting).

[0105] Analyze the previous night's diet record (find that 15 g of carbohydrates were ingested during the 21:00 snack).

[0106] Suggested adjustment:

[0107] Immediate: Administer 2 U of rapid-acting insulin.

[0108] Long-term: Change the night snack to a protein food (push a recipe link).

[0109] Example 2: Hypertension management.

[0110] The systolic blood pressure > 135 mmHg continuously for 3 days from 07:00 to 09:00.

[0111] Associated environmental data: It is found that the blood pressure rises more significantly when the room temperature < 18 degrees Celsius.

[0112] Intervention measures:

[0113] Start the bedroom air conditioner 30 minutes in advance to heat up to 22 degrees Celsius.

[0114] Adjust the medication time: change the taking time of amlodipine from 08:00 to 06:00.

Claims

1. A chronic disease dynamic monitoring and intervention system based on multi-modal data fusion and deep learning, characterized in that, It includes data collection, intelligent analysis, and dynamic intervention. The core innovation lies in the efficient fusion of multi-modal data and the real-time decision-making mechanism based on the DeepSeek model. The specific implementation is as follows: At the technical implementation level, the system includes the following core directions: Biomedical sensing technology: Through high-precision PPG (photoplethysmography), ECG (electrocardiogram), and bioimpedance sensor arrays, continuous non-invasive monitoring of key physiological parameters such as blood glucose, blood pressure, and blood oxygen saturation is achieved, breaking through the limitations of traditional single-point measurement; Edge computing and cloud computing collaborative architecture: Based on an improved Transformer-UNet hybrid neural network, a two-layer inference framework of a lightweight model at the edge side (parameter count < 1MB) and a high-precision model in the cloud is constructed, reducing device power consumption while ensuring prediction accuracy; Multi-modal time-series data analysis: For multi-source heterogeneous data such as physiological signals (sampling rate 100 - 500Hz), environmental data (temperature, humidity, air quality), and user behavior (diet, exercise), a spatio-temporal attention mechanism and a Kalman filter fusion algorithm are designed to improve the efficiency of feature extraction; Closed-loop intervention mechanism: Combining reinforcement learning and a medical knowledge graph, personalized intervention plans (such as drug dose adjustment, exercise recommendations) are dynamically generated, and a closed-loop control of warning - feedback - evaluation is achieved through a low-latency communication protocol (BLE 5.0 + LoRa).

2. A chronic disease dynamic monitoring and intervention system based on multi-modal data fusion and deep learning according to claim 1, characterized in that: Optical and electrophysiological combined monitoring: Dual-wavelength PPG (660 / 940nm) and four-electrode bioimpedance are synchronously measured to achieve non-invasive continuous monitoring of blood glucose / blood pressure (error < clinical standard); Environmental compensation algorithm: Temperature and humidity data dynamically correct sensor drift, with a 40% improvement in accuracy; Spatio-temporal feature decoupling: 1D-CNN extracts local features + Transformer captures long-range dependencies Cross-modal attention gating: Dynamically allocate weights for physiological / environmental / behavioral data; Three-level response mechanism: From vibration warning to automatic linkage of medical devices (delay < 3 seconds); Reinforcement learning dose optimization: Personalized medication strategies based on patient digital twins; Edge lightweighting: The model is compressed to 0.8MB (accuracy loss < 2%); Federated learning privacy protection: Differential privacy (ε = 0.7) jointly trains multi-center data; Event-driven sampling: The ECG module is only activated during R-wave detection (power consumption reduced by 65%); Dual-mode communication: BLE 5.2 + LoRa adaptively switches, with a battery life > 7 days.

3. A chronic disease dynamic monitoring and intervention system based on multi-modal data fusion and deep learning, characterized in that: It is completed using the system described above in the claims and includes the following steps: This system integrates wearable sensing technology, multimodal data fusion algorithms, deep learning models (DeepSeek), and dynamic intervention strategies to achieve real-time monitoring, risk prediction, and personalized management of chronic diseases. By inputting user information to establish a file, after data collection, the deep model is used for inference processing, and short-term prediction and risk assessment results are output, and dynamic personalized intervention plans (including emergency event handling, drug dose optimization, and behavior intervention strategies) are carried out.

Images