Cerebral stroke risk monitoring and identification method
Through the stroke risk identification method of multimodal physiological data acquisition and dual-branch hybrid model, the problems of data set imbalance and low computational efficiency are solved, efficient and accurate stroke risk monitoring and identification are achieved, and clinical interpretability and real-time early warning are provided.
Patent Information
- Application Number
- CN202511009038.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-22
- Publication Date
- 2025-10-17
AI Technical Summary
The imbalance of stroke datasets causes traditional classifiers to be biased towards the majority class, resulting in insufficient model generalization and low computational efficiency. Single machine learning models perform poorly on unbalanced data, and existing technologies are sensitive to unbalanced data and are prone to overfitting or ignoring noise samples.
Multimodal physiological data collection and fusion are used to identify stroke risk through a two-branch hybrid model, including a temporal processing branch and a non-temporal processing branch. 1D-CNN, BiLSTM, and XGBoost classifiers are combined, and sliding window segmentation and entity embedding technology are used for feature extraction and fusion. Edge computing is used to reduce cloud transmission delays.
It improves the accuracy and computational efficiency of stroke risk identification, overcomes the limitations of a single data source, provides clinical interpretability and real-time early warning capabilities, and reduces cloud transmission latency.
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of data processing, and particularly relates to a stroke risk monitoring and identification method. BACKGROUND
[0002] Data imbalance: the positive and negative sample ratio of the stroke dataset is greatly different (for example, the positive sample in the Kaggle dataset accounts for only 4.9%), which leads to the bias of the traditional classifier to the majority class.
[0003] Model generalization deficiency: the AUC value of a single machine learning model (such as logistic regression) on unbalanced data is usually less than 0.85.
[0004] Low computing efficiency: when the grid search hyperparameter optimization is performed, the traditional brute force search method is time-consuming.
[0005] For the logistic regression technology, the defect is sensitivity to unbalanced data. The defect of the naive Bayes + SMOTE technology is that noise samples are easily ignored; the defect of the unoptimized single XGBoost technology is high risk of overfitting.
[0006] Therefore, the application is proposed. SUMMARY
[0007] The application aims to provide a stroke risk monitoring and identification method to solve the above problems.
[0008] The stroke risk monitoring and identification method comprises the following steps:
[0009] Step 1, data acquisition and preprocessing
[0010] Multi-modal physiological data acquisition is adopted, and physiological signals, biochemical indicators and behavior data of the monitoring object are collected through wearable devices, medical IoT terminals and electronic health record systems
[0011] Step 2, data encoding and standardization
[0012] Time series data encoding is adopted, and the frequency domain features of the ECG / EEG signal are extracted through Fourier transform by sliding window segmentation;
[0013] Or,
[0014] Category data encoding is adopted, and the categorical variable of the history of hypertension is mapped to a 3-dimensional continuous vector by using entity embedding technology.
[0015] Step 3, feature extraction and fusion
[0016] The physiological signal features and behavior features are fused according to the hierarchical fusion architecture.
[0017] Step 4, construction and reasoning of stroke identification model
[0018] Adopt a double-branch hybrid model:
[0019] The time sequence processing branch: 1D-CNN cooperates with BiLSTM to process physiological signals;
[0020] The non-time sequence processing branch: a fully connected layer processes biochemical and behavioral characteristics;
[0021] The fusion layer: the outputs of the double branches are input to an XGBoost classifier after being concatenated;
[0022] Real-time inference process:
[0023] Input: data of the monitoring object on the day;
[0024] Feature extraction: real-time calculation of HRV SDNN and gait asymmetry;
[0025] Model prediction: if the output probability is greater than a threshold, an early warning is triggered to the doctor end.
[0026] Further improvement, the physiological signals include one or more of blood pressure, electrocardiogram, electroencephalogram and blood oxygen saturation.
[0027] Further improvement, the biochemical indicators include one or more of serum homocysteine and fasting blood glucose.
[0028] Further improvement, the behavioral data include one or more of gait stability and speech fluency.
[0029] Further improvement, the physiological signal features include one or more of blood pressure variability and heart rate variability.
[0030] Further improvement, the behavioral features include gait asymmetry.
[0031] Further improvement, the hierarchical fusion architecture includes:
[0032] 1. Low-level fusion: concatenation of features of the same type of data;
[0033] 2. Intermediate fusion: cross-modal features are weighted through an attention mechanism;
[0034] 3. High-level fusion: the intermediate fusion result is input to a graph neural network to construct a feature relationship graph.
[0035] Compared with the prior art, the present application has the following beneficial effects:
[0036] 1. Multi-modal fusion is adopted: the limitation of a single data source (such as relying only on blood pressure, which is easy to miss the diagnosis of asymptomatic patients) is overcome.
[0037] 2. Edge computing compatibility: feature extraction module can be deployed to mobile phone to reduce cloud transmission delay.
[0038] 3. Clinical interpretability: output key risk factors through SHAP values (example: 80% of a patient's risk comes from Hcy exceeding the standard). DETAILED DESCRIPTION
[0039] The application will be further described in detail below through specific examples.
[0040] Example 1
[0041] 1. Data collection and preprocessing
[0042] 1.1 Multimodal physiological data collection
[0043] The following data is collected by the monitoring object through wearable devices, medical IoT terminals and electronic health record (EHR) systems:
[0044] Physiological signals:
[0045] Blood pressure (dynamic 24-hour monitoring, sampling frequency 5 minutes / time);
[0046] Electrocardiogram (ECG, single lead, sampling rate 250 Hz);
[0047] Electroencephalogram (EEG, 8 channels, focusing on frontal and temporal regions);
[0048] Oxygen saturation (SpO2, collected by infrared photoelectric sensor);
[0049] Biochemical indicators:
[0050] Serum homocysteine (Hcy, detected by microfluidic chip);
[0051] Fasting blood glucose (daily morning fingertip blood measurement);
[0052] Behavioral data:
[0053] Gait stability (three-axis acceleration sensor, 100 Hz sampling);
[0054] Speech fluency (smartphone audio recording, NLP analysis of syllable interval);
[0055] 1.2 Data encoding and standardization
[0056] Time series data encoding:
[0057] The ECG / EEG signal is subjected to Fourier transform to extract frequency domain features by using sliding window segmentation (window length 60 seconds, overlap rate 30%);
[0058] Category data encoding:
[0059] Categorical variables such as hypertension history are mapped to 3-dimensional continuous vectors using Entity Embedding technique (better than One-Hot encoding, preserving semantic relevance).
[0060] 2. Feature extraction and fusion
[0061] 2.1 Core feature extraction algorithm
[0062] Physiological signal features:
[0063] Blood pressure variability (BPV): Calculate 24-hour systolic blood pressure standard deviation (SD) and coefficient of variation (CV).
[0064] Heart rate variability (HRV): Detect R-waves using Pan-Tompkins algorithm, extract SDNN (time domain) and LF / HF ratio (frequency domain).
[0065] Behavioral features:
[0066] Gait asymmetry: Left / right stride difference ratio (threshold > 15% is abnormal).
[0067] 2.2 Feature fusion strategy
[0068] Hierarchical fusion architecture:
[0069] 1) Low-level fusion: Concatenate features of the same type (e.g., merge ECG time domain + frequency domain features into a 10-dimensional vector).
[0070] 2) Mid-level fusion: Cross-modal features are weighted through an attention mechanism.
[0071] 3) High-level fusion: The mid-level fusion results are input into a graph neural network (GNN) to construct a feature relationship graph (nodes = features, edges = Pearson correlation coefficients).
[0072] 3. Stroke recognition model construction and inference
[0073] 3.1 Model architecture
[0074] Dual-branch hybrid model:
[0075] Branch 1 (time series processing): 1D-CNN (convolution kernel size 5, step 2) + BiLSTM (hidden layer 64 units) process physiological signals.
[0076] Branch 2 (non-time series processing): Fully connected layer (ReLU activation) processes biochemical and behavioral features.
[0077] Fusion layer: The outputs of the two branches are concatenated and input into an XGBoost classifier.
[0078] 3.2 Real-time inference procedure
[0079] Input: Daily data of monitoring subject (blood pressure + ECG + gait).
[0080] Feature extraction: Real-time calculation of HRV SDNN and gait asymmetry.
[0081] Model prediction: If the output probability > 0.85 (threshold determined by ROC curve), trigger an early warning to the doctor's end.
[0082] 4. Verification and performance optimization
[0083] 4.1 Cross-validation design
[0084] Dataset division:
[0085] Training set (70%): 2000 cases (980 cases of stroke positive samples, balanced after SMOTE-ENN).
[0086] Test set (30%): 858 cases (original unbalanced distribution is reserved to verify the generalization).
[0087] Evaluation indicators:
[0088] Main indicators: AUC-ROC, sensitivity (recall rate).
[0089] Secondary indicators: F1-score (key indicator for handling unbalanced data).
[0090] 4.2 The results of comparative experiments are shown in Table 1
[0091] AUC sensitivity specificity The invention (GNN+XGBoost) 0.99 96.2％ 98.1％ Single LSTM 0.91 82.4％ 89.3％ Traditional Logistic Regression 0.76 68.5％ 79.0％
[0092] 4.3 Ablation experiment
[0093] Feature fusion necessity: After removing the attention mechanism, the AUC decreased by 7.2%.
[0094] Sampling technique impact: SMOTE-ENN improved sensitivity by 21.5% compared to original data.
[0095] 5. Application scenario example
[0096] Community hospital screening procedure:
[0097] Residents wear smart bracelets for 7 days to collect basic physiological data.
[0098] The system automatically generates a risk score (0-100 points), and ≥85 points are recommended for further CT examination.
[0099] High-risk patients are included in a follow-up list, and personalized health recommendations (such as a reminder of antihypertensive drugs) are pushed every week.
[0100] The above merely provides the preferred embodiment of the present application, but not for limiting the present application. For the person skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.
Claims
1. A method for monitoring and identifying stroke risk, characterized by: The following steps are involved: Step 1: Data collection and preprocessing Using multimodal physiological data collection, the monitored subjects collect physiological signals, biochemical indicators, and behavioral data through wearable devices, medical IoT terminals, and electronic health record systems Step 2: Data coding and standardization Using time series data encoding and sliding window segmentation, the ECG / EEG signal is subjected to Fourier transform to extract frequency domain features; or, Categorical data encoding was adopted, and the categorical variable of hypertension history used entity embedding technology to map discrete values into 3-dimensional continuous vectors; Step 3: Feature extraction and fusion The physiological signal features and behavioral features are integrated according to the hierarchical fusion architecture; Step 4: Stroke recognition model construction and reasoning Using a two-branch mixture model: Time series processing branch: 1D-CNN combined with BiLSTM to process physiological signals; Non-temporal processing branch: fully connected layers process biochemical and behavioral features; Fusion layer: The dual-branch output is concatenated and then input into the XGBoost classifier; Real-time inference process: Input: daily data of the monitored object; Feature extraction: Real-time calculation of HRV SDNN and gait asymmetry; Model prediction: If the output probability is greater than the threshold, an early warning is triggered to the doctor.
2. The stroke risk monitoring and identification method according to claim 1, characterized in that: The physiological signal includes one or more of blood pressure, electrocardiogram, electroencephalogram, and blood oxygen saturation.
3. The stroke risk monitoring and identification method according to claim 1, characterized in that: The biochemical indicators include one or more of serum homocysteine and fasting blood glucose.
4. The stroke risk monitoring and identification method according to claim 1, wherein: The behavioral data includes one or more of gait stability and speech fluency.
5. The stroke risk monitoring and identification method according to claim 1, characterized in that: The physiological signal characteristics include one or more of blood pressure variability and heart rate variability.
6. The stroke risk monitoring and identification method according to claim 1, characterized in that: The behavioral characteristics include gait asymmetry.
7. The stroke risk monitoring and identification method according to claim 1, characterized in that: The hierarchical fusion architecture includes:
1. Low-level fusion: splicing of similar data features; 2. Intermediate fusion: cross-modal features are weighted through the attention mechanism; 3. Advanced fusion: Input the intermediate fusion results into the graph neural network to construct a feature relationship graph.