AI Monitoring of Thrombosis Risk on Wearable Devices: A Clinical-Grade Breakthrough from Traditional Machine Learning to a Spatial-Relation-Temporal Hybrid Architecture

This project develops a clinical-grade AI system that continuously monitors thrombosis risk using wearable sensors. It employs a CNN-Transformer-BiLSTM hybrid architecture and Bayesian uncertainty quantification technology, achieving a 100% emergency recall rate and a 63.52% deterministic accuracy. It addresses core challenges in clinical AI such as class imbalance, while balancing the feasibility of edge deployment and prioritizing patient safety.

Deep Vein Thrombosis (DVT) and Pulmonary Embolism (PE) are among the leading causes of preventable deaths globally; approximately 100,000 people die from thrombosis-related diseases in the U.S. each year. Traditional detection relies on hospital-based ultrasound or D-dimer tests, missing opportunities for early intervention. Wearable devices can collect multi-modal signals such as heart rate and blood oxygen, but face machine learning challenges like noisy data and class imbalance in thrombosis events.

The project aims to solve the classic clinical AI problem of the 'accuracy paradox' (high overall accuracy under class imbalance but missed detection of rare emergency events). It evolved from traditional machine learning (Phase1) to a spatial-relation-temporal hybrid architecture (version v6), achieving stable generalization across patient groups.

Data Flow: Three-stage preprocessing (0.5Hz Butterworth high-pass filtering, 4-level db4 wavelet denoising, 30-second windowing enhancement) + subject-level stratification to avoid data leakage; Core Model: 1D-CNN extracts local morphological features, Transformer models cross-sensor relationships, Bi-LSTM captures long-term temporal dependencies, and the outputs of the three encoders are fused for decision-making.

• Bayesian Inference: Uses Monte Carlo Dropout to provide prediction confidence; high-uncertainty samples are marked for manual review. - Class Imbalance: Weighted focal loss (20x penalty for emergency class) + WGAN-GP to synthesize minority class samples. - Clinical Gating: Sets emergency trigger (≥3%) and high-risk trigger (≥10%) thresholds to ensure minimal missed detection rate.

Clinical Performance: 100% emergency recall rate on the test set, 63.52% overall deterministic accuracy; Edge Optimization: After INT8 quantization, the model is 4.2MB with 541k parameters, and the inference latency per 30-second window is 12.4ms, supporting deployment on Cortex-M4/M7 microcontrollers.

This project combines cutting-edge deep learning technology with clinical safety requirements; all technical choices serve the core goal of 'no missed emergency detection'. It provides technical references for medical AI developers, emphasizing the balance between performance metrics and patient safety, with patient safety always as the top priority.