Constructing a Multi-modal Intelligence Audit System Based on Biosensors and Deep Fusion

This study presents a multimodal intelligence audit system based on biosensors and deep fusion for safety, quality, and compliance assurance in biomedical systems. The system aligns heterogeneous evidence from physiological biosensor streams, electronic health records, medical-device logs, clinical text, images, audio, and procedural video, and then applies an attention-enhanced long short-term memory (LSTM) network for temporal anomaly detection. A constrained agentic layer converts model evidence into traceable audit tasks, retrieves supporting records, checks predefined clinical-engineering rules, and drafts reports subject to human approval. On the available enterprise audit dataset (12,600 samples), the implementation reduced total audit time by approximately 80% and achieved 96.0%-99.0% task-level accuracy. The biomedical extension is presented as a method-transfer framework rather than as a claim of completed clinical validation: patient-level evaluation, calibration, and prospective clinical testing remain required. The work therefore contributes a reproducible alignment interface between multimodal sensing, temporal fusion, and human-governed audit agents for biomedical engineering and clinical operations.