EDGE AI–ENABLED INTELLIGENT HEALTHCARE FRAMEWORK USING LIGHTWEIGHT DEEP LEARNING, MULTIMODAL SENSOR FUSION, AND REAL-TIME ON-DEVICE ANALYTICS FOR CONTINUOUS PATIENT MONITORING AND EARLY HEALTH RISK DETECTION

Abstract

The rapid growth of chronic disease burden and the shortage of clinical personnel have intensified the demand for continuous, low-latency patient monitoring that does not depend on constant cloud connectivity. This paper presents an edge artificial intelligence-enabled intelligent healthcare framework that integrates lightweight deep learning, multimodal sensor fusion, and real-time on-device analytics for continuous patient monitoring and early health risk detection. The proposed architecture acquires electrocardiogram, photoplethysmography, respiration, skin temperature, blood oxygen saturation, and inertial motion signals from wearable and bedside sensors, and combines them through an attention-based fusion module operating on time-aligned feature embeddings. A compact convolutional-recurrent network, optimized through structured pruning, knowledge distillation, and 8-bit post-training quantization, performs inference directly on resource-constrained edge hardware, while a federated learning layer enables privacy-preserving model updates without transferring raw physiological data. An adaptive risk-scoring engine converts predictions into graded alerts for clinicians and caregivers, addressing the latency, bandwidth, energy, and privacy limitations of conventional cloud-centric monitoring systems.The framework was evaluated on benchmark physiological datasets, including MIMIC-III waveform records and the PhysioNet challenge collections, together with a prospective pilot involving 120 patients. The fused model achieved an accuracy of 97.3%, a sensitivity of 96.8%, a specificity of 97.6%, and an F1-score of 96.9%, outperforming single-modality baselines by 5.2 to 9.4 percentage points. Compression reduced the model size by 78% to 0.84 megabytes and lowered inference latency to 11.6 milliseconds on a Raspberry Pi 4 class device, with energy consumption per inference decreasing by 73% compared with the uncompressed network. Early deterioration events were identified on average 37.7 minutes before clinician-identified onset, compared with less than 5 minutes for conventional threshold-based alarms, and the false alarm rate declined from 64% to 26%. Performance remained consistent across demographic subgroups and varying noise and signal-quality conditions. Federated training retained accuracy within 1.1% of centralized learning while preserving data locality. These results demonstrate that edge-native multimodal analytics can deliver clinically reliable, energy-efficient, and privacy-aware monitoring across diverse patient populations suitable for hospital wards, remote care, and home-based settings.