The AIoT-HealthSense CNN-GRU model achieved 96.3% accuracy, 0.94 F1-score, and 0.97 ROC-AUC, outperforming baseline models in real-time physiological signal classification with reduced latency by 28%.
The proposed AIoT-HealthSense framework provides a highly accurate, low-latency, and privacy-preserving solution for real-time health monitoring and telemedicine diagnostics.
Effect estimate: Accuracy 96.3% vs baseline models
Modern healthcare systems (particularly those with limited resources and in remote locations) require real-time monitoring and telemedicine. Conventional cloud-reliant solutions face challenges with latency, data privacy, and trust. The system integrates a CNN-GRU hybrid deep learning model for classifying physiological signals, a Trust-Aware Diagnostic Engine based on EMA scoring, and a Federated Learning system augmented with Differential Privacy (DP) to ensure decentralized training. The system was extensively tested on PhysioNet MIT-BIH, MIMIC-III, and a simulated H-IoT dataset to assess real-time performance, accuracy, and efficiency. The proposed CNN-GRU model outperformed the baseline models (LSTM, Transformer-only) in average accuracy (96.3), F1-score (0.94), and ROC-AUC (0.97). In the hybrid edge-fog, latency was reduced by 28%. When device signal noise was intermittent, the trust score remained stable at over 0.85. Under DP noise (epsilon = 0.1), the federated learning model converged in 40 rounds with a negligible loss of accuracy. SHAP feature attribution made the predictions interpretable. The framework presents a new combination of federated learning, privacy-preserving and trust-based decision-making, and adaptive edge-fog orchestration for smart healthcare. This work provides a standard for the design of clinically relevant, safe, and interpretable systems based on AI-powered telemedicine and strongly aligns with translational impact.
Naik et al. (Thu,) conducted a other in Adults using wearable Raspberry Pi-based health monitoring devices capturing physiological signals including ECG, SpO2, temperature, and motion for real-time anomaly detection (n=20). AIoT-HealthSense system featuring CNN-GRU hybrid model with transformer-based multimodal fusion, trust-aware diagnostic engine, federated learning with differential privacy, and adaptive edge-fog orchestration vs. Baseline models including LSTM and Transformer-only models was evaluated on Diagnostic accuracy of real-time anomaly detection using physiological signal classification (Accuracy 96.3% vs baseline models). The AIoT-HealthSense CNN-GRU model achieved 96.3% accuracy, 0.94 F1-score, and 0.97 ROC-AUC, outperforming baseline models in real-time physiological signal classification with reduced latency by 28%.