A lightweight hybrid LSTM–GNN framework with context-aware risk calibration for physiological monitoring

Abstract The Internet of Medical Things requires physiological monitoring systems that balance temporal signal dynamics, relational structure, and computational efficiency. This study presents a lightweight hybrid LSTM–GNN framework for binary ECG screening, integrating explicit graph message passing, temporal modelling, and an ECG-derived proxy-context node. The framework was evaluated on 47 MIT-BIH records using strict record-wise splitting, preprocessing designed to avoid information leakage, and a fixed 0.50 decision threshold across five random seeds. In addition, an external cross-database evaluation was conducted on the INCART database without fine-tuning or threshold tuning, using a harmonized binary label mapping. On MIT-BIH, the Full Hybrid condition achieved the highest mean Accuracy (0.8991 ± 0.0363) and F1-score (0.8121 ± 0.0573), together with the lowest false-alarm rate (0.1129 ± 0.0523) among the evaluated conditions. On the external INCART database, the Full Hybrid condition achieved a mean Accuracy of 0.8801 ± 0.0151 and ROC-AUC of 0.9065 ± 0.0227 under the fixed external-test protocol. The ablation analysis indicated measurable contributions from the combined representations, while the context-permutation analysis showed that the proxy context should not be interpreted as an independent modality. In addition, software-level runtime profiling using ONNX Runtime demonstrated software-level portability characteristics and runtime efficiency on a standard CPU environment, with 8066 parameters, an approximately 36.1-kB model file, and a median CPU latency of 0.076 ms. The reported variability summarizes performance variation across five random seeds reflecting training stability, and should not be interpreted as patient-level inferential confidence intervals. These findings support reproducible, lightweight cross-database evaluation under the specified harmonization protocol, while limiting the claims to software-level computational feasibility rather than physical edge deployment, independent multimodal sensing, autonomous adaptive behavior, or established clinical utility.