Underwater IoT body temperature-heart rate dual-mode terminal and low temperature stress risk model for winter swimming enthusiasts

To address the issues of missed detections, false alarms, and lack of personalization in existing low-temperature stress risk models due to their neglect of the dynamic coupling between body temperature and heart rate and their reliance on fixed thresholds, this paper constructs a dual-mode body temperature-heart rate monitoring and risk modeling framework that integrates end-to-cloud collaboration. In this framework, the end-to-end uses a rule engine combined with a quantized Tiny-GRU (Gated Recurrent Unit) to provide low-latency initial screening and safety triggering based on the individual's baseline before water entry, ensuring immediate response when communication or cloud access is unavailable. The cloud uses a TCN (Temporal Convolutional Network) for local temporal feature extraction and a Transformer to model long-term coupling relationships, while injecting individual embeddings to achieve personalized adaptation for small sample sizes. The decision layer employs uncertainty decomposition and inverse variance weighted fusion of end-to-cloud outputs, and includes highly uncertain samples in an active annotation queue to continuously optimize the model and support fine-tuning of class adaptation strategies to address differences in devices and scenarios. Experimental results on field validation and test data from winter-swimming sessions show that the edge-cloud collaborative model achieved a mean AUC of 0.94 with a fold-wise SD of 0.010 under subject-disjoint cross-validation, an F1 score of 0.91 with a fold-wise SD of 0.020, an average early warning lead time of 5.91 min, and a FAR of 0.035 times/hour. Personalization further improved risk-window sensitivity, probability calibration, and false-alarm control, with recall-related improvement interpreted at the risk-window level rather than as per-subject event-level Recall across all 60 participants. The overall framework demonstrates promising potential for interpretable personalized early warning in real-world winter swimming scenarios.