Degrees of Uncertainty: Conformal Deep Learning for Core Body Temperature Prediction in Extreme Environments

We introduce a conformal deep learning framework for real-time, non-invasive core body temperature (CBT) prediction, designed for deployment in diverse, high-risk heat environments. The framework combines bidirectional Long Short-Term Memory (LSTM) networks with dense layers and leverages Stratification of Inductive Conformal Prediction Estimates (SCOPE) to deliver reliable uncertainty calibration and adaptability across varied demographics, workloads, protective gear, and environmental conditions. Calibrated over a CBT range of 36.15°C to 40.25°C, the model delivers a test error of 0.29°C, outperforming the widely used ECTemp™, with a 12-fold improvement in probabilistic accuracy and statistically guaranteed prediction intervals. Designed for seamless integration with wearable devices, the framework leverages accessible demographic, physiological, and environmental metrics for practical, non-invasive monitoring. SCOPE powers a customizable alert system, enabling proactive safety management through precise, real-time interventions. Adaptable to user-defined thresholds and confidence levels, the system is well-suited for applications in healthcare, industry, and other high-risk settings prone to heat stress. By combining accurate predictions with reliable uncertainty estimates, our framework sets a new standard for non-invasive, real-time CBT monitoring in safety-critical applications.