Fig. 1: End-to-end architecture for core body temperature (CBT) prediction, from raw data input to alert generation. | Communications Engineering

Fig. 1: End-to-end architecture for core body temperature (CBT) prediction, from raw data input to alert generation.

From: Degrees of uncertainty: conformal deep learning for non-invasive core body temperature prediction in extreme environments

Fig. 1: End-to-end architecture for core body temperature (CBT) prediction, from raw data input to alert generation.

Sequential features (e.g., physiological time-series) and non-sequential features (e.g., demographic and environmental data) are first preprocessed through scaling and imputation. These processed inputs feed into a hybrid model combining Long Short-Term Memory (LSTM) and Dense layers. The model’s outputs are refined by a conformal prediction layer, which provides statistically guaranteed prediction bounds at a user-defined confidence level (e.g., 95% probability that the actual CBT lies between CBT High and CBT Low). The alert layer then classifies predictions into user-defined safety states; for example, following Lousada et al.38: Nominal (36.5–37.75 °C, green), Advisory (36–36.5 °C or 37.75–38 °C, yellow), Caution (35–36 °C or 38–38.9 °C, orange), and Warning ( < 35 °C or > 38.9 °C, red). This modular architecture enables efficient real-time monitoring and confidence-driven alerts across diverse, safety-critical operational environments, and can be readily updated with new data or features.

Back to article page