LSTM-based safety-oriented prediction of big toe skin temperature in extreme cold conditions for mountaineering footwear evaluation.

This study proposes a safety-oriented framework for the assessment of mountaineering footwear insulation in extreme cold environments. A novel metric, the Duration of Safe Exposure (DSE), is introduced and defined as the time required for big toe skin temperature to reach the conservative critical threshold of 15 °C under a given combination of footwear insulation, ambient temperature, and physical activity. Four structured experimental campaigns were conducted in a controlled climatic chamber to investigate the interaction between footwear thermal resistance, ambient temperature, and physical activity on peripheral thermoregulation. Big toe temperature was selected as a benchmark indicator due to its pronounced vasoconstrictive response during cold exposure. In total, the dataset comprised 96 time series collected across 10 experimental protocols, each corresponding to subject-specific tests under defined conditions. An LSTM-based neural network was trained to predict the temporal evolution of big toe temperature from mean skin temperature and the three operational parameters. The evaluation framework was explicitly structured to reflect the safety-oriented objective of the study and was fully consistent with the custom loss function adopted during training, which jointly penalises trajectory deviation (RMSE) and errors in the prediction of the threshold-crossing event. Beyond regression-based performance indicators, the model was evaluated through a binary classification of DSE (Duration of Safe Exposure) occurrence, defined as the network's ability to correctly predict whether and when the big toe temperature reaches the critical 15 °C threshold. Based on true and false positive/negative outcomes, classification metrics including Accuracy, Precision, Recall, and F1-score were computed to quantify the reliability of threshold detection within a predefined [Formula: see text]10-minute tolerance window. To further address safety implications, a dedicated safety-specific indicator, the Unsafe Error Rate (UER), was introduced to explicitly identify non-conservative DSE overestimations, i.e., cases in which the model predicts a longer safe exposure time than physiologically observed. Results indicate that unsafe predictions were rare, with most estimates being either within tolerance or conservative. Finally, the trained network was integrated with the JOS-3 thermoregulation model to enable exploration of physiologically consistent virtual scenarios. This hybrid framework provides a quantitative, physiologically grounded approach that may support the investigation of safety-related aspects of mountaineering footwear under controlled conditions. However, its applicability remains limited to the specific experimental setup and population considered, since the experimental data were collected exclusively from young male participants, and cannot be generalised to other populations without further validation.