An investigation of the suitability of Artificial Neural Networks for the prediction of core and local skin temperatures when trained with a large and gender-balanced database

Neural networks have been proven to successfully predict the results of complex non-linear problems in a variety of research fields, including medical research. Yet there is paucity of models utilising intelligent systems in the field of thermoregulation. They are under-utilized for predicting seemingly random physiological responses and in particular never used to predict local skin temperatures; or core temperature with a large dataset. In fact, most predictive models in this field (non-artificial intelligence based) focused on predicting body temperature and average skin temperature using relatively small gender-unbalanced databases or data from thermal dummies due to a lack of larger datasets.This paper aimed to address these limitations by applying Artificial Intelligence to create predictive models of core body temperature and local skin temperature (specifically at forehead, chest, upper arms, abdomen, knees and calves) while using a large and gender-balanced experimental database collected in office-type situations.A range of Neural Networks were developed for each local temperature, with topologies of 1–2 hidden layers and up to 20 neurons per layer, using Bayesian and the Levemberg-Marquardt back-propagation algorithms, and using various sets of input parameters (2520 NNs for each of the local skin temperatures and 1760 for the core temperature, i.e. a total of 19400 NNs). All topologies and configurations were assessed and the most suited recommended. The recommended Neural Networks trained well, with no sign of over-fitting, and with good performance when predicting unseen data. The recommended Neural Network for each case was compared with previously reported multi-linear models. Core temperature was avoided as a parameter for local skin temperatures as it is impractical for non-contact monitoring systems and does not significantly improve the precision despite it is the most stable parameter. The recommended NNs substantially improve the predictions in comparison to previous approaches. NN for core temperature has an R-value of 0.87 (81% increase), and a precision of ±0.46 °C for an 80% CI which is acceptable for non-clinical applications. NNs for local skin temperatures had R-values of 0.85-0.93 for forehead, chest, abdomen, calves, knees and hands, last two being the strongest (increase of 72% for abdomen, 63% for chest, and 32% for calves and forehead). The precision was best for forehead, chest and calves, with about ±1.2 °C, which is similar to the precision of existent average skin temperature models even though the average value is more stable.