Optimal Sensor Configuration for Body Temperature Estimation in Smart Clothing Using Machine Learning Models

Abstract

: Modern wearable technologies enable the collection of physiological and environmental data, which represents a significant potential for application in sports monitoring, analytics, performance modeling and even injury prevention. Although sensors allow for direct measurement, their values often deviate due to the influence of external conditions and sensor position. The unique materials, sensor technology, ethical issues, the benefits and drawbacks of using smart clothing in sports and medicine have all been discussed in the presented research, with particular emphasis on the fact that the analysis of data obtained from multiple sensors is challenging due to the existence of nonlinearity, variability and the presence of various noise. In order to improve the accuracy of the estimation, RF, SVM and FFNN have been applied in this paper. Machine learning models have been developed with different configurations of input variables, including individual temperature sensors, as well as their combinations and the performance of all models was evaluated using RMSE and R 2 . The assessment was carried out in a regression framework, while the results were additionally classified into three classes based on medically defined thresholds. The obtained results confirm that a high estimation accuracy can be achieved with a reduced number of sensors, which enables the simplification of the system and the reduction of implementation costs. Also, it is shown that the conclusions are consistent across different models, which indicates the robustness of the approach. The paper discusses the application of systems in sports for monitoring athletes' physiological conditions, optimizing training, and detecting risky conditions.