Robust Body Temperature Estimation for Smartwatch using Ensemble of DL Models

In this paper we present a method of core body temperature (BT) estimation using sensors available in smartwatch. An ensemble of deep learning (DL) models was trained and validated with large and diverse dataset. Totally 304 human subjects participated in 467 tests (including 55 febrile subjects with oral temperature ≥ 37.5°C). The algorithm allows to estimate user’s BT with root mean square error (RMSE) = 0.26°C and detect febrile illness with F1-score = 0.84 at the test dataset. Data augmentation is one of the key development point of our algorithm. Three techniques were used to augment the training dataset to reduce RMSE of BT estimation significantly at the test dataset. The best metrics were achieved with ensembling of DL models (time and tabular domains). The limit of metrics for our task was assessed using a machine learning (ML) approach based on two medical thermometers (oral and axillary) data and general parameters of user profile. Finally, we found that the performance of smartwatch-based solution is very close to the clinically widespread method of fever detection (compare F1-score 0.84 and 0.90 respectively). To the authors' knowledge, the proposed algorithm provides the highest performance among existing solutions or ever described in the literature in the area of smartwatch-based true body temperature estimation.