Abstract
Sweat biomarkers offer valuable insights into the health conditions of individuals. Despite the recent advances in wearable technologies that enable real-time monitoring of sweat biomarkers, their potential to infer health conditions remains largely unexplored. This thesis, conducted as part of the WeCare project, leverages machine learning (ML) models including deep neural networks (DNNs), for real-time predictive health monitoring using these sweat biomarkers. Our research primarily focuses on predicting physiological states such as hydration status and core body temperature during exercise.
One version of the wearable sweat patch developed by our WeCare partner at the Instituto de Microelectrónica de Barcelona (IMB-CNM) uses ion-sensitive field-effect transistors (ISFETs). While these sensors are sensitive, lightweight, and cost-effective, they are prone to sensor drift. Previous work shows that DNNs are promising for predicting ionic concentration from ISFET sensor readings with the presence of sensor drift. However, training DNNs requires large labeled datasets that are difficult to collect. To address this, we first construct a physical model for ISFET sensors that simulates sensor readings and takes into account sensor drift. We then train an end-to-end prediction neural network as a sensor calibration tool on these simulated readings. Our prediction network outperforms two manual calibration methods in predicting sodium concentration from uncalibrated real-world sodium ISFET readings, suggesting its promise for future calibration of wearable patches using ISFETs.
Next, we carry out a study aimed at designing personalized hydration strategies based on noninvasive biomarkers. We examine the feasibility of using ML models to predict the hydration status of an athlete using physiological and sweat biomarker recordings collected from a subject during a set of indoor cycling sessions supervised by the Lausanne University Hospital (CHUV). Because the wearable sweat patches were still under development at that stage, absorbent patches were used for sweat sample collection. We also compared the performance of nonlinear ML models with the linear model on this predictive task. This investigation provides insights for future hydration status predictions using ML models on sweat biomarker data collected from wearable sweat patches.
Finally, following the available printed sensor patch developed by the Soft Transducers Lab at École Polytechnique Fédérale de Lausanne (EPFL-LMTS), we determine the prediction accuracy of core body temperature during exercise using real-time sweat biomarkers measured with this wearable prototype and with the addition of other biomarker data collected with commercial devices. All experimental sessions were conducted at CHUV. Our results indicate that DNNs can accurately and continuously predict core body temperature solely from sweat biomarker data, specifically sweat sodium and potassium concentrations collected from the wearable patch. Moreover, our analysis of the collected sweat biomarker data shows that they can be used to predict future core body temperature values. Our findings highlight the potential of integrating advanced predictive models with wearable sweat patches for real-time and accurate prediction of physiological states.
Wang, Shu