Comparison of feature learning methods for non-invasive interstitial glucose prediction using wearable sensors in healthy cohorts: a pilot study

• Published in: Intelligent medicine Vol. 4; no. 4; pp. 226 - 238
• Main Authors: Huang, Xinyu, Schmelter, Franziska, Uhlig, Annemarie, Irshad, Muhammad Tausif, Nisar, Muhammad Adeel, Piet, Artur, Jablonski, Lennart, Witt, Oliver, Schröder, Torsten, Sina, Christian, Grzegorzek, Marcin
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.11.2024; Institute of Medical Informatics, University of Lübeck, Germany%Institute of Nutritional Medicine, University of Luebeck and University Medical Center Schleswig-Holstein, Lübeck, Germany%Department of IT, University of the Punjab, Lahore, Pakistan%Perfood GmbH, Research & Development, Lübeck, Germany%Fraunhofer Research Institution for Individualized and Cell-Based Medical Engineering (IMTE), Lübeck, Germany%German Research Center for Artificial Intelligence (DFKI), Lübeck, Germany
• Subjects: Clarke error grid; Deep learning; Interstitial glucose prediction; Non-invasive glucose monitoring; Physiological signal processing; Wearable sensors; Deep learning; Clarke error grid; Physiological signal processing; Interstitial glucose prediction; Wearable sensors; Non-invasive glucose monitoring
• ISSN: 2667-1026
• DOI: 10.1016/j.imed.2024.05.002

Alterations in glucose metabolism, especially the postprandial glucose response (PPGR), are crucial contributors to metabolic dysfunction, which underlies the pathogenesis of metabolic syndrome. Personalized low-glycemic diets have shown promise in reducing postprandial glucose spikes. However, current methods such as invasive continuous glucose monitoring (CGM) or multi-omics data integration to assess PPGR have limitations, including cost and invasiveness that hinder the widespread adoption of these methods in primary disease prevention. Our aim was to assess machine learning algorithms for predicting individual PPGR using non-invasive wearable devices, thereby, circumventing the limitations associated with the existing approaches. By identifying the most accurate model, we sought to provide a more accessible and efficient method for managing glucose metabolic dysfunction. This data-driven analysis used the experimental dataset from the SENSE (”Systemische Ernährungsmedizin”) study. Healthy participants used an Empatica E4 wristband and Abbott Freestyle Libre 3 CGM for 10 days. Blood volume pulse, electrodermal activity, heart rate, skin temperature, and the corresponding CGM values were measured. Subsequently, four data-driven deep learning (DL) models-convolutional neural network, lightweight transformer, long short-term memory with attention, and Bi-directional LSTM (BiLSTM) were implemented and compared to determine the potential of DL in predicting interstitial glucose levels without involving food and activity logs. The proposed BiLSTM achieved the best interstitial glucose prediction performance, with an average root mean squared error of 13.42 mg/dL, an average mean absolute percentage error of 0.12, and only 3.01% values falling within area D in Clarke error grid analysis, incorporating the leave-one-out cross-validation strategy for a five-minute prediction horizon. The findings of this study may demonstrate the feasibility of transferring knowledge gained from invasive glucose monitoring devices to non-invasive approaches. Furthermore, it could emphasize the promising prospects of combining DL with wearable technologies to predict glucose levels in healthy individuals.