Sensor-based evaluation of intermittent fasting regimes: a machine learning and statistical approach

Abstract

The primary aim was to develop and assess the performance and applicability of different models utilizing sensor data to determine dietary adherence, specifically within the context of intermittent fasting. Our approach utilized time-series data from two completed human trials, which included continuous glucose monitoring, acceleration data, and food diaries, and a synthetic data set. Machine learning models achieved an average F1-score of 0.88 in distinguishing between fasting and non-fasting times, indicating a high level of reliability in classifying fasting states. The Hutchison Heuristic statistical method, while more moderate in performance, proved to be robust across different cohorts, including individuals with and without type 1 diabetes. A dashboard was developed to visualize results efficiently and in a user-friendly manner. The findings highlight the effectiveness of using sensor data, combined with advanced statistical and machine learning approaches, to passively evaluate dietary adherence in an intermittent fasting context.