AI-driven actigraphy data reconstruction

Abstract

Actigraphy is a non-invasive technique that uses a wrist-worn device to track sleep patterns and circadian rhythms, providing valuable data for diagnosing sleep disorders such as insomnia and hypersomnia. However, gaps in actigraphy data often occur when the device is ”off-wrist.” This study explores the use of advanced Machine Learning and Deep Learning techniques to impute missing data, aiming to improve the accuracy and reliability of the analyses. The main objective is to develop robust data imputation pipelines utilizing a range of algorithms, including regression models (Random Forests, XGBoost, and Support Vector Regression), with hyperparameter optimization strategies such as Grid Search and Randomized Search and autoencoders optimized by Keras Tuner. Along with these regression strategies, we have vital data filtering strategies, including Dynamic Time Warping, leveraging the M10 variable, which represents the 10 hours of highest daily activity, and considering data surrounding off-wrist periods. In this study, the methodology preprocesses actigraphy datasets from the ActTrust device, incorporating variables such as temperature, light exposure, and activity metrics. Preliminary results demonstrate that autoencoder models outperform traditional methods in imputing missing data, significantly reducing Mean Squared Error. Combined models tailored to specific activity patterns further enhance predictive accuracy, as validated by statistical tests such as the Wilcoxon test. These findings highlight the potential of these models to improve clinical practice by enabling personalized sleep assessments and supporting targeted interventions.