Identifying activity patterns from different actigraphy devices in people with schizophrenia spectrum disorders: a comparative study

Abstract Actigraphy is a reliable and non-invasive method for assessing motor activity and how disease affects the daily activity of patients. Machine learning models applied to actigraphy data have been used to identify activity patterns in patients suffering from either major depressive disorder or Schizophrenia Spectrum Disorders (SSDs). However, heterogeneous collection protocols across different datasets significantly hinder generalizability, making direct comparison difficult. In this work, we used actigraphy data to inform a machine learning model that distinguishes between the activity patterns of healthy controls (HC) and patients with SSDs. Actigraphy recordings from a total of 258 subjects (126 HC and 132 with SSDs) from the DiAPASon and PSYKOSE datasets, each spanning one week, were used. Actigraphy patterns were classified using principal component analysis (PCA) followed by a support vector machine (SVM) classifier trained on DiAPASon data. The model was validated using independent data from PSYKOSE, collected with a different actigraphy device, after a rescaling procedure to match DiAPASon. After rescaling, the concordance between DiAPASon and PSYKOSE was 99.9% (30.7% pre-rescaling). The best performing model achieved a classification accuracy of 0.860 for DiAPASon data and 0.80 for PSYKOSE, demonstrating the model’s generalizability. Using Shapley analysis, distinctive activity patterns driving classifications towards HC or SSDs classes were detected. Our results stand as a proof of concept that machine learning models, able to identify activity patterns, can be generalized to other actigraphy devices, but further testing on more diverse datasets and devices is required.