Managing Data Uncertainty and Machine Learning for Adult ADHD Classification Using Accelerometry:OBF-Psychiatric Case Study

This study aims to enhance our understanding of ADHD individuals through accelerometer analysis while developing a framework for managing data uncertainty in digital biomarker research. Our primary emphasis is on identifying and mitigating biases within the OBF-Psychiatric dataset for ADHD and CONTROL groups, exploring how these biases influence machine learning model performance and generalizability. To balance patient inclusion with data quality, we applied innovative Pareto optimization and rigorous quality criteria, establishing optimal temporal windows, selecting 16:00 to 23:00 hours involving 53 patients from 77 available. Statistical analysis used robust Brunner-Munzel tests with False Discovery Rate correction, examining 34 comprehensive features spanning statistical, complexity, and frequency-based domains. Following thorough corrections, no significant differences in motor activity features emerged between the ADHD and CONTROL groups in quality-controlled data. Multidimensional scaling confirmed considerable overlap between groups. We assessed six traditional supervised machine learning algorithms through Leave-One-Patient-Out cross-validation: Logistic Regression, Random Forest, Support Vector Machine, Multilayer Perceptron, K-Nearest Neighbors, and XGBoost, plus baseline classifiers. Performance was evaluated across three data quality configurations to assess data processing consequences. Notably, performance systematically declined as data quality improved, with ROC AUC dropping from 75% (uncleaned) to 54% (quality-controlled). Our analysis suggests that the ADHD and CONTROL groups are indistinguishable using traditional feature-based motor activity patterns with these data collected in naturalistic conditions. However, the methodological contributions of this study provide foundations for future appropriately powered research and underscore key considerations for accelerometry utility in detecting adult ADHD and other psychiatric disorders in real-world scenarios.