Accurate classification of depression through optimized machine learning models on high-dimensional noisy data

Motivation: Depressive disorders are highly prevalent and impairing psychiatric conditions with neurocognitive abnormalities, including reduced event-related potential (ERP) measures of reward processing and emotional reactivity. Accurate classification of Major Depressive Disorder (MDD) based on ERP data could help improve our understanding of these alterations and propel novel diagnostic or screening measures. However, it has been particularly challenging due to the lack of generalization for noisy raw data with small sample sizes. We aim to improve classification performance for MDD using noisy ERP datasets using machine learning (ML) techniques. Results: We have developed two optimizations in our ML-based analysis of ERP datasets: effective feature extraction in the preprocessing of high-dimensional noisy data and enhanced classification through ensemble ML models. Together with a carefully designed validation strategy, our techniques provide a highly accurate method for MDD classification even for ERP data that are limited in sample size, inherently noisy and high-dimensional in nature. Our experimental results demonstrate that our ML optimizations achieve great accuracy and nearly perfect sensitivity simultaneously, particularly in classifying data samples unseen during the training process, compared to prior studies that perform regression-based classifications. Supplementary information: A supplementary document on ERP data collection is available.