Predicting the longitudinal efficacy of medication for depression using electroencephalography and machine learning

At present, determining an effective antidepressant regimen for major depressive disorders remains largely a matter of trial-and-error. This study assessed the viability of using machine learning methods to predict short- and long-term responses to antidepressants based primarily on electroencephalographic (EEG) recordings.

This study assessed the effectiveness of several machine learning-based algorithms in predicting treatment outcomes at 4, 6, and 8 weeks after medication initiation. Training data included clinical features, electroencephalographic (EEG) recordings, and functional connectivity metrics collected from 77 patients at baseline and after taking medication for 1 week, as well as change index features computed as the difference between the two time points. A generalized estimation equation (GEE) model was also utilized to integrate longitudinal data, providing insights into the evolution of model performance and experiment variables over time.

The machine learning models achieved prediction accuracies of 83.1 % (Week 4), 73.3 % (Week 6), and 80.0 % (Week 8). Functional connectivity analysis and phase synchronization emerged as critical contributors to prediction accuracy.

These findings highlight the potential of using non-invasive EEG recordings and data-driven methods to optimize clinical decision-making. Functional connectivity metrics could be used by clinicians to identify responders early in the treatment process, thereby reducing the burden of lengthy trial-and-error methods. Future research should focus on larger datasets, alternative feature extraction techniques, and the integration of multimodal biomarkers. These advancements could pave the way for personalized treatment strategies to enhance patient outcomes and mitigate the socio-economic impact of MDD. The institutional review board of Taipei Veterans General Hospital (2021-07-01B) approved this study.