Multimodal machine learning models for predicting remission in major depressive disorder using clinical data, blood biomarkers, and DNA methylation

초록

Major depressive disorder (MDD) is a leading global health burden, yet only one-third of patients achieve remission with initial antidepressant therapy. Inflammatory biomarkers and epigenetic signatures such as DNA methylation have been implicated in treatment response, but their temporal predictive utility remains unclear. We analyzed 821 Korean patients with MDD from the MAKE BETTER study, integrating clinical variables, serum inflammatory biomarkers, and DNA methylation profiles into machine-learning models. Twelve-month remission was modeled as prospective prediction using baseline and week-12 treatment data, whereas early improvement (2 weeks) and 12-week remission were assessed as exploratory classification tasks using all 12week data (look-ahead bias). For 12-month remission, XGBoost achieved AUROC 0.728 and AUPRC 0.840. For 12-week remission, logistic regression achieved AUROC 0.742 and AUPRC 0.595. Predictive drivers shifted over time, from baseline clinical severity (early response) and antidepressant dosage (12-week remission) to inflammatory/epigenetic markers (hs-CRP and epigenetic inflammation score, EIS) for 12-month remission. Differential methylation results showed increasing numbers of significant CpGs over time, with inflammation-linked CpGs providing stable contributions. A two-CpG signature (cg10636246 in AIM2; cg02650017 near ABCG1/PHOSPHO1) achieved AUROC 0.757 and AUPRC 0.854, supporting compact epigenetic signatures for long-term risk stratification. Clinical and treatment features were most informative for short-term outcomes, whereas inflammatory and epigenetic markers became increasingly important for long-term remission. These findings support precision psychiatry approaches integrating dynamic multimodal features, emphasizing treatment exposure for acute management and inflammation-related markers for long-term planning.

키워드