Identifying clinical features associated with electroconvulsive therapy response in adolescents with major depressive disorder using machine learning.

Background: Electroconvulsive therapy (ECT) is an effective treatment for adolescent major depressive disorder (MDD), but its efficacy varies. This study utilized machine learning (ML) to identify baseline clinical factors associated with poor ECT response. Methods: We retrospectively enrolled 503 adolescent MDD patients. A poor response was defined as a <50% reduction on the Hamilton Depression Scale (HAMD-24). The optimal ML algorithm (Random Forest, RF) was selected from nine candidates and then simplified using recursive feature elimination (RFE) and interpreted via Shapley Additive Explanations (SHAP). Results: A simplified model using two baseline features-the neutrophil-to-platelet ratio (NPR) and pre-treatment HAMD score-achieved an AUC of 0.731 on the testing set, comparable to the full-feature model (AUC: 0.751). SHAP analysis revealed that a lower baseline NPR and a lower pre-treatment HAMD score were associated with a poor response. Furthermore, retrospective statistical comparisons revealed that patients in the poor response group completed significantly fewer ECT sessions than those in the good response group. Conclusions: We developed a concise explanatory model demonstrating that routine clinical data available at admission (blood NPR and HAMD score) can effectively stratify the risk of poor ECT efficacy. Crucially, identifying these high-risk patients early empowers clinicians to implement targeted management, ensuring they complete a full and adequate course of ECT to maximize therapeutic benefits and prevent premature termination.