Predicting RTMS Treatment Response in Depression: Use of Machine Learning Models to Identify the Roles of Metabolic and Clinical Factors

Background Depression is a prevalent and challenging mental health condition, often resistant to traditional therapies. Repetitive transcranial magnetic stimulation (rTMS) is an FDA-approved treatment for resistant depression, but predicting treatment response remains difficult. Metabolic factors, such as obesity, diabetes, and associated inflammatory pathways, are increasingly recognized as influencing depression treatment outcomes. Leveraging machine learning may provide insights into how these factors predict rTMS efficacy, enabling more personalized treatments. Aims & Objectives This study aims to identify key metabolic and clinical factors that predict rTMS treatment outcomes in patients with major depressive disorder (MDD) and bipolar disorder (BD). To achieve this, machine learning models were employed to develop predictive frameworks tailored for small participant groups. Method Sixty-eight patients (28 with BD) participated, undergoing 12 sessions of rTMS targeting the left dorsolateral prefrontal cortex. Participants were assessed at baseline, Week 4, and Week 12 using the Hamilton Depression Rating Scale (HDRS). Predictive models were developed using recursive feature elimination (RFE) and four machine learning algorithms: Random Forest, Gradient Boosting, Extra Trees, and XGBoost. Key features identified included body mass index (BMI), cholesterol ratios, and insulin resistance markers. SHAP (Shapley Additive Explanation) values were applied for model interpretation. Results At Week 4, the remission rate was 59.6%, with BMI, systolic blood pressure (SBP), and LDL/HDL ratio emerging as significant predictors. The Extra Trees model achieved the highest performance (AUC = 0.766, accuracy = 0.726). At Week 12, remission was 61.5%, and predictors included insulin resistance (HOMA-IR), brain-derived neurotrophic factor (BDNF), and LDL/HDL ratio. The Gradient Boosting model excelled (AUC = 0.862, accuracy = 0.773). SHAP analysis highlighted BMI and HOMA-IR as the most influential factors at Weeks 4 and 12, respectively. Discussion & Conclusions The study underscores the impact of metabolic health on depression treatment outcomes. Metabolic dysregulation, particularly obesity and insulin resistance, may impair neuroplasticity and inflammation pathways, complicating treatment. BDNF's role in neurogenesis and metabolism was also highlighted as pivotal for recovery. Machine learning models demonstrated robust predictive performance, emphasizing the potential of integrating clinical and metabolic data to refine rTMS treatment approaches. Metabolic and clinical factors significantly influence rTMS outcomes in depression. Machine learning models, especially Extra Trees and Gradient Boosting, offer promising tools for individualized treatment planning. Future research should validate these findings in larger populations and explore additional biomarkers. Personalized approaches that account for metabolic health could reduce the trial-and-error nature of depression treatment and improve patient outcomes.