Resting-state fMRI-based machine learning for predicting SSRI treatment response in major depressive disorder.

backgroundMajor Depressive Disorder (MDD) is a prevalent mental health condition with significant societal impact. Although prior research has highlighted the brain changes modulated by antidepressant therapy, their efficacy and effectiveness are debated. The low rates of treatment response still existed in the pharmacological therapy of MDD. Exploring an optimal neurological predictor of symptom improvement caused by pharmacotherapy is urgently needed for improving response to treatment. Our purpose is to develop a predictive model for MDD therapy using machine learning techniques based on resting-state fMRI metrics.

methodsA total of 116 MDD patients underwent 3.0T resting-state magnetic resonance image scanning. Demographic data and the 24-item Hamilton Depression Rating Scale (HAMD-24) were collected from all participants. An additional independent cohort of 25 MDD patients was included for external model validation. Based on the reduction rate of HAMD-24 scores at different time points, the patients were divided into an early improvement group, an early response group, and a clinical response group: (1) Early improvement group: HAMD-24 reduction rate ≥ 20% after 1 week of treatment; (2) Early response group: HAMD-24 reduction rate ≥ 25% after 2 weeks of treatment; (3) Clinical response group: HAMD-24 reduction rate ≥ 50% after 4 weeks of treatment. For model construction, LASSO regression was applied for feature selection, integrating identified brain regions, HAMD-24 symptom clusters, and clinical variables. Five machine learning models were established based on features screened by the LASSO regression model to predict treatment response in MDD. All models were trained and assessed using 10-fold cross-validation, the most stable logistic regression model was selected for external validation in a temporally independent cohort (n = 25).

resultsAt 1 week, alterations were mainly observed in the frontal and sensorimotor regions. At 2 weeks, differences were found in the opercular, postcentral, orbitofrontal, temporal, and angular areas. At 4 weeks, additional abnormalities appeared in the frontal, occipital, and postcentral cortices. Moreover, right Postcentral gyru was found to be present in the differential brain regions observed at 1 week, 2 weeks, and 4 weeks comparisons. In addition, ten features related to the treatment response were identified using the LASSO regression model, including age, years of education, core depressive symptoms, and several brain region features primarily involving the postcentral gyrus. Among the constructed machine learning models, the Logistic Regression model based on LASSO-selected features (including age, education, core depressive symptoms, and the postcentral gyrus) demonstrated the most stable performance. It achieved an area under the curve (AUC) of 0.801 in the internal validation and maintained robust predictive performance in the independent external validation set, indicating promising generalizability. Following independent external validation, the results showed that the predictive performance of the Logistic regression model on the external validation set yielded an AUC value of 0.643 and an accuracy of 0.60. This indicates that the brain regions involved in the model-namely the right postcentral gyrus, right precentral gyrus, left superior frontal gyrus, left middle occipital gyrus, and the orbital part of the right inferior frontal gyrus-hold promise as exploratory neural correlates for predicting the efficacy of SSRIs in depression.

conclusionMDD exhibits early functional alterations in brain regions involved in emotion regulation, cognition, and sensorimotor processing. Integrating these multi-metric neural features with clinical variables via machine learning models suggests potential clinical utility for evaluating and predicting early SSRI treatment outcomes.