TITLE:
Cross-Population Transfer Learning for Antidepressant Treatment Response Prediction: A SHAP-Based Explainability Approach Using Synthetic Multi-Ethnic Data
AUTHORS:
Rocco de Filippis, Abdullah Al Foysal
KEYWORDS:
Antidepressant Response, Transfer Learning, Cross-Population Modelling, SHAP Explainability, Synthetic Clinical Data
JOURNAL NAME:
Open Access Library Journal,
Vol.13 No.1,
January
14,
2026
ABSTRACT: Accurate prediction of antidepressant treatment response remains a major challenge in psychiatry, particularly across diverse patient populations where genetic, demographic, and clinical characteristics vary substantially. In this study, we evaluate the potential of transfer learning to enhance predictive performance across heterogeneous cohorts. We generated a synthetic, population-stratified dataset representing four major demographic groups, European, East Asian, African, and Latin American, each characterized by clinical variables (age, gender, BMI, baseline Hamilton Depression Rating Scale [HAMD] score) and genetic factors (SNP1, SNP2, CYP2D6 metabolizer status). A baseline feedforward neural network was trained exclusively on the European cohort and assessed for zero-shot generalization to the remaining populations. Transfer learning was then applied by fine-tuning the base model on small samples from each target cohort. Model performance was quantified using AUROC, accuracy, and bootstrap-derived 95% confidence intervals. Explainability was incorporated via SHAP KernelExplainer to produce global feature importance rankings and local, instance-level explanations. The baseline model achieved high discrimination in European (AUROC 0.746) and African (0.714) cohorts but exhibited markedly reduced performance in East Asian (0.501) and Latin American (0.658) populations. SHAP analysis consistently identified gender, age, and baseline HAMD as top predictors, with CYP2D6 metabolizer status and SNP1 allele frequency contributing variably across populations. These results underscore the importance of population-specific fine-tuning to mitigate performance degradation when applying models beyond their source domain. Furthermore, the integration of SHAP explanations facilitates model interpretability, enabling clinicians to assess feature-level contributions and identify potential biases. While demonstrated here on synthetic data, this methodological framework provides a robust foundation for future validation using real-world, multi-ethnic patient datasets.