TITLE:
Predicting Treatment Response in Bipolar Disorder Using Biomarker Profiles and Machine Learning Models
AUTHORS:
Rocco de Filippis, Abdullah Al Foysal
KEYWORDS:
Bipolar Disorder, Treatment Response Prediction, Machine Learning, Deep Learning, Biomarkers, Precision Psychiatry, Neuroimaging, Genomics, Feature Importance, SHAP Values, Model Interpretability, Computational Psychiatry
JOURNAL NAME:
Open Access Library Journal,
Vol.12 No.8,
August
18,
2025
ABSTRACT: Bipolar disorder is a multifaceted psychiatric illness characterized by unpredictable mood episodes and highly variable treatment responses across individuals. Predicting response to specific pharmacological treatments remains a key challenge in personalized psychiatry. This study aims to develop predictive models for treatment response subtypes—non-responders, lithium responders, and anticonvulsant responders—using a diverse array of biomarkers, including genetic variants, serum levels, neuroimaging-derived features, and clinical history. A dataset of 2000 patients was analyzed, containing 31 features spanning single nucleotide polymorphisms (SNPs), inflammatory and neurochemical markers, structural and functional brain imaging variables, and illness course descriptors. Initial exploratory data analysis revealed two variables with missing values, and class imbalance across response types. Correlation analysis highlighted strong associations between GABA, DLPFC_connectivity, and treatment outcomes. Dimensionality reduction with UMAP illustrated overlapping distributions among classes, justifying the need for non-linear classifiers. Five models—logistic regression, SVM, random forest, XGBoost, and a deep neural network—were trained and evaluated. The deep learning model achieved the highest validation accuracy (46%) and ROC AUC (0.65). Feature importance analysis across models identified BDNF_serum, COMT_Val158Met, and DLPFC_connectivity as top contributors. Despite comparable performance among classical models, deep learning showed superior generalization and interpretability through its learning curve. Our findings underscore the feasibility of integrating multimodal biomarkers and deep learning for accurate stratification of bipolar disorder treatment response. The results support the future development of decision-support tools that incorporate genetic, proteomic, and neurobiological data to guide personalized psychiatry. Future work will include external validation, imputation strategies, and further interpretability using SHAP values.