TITLE:
Interpretable Machine Learning for Mood State Classification and Treatment Response Analysis Using Clinical and Biomarker Data
AUTHORS:
Rocco de Filippis, Abdullah Al Foysal
KEYWORDS:
Bipolar Disorder, Mood Classification, Machine Learning, Precision Psychiatry, Ensemble Models
JOURNAL NAME:
Open Access Library Journal,
Vol.12 No.8,
August
21,
2025
ABSTRACT: Mood disorders—particularly bipolar spectrum conditions—pose enduring diagnostic and therapeutic challenges due to their episodic nature, heterogeneous presentations, and reliance on subjective assessments. Conventional frameworks often delay accurate diagnosis and appropriate treatment. This study proposes an interpretable machine learning (ML) framework that integrates biological and behavioural data to support mood state classification and treatment response prediction in bipolar disorder. We constructed a multimodal dataset combining seven genetic polymorphisms (SNPs), ten serum biomarkers, and clinically relevant behavioural features such as stress levels, sleep deviation, and lithium adherence. Using this dataset, we developed and evaluated an ensemble-based ML pipeline incorporating Random Forest, XGBoost, and neural network classifiers, enhanced with SHAP-based interpretability and dimensionality reduction for decision visualization. The model demonstrated robust performance, achieving a macro-average ROC-AUC of 0.81, with the manic class yielding an AUC of 0.87. While manic states were well classified, euthymic states were frequently misclassified, reflecting the clinical ambiguity and low separability of remission features. Feature attribution consistently highlighted stress, lithium adherence, and sleep deviation as top predictors. Dimensionality reduction techniques (PCA, t-SNE, UMAP) revealed substantial class overlap, but improved separability for manic clusters. Stratification of biomarker distributions by treatment response further revealed clinical relevance in BDNF, DLPFC connectivity, and IL-6 levels. This work demonstrates the feasibility of using interpretable ML to enhance psychiatric diagnostics. By providing transparent, biologically grounded predictions, such systems may complement clinical judgment, enable early interventions, and support the development of precision psychiatry.