Interpretable Mental Health Screening Using Bayesian PCA-Enhanced Logistic Regression for Decision Support

Resumen

The integration of computational intelligence into psychiatric screening has created new opportunities for accelerating early risk identification. Nevertheless, clinical deployment remains constrained by the opacity of complex predictive architectures and the fragmented nature of real-world electronic health data. Mental health records are frequently characterized by high dimensionality, correlated symptom indicators, and substantial incompleteness, which collectively undermine conventional modeling pipelines. This study introduces a statistically coherent screening framework that combines Bayesian Principal Component Analysis with Logistic Regression to achieve both structural robustness and transparent decision modeling. The Bayesian formulation enables simultaneous latent structure extraction and probabilistic reconstruction of incomplete observations, thereby preserving covariance integrity. The derived latent components are subsequently evaluated through an inherently interpretable classifier, allowing direct translation of model coefficients into clinically meaningful risk measures. Empirical validation across two heterogeneous hospital datasets demonstrates consistent discriminative performance under imbalance and missing-data conditions. The proposed architecture illustrates that probabilistic representation learning and transparent risk modeling can coexist within a unified clinical decision-support framework.