From Traditional Inspection to Quantitative Imaging: Tongue and Facial Color Features for Automated Machine Learning-Driven Depression and Schizophrenia Classification

Abstract

Objective and non-invasive markers for psychiatric assessment remain limited. This study evaluated whether standardized tongue and facial color features provide measurable information relevant to depression and schizophrenia classification. Tongue and facial images were collected from 749 participants, including healthy controls (n = 84), patients with depression (n = 246), and patients with schizophrenia (n = 419). Color characteristics were quantified in predefined tongue and facial regions using the LAB color space. Group differences were examined statistically, and machine-learning models were evaluated across five repeated stratified splits. Most LAB-derived features differed significantly across groups, with luminance-related measures showing the largest effect sizes and more consistent shifts in schizophrenia than in depression. In multiclass classification, LAB plus demographic variables achieved strong performance (accuracy = 0.782 ± 0.035, macro-F1 = 0.697 ± 0.038, AUC = 0.912 ± 0.027), similar to LAB plus demographic and traditional variables (AUC = 0.912 ± 0.032). LAB-only models showed comparable AUC to demographic-only models but lower macro-F1. In pairwise analyses, discrimination was strongest for healthy control versus schizophrenia (AUC = 0.952 ± 0.030) and depression versus schizophrenia (AUC = 0.926 ± 0.019), and lower for healthy control versus depression (AUC = 0.817 ± 0.054). These findings suggest that LAB-derived tongue and facial color features may provide complementary group-level information, particularly when combined with demographic variables, but should not be interpreted as standalone diagnostic biomarkers.