Eye‐Tracking Phenotypes Across the Depressive Spectrum: Joint Modeling of Diagnostic Group and Symptom Severity

Aims To characterize eye‐tracking alterations across the depressive spectrum and evaluate their value for diagnostic stratification and symptom severity estimation. Methods In this cross‐sectional study, 183 participants completed a multi‐paradigm eye‐tracking battery. Group differences and associations with 17‐item Hamilton Depression Rating Scale (HAMD‐17) scores were examined with multiple‐comparison correction and covariate adjustment. Classification, regression, and multi‐task models used repeated nested resampling with fold‐specific preprocessing. Results False discovery rate (FDR)‐significant between‐group differences were observed for two first‐fixation latency measures from lateral‐gaze orienting and for antisaccade accuracy. Among 35 candidate features, 33 had covariate‐adjusted group associations at p < 0.05. Across 539 retained features, covariate‐adjusted generalized additive models identified eight FDR‐significant associations with HAMD‐17. In multi‐task analysis, Route B had numerically higher mean internal performance than Route A, although paired between‐route differences were imprecisely estimated (accuracy, 0.929 vs. 0.864; macro F1, 0.928 vs. 0.859; R 2 , 0.834 vs. 0.596). Regression‐only performance was similar to Route B (R 2 , 0.847). Conclusion Eye‐tracking alterations spanned the depressive spectrum and were associated with symptom severity after covariate adjustment. Between‐route differences were imprecisely estimated, and joint‐task training showed no clear improvement in severity estimation. External and temporal validation are required before clinical use.