Antisaccade and prosaccade performance in bipolar disorder and unipolar depression: a three-level meta-analysis

Differentiating unipolar depression (UD) from bipolar disorder (BD) is clinically challenging because depressive presentations overlap and objective markers are lacking. We quantified antisaccade (AS) and prosaccade (PS) accuracy and latency in UD and BD versus healthy controls and indirectly compared the diagnostic evidence sets. Following PRISMA 2020, PubMed, Scopus, Web of Science, and PsycINFO were searched from inception through 31 August 2026. Thirty-four studies (2,504 participants) were included. Clinically diagnosed datasets were synthesized with three-level random-effects models and study-level cluster-robust inference. Meta-regressions indirectly compared BD-versus-control and UD-versus-control effects. Moderator analyses explored heterogeneity, sensitivity analyses assessed robustness, and certainty was evaluated with an adapted GRADE framework. AS accuracy was poorer in BD (g = − 0.69, 95% CI [− 0.87, − 0.50]) and UD (g = − 0.52, 95% CI [− 0.74, − 0.31]) than in controls. AS latency was prolonged in BD (g = 0.85, 95% CI [0.52, 1.17]) but did not meet the conventional significance threshold in UD (g = 0.36, 95% CI [− 0.0006, 0.72], p = 0.0503). BD showed a smaller PS latency effect (g = 0.29, 95% CI [0.03, 0.56]); PS accuracy in BD and both PS outcomes in UD were non-significant. Stimulus type moderated PS latency in BD, with larger effects for faces than simple targets ( p = 0.031). Only AS latency differed nominally between BD-versus-control and UD-versus-control effects (β = 0.46, 95% CI [0.06, 0.86], p = 0.027); this contrast was no longer statistically significant after excluding two studies that enrolled both diagnostic groups ( p = 0.067). Sensitivity analyses did not reverse effect directions, although significance varied for AS latency in UD and both PS outcomes in BD. Certainty was very low for all eight primary outcomes. Poorer AS accuracy in both disorders suggests shared difficulty suppressing reflexive orienting and implementing rule-guided responses, whereas the larger BD AS latency effect may reflect additional slowing in response selection or preparation. Smaller, less consistent PS effects suggest relatively automatic orienting is less affected or more context-dependent. Direct prospective comparisons using standardized tasks are required before individual-level diagnostic or clinical applications can be considered.