Abstract
<title>Abstract</title> <p>Facial fear recognition is a behavioural marker of theoretical and clinical relevance, but the literature is fragmented by methodological and cultural heterogeneity. This PRISMA 2020 systematic review and meta-analysis synthesises 40 studies (2016–2025) on fear recognition in healthy and clinical adults. Risk of bias was assessed with a hybrid PROBAST/AXIS tool. Pooled fear accuracy in healthy controls, from a DerSimonian–Laird random-effects logit model across 21 studies (n = 1369), was 54.5% (95% CI 45.8–62.9), with very large heterogeneity (I² = 87.6%) and a range of 19% to 94%. This dispersion was largely methodological: paradigms with fewer than six response options yielded a far higher raw pool (80.7%; k = 4) than those with six or more (47.8%; k = 17), reflecting the higher chance baseline and the exclusion of surprise rather than a true difference in ability. Across three studies with complete confusion matrices, the pooled fear → surprise rate was 32.3% (95% CI 25.5–39.8; I² = 4%), conserved across method and culture. Aggregate estimates should be read alongside subgroup analyses. The fear–surprise confusion may be a differential indicator worth assessing with standardized procedures and detailed error analyses.</p>