Diagnostic accuracy of tongue diagnosis for stroke syndrome differentiation in traditional Chinese medicine: a systematic review, meta-analysis, and narrative synthesis of artificial intelligence-assisted models

Background Tongue examination is central to traditional Chinese medicine (TCM) syndrome differentiation after stroke, but its diagnostic accuracy has not been evaluated using contemporary diagnostic test accuracy (DTA) methods. Methods We searched PubMed, Embase, Web of Science, CNKI, Wanfang, and VIP from inception to 23 February 2026 (PROSPERO: CRD420261288502). Track 1 studies of manual tongue diagnosis were pooled using bivariate or univariate random-effects models; Track 2 studies of artificial intelligence (AI)-assisted models were synthesized narratively according to SWiM. Risk of bias was assessed with QUADAS-2 and the certainty of Track 1 evidence with GRADE-DTA. Results Ten studies from mainland China included 3,437 participants: six manual-diagnosis studies ( N = 2,191) and four AI-assisted studies ( N = 1,246). Because evidence certainty was very low and reference-standard independence was not assured, all estimates were treated as exploratory agreement with existing syndrome-classification processes rather than clinically validated diagnostic accuracy. The Deficient-Cold Tongue cluster showed pooled specificity of 0.97 (95% CI: 0.87–1.00) and a positive likelihood ratio (PLR) of 31.0 (95% CI: 2.8–341.6). The Greasy Coating cluster showed pooled sensitivity of 0.83 (95% CI: 0.52–0.96), specificity of 0.80 (95% CI: 0.68–0.88), PLR of 4.1 (95% CI: 2.5–6.8), and a negative likelihood ratio of 0.21 (95% CI: 0.06–0.71). Blood Stasis Tongue showed low sensitivity and unstable, heterogeneous estimates. The small analyzable evidence base contrasted with the much larger descriptive literature. Conclusion Current evidence remains too uncertain to support manual tongue diagnosis for routine or standalone clinical use, and applicability is largely limited to mainland Chinese TCM stroke-care settings. Manual and AI-assisted approaches require prospective multicenter validation using independent syndrome adjudication that excludes the index tongue feature. Future studies should stratify analyses by stroke etiology and report calibration and clinical utility for AI models. Systematic review registration PROSPERO , CRD420261288502.