Artificial intelligence for the data-driven diagnosis of ADHD: a systematic review and meta-analysis

Attention-deficit/hyperactivity disorder (ADHD) is clinically heterogeneous, which complicates the development of reproducible data-driven diagnostic models. Following PRISMA guidelines, this systematic review included 54 eligible studies. The pooled sensitivity was 0.87 (95% CI: 0.83–0.91) and the pooled specificity was 0.91 (95% CI: 0.88–0.93), but residual heterogeneity remained above 96% in meta-regression. PROBAST classified 11 studies (20.4%) as high risk of bias, 11 (20.4%) as unclear risk, and 32 (59.3%) as low risk; the analysis domain was the main source of high-risk judgments. Publication bias was detected for sensitivity but not specificity. No statistically significant between-modality differences were observed for sensitivity or specificity. The pooled values therefore describe a highly heterogeneous literature and should not be interpreted as evidence that one modality is superior or that current models are ready for clinical use. Standardized reporting, external validation, and prospective evaluation are needed before clinical translation. A sensitivity analysis excluding the 11 studies at high overall risk of bias produced similar pooled sensitivity (0.87, 95% CI: 0.82–0.91) and specificity (0.92, 95% CI: 0.88–0.94), while heterogeneity remained substantial ( I ² = 97.9% and 98.5%, respectively).