Dokładność sztucznej inteligencji w przewidywaniu wyników operacji epilepsji: przegląd systematyczny i meta-analiza
Diagnostic accuracy of artificial intelligence models for seizure outcome prediction after epilepsy surgery: A systematic review and meta-analysis
W skrócie
Badacze sprawdzili, jak dobrze komputerowe programy (sztuczna inteligencja) potrafią przewidzieć, czy pacjent będzie wolny od napadów epilepsji po operacji. Przeanalizowali 10 badań obejmujących 721 pacjentów i stwierdzili, że programy te są prawidłowe w 85% przypadków, ale słabiej radzą sobie z rozpoznawaniem pacjentów, u których napady będą się powtarzać. Wyniki pokazują, że sztuczna inteligencja może być pomocnym narzędziem wspierającym decyzję lekarza, ale nie powinna być jedynym kryterium do podjęcia decyzji o operacji.
Oryginalny abstract (angielski)
Drug-resistant epilepsy (DRE) affects approximately one-third of individuals with epilepsy. Although surgery offers the best chance of durable seizure freedom in appropriately selected patients, postoperative outcomes remain difficult to predict. Artificial intelligence (AI) has been proposed to improve individualized presurgical prediction, but its diagnostic performance has not been quantitatively synthesized. We conducted a systematic review and meta-analysis to evaluate the diagnostic accuracy of AI models for predicting 12-month seizure freedom after surgery in patients with focal DRE. PubMed, Embase, Scopus, Web of Science, and IEEE Xplore were searched from January 2015 to October 2025, and studies reporting threshold-dependent metrics enabling reconstruction of 2 × 2 contingency tables were included. Ten retrospective studies comprising 721 evaluable patients were analyzed. Mean prevalence of seizure freedom was 60.5%. Pooled sensitivity was 0.85 (95% CI, 0.80-0.89), whereas specificity was 0.39 (95% CI, 0.24-0.58). The area under the hierarchical summary receiver operating characteristic curve was 0.82, and the pooled diagnostic odds ratio was 3.66 (95% CI, 1.45-9.20). Between-study heterogeneity was low for sensitivity (τ² = 0.06) and substantial for specificity (τ² = 1.11), with a strong negative correlation between sensitivity and false-positive rate (ρ ≈ -1.00). Externally validated models showed lower sensitivity than internally validated models (0.79 vs 0.87). Overall, AI-based models showed moderate discriminatory performance, with greater ability to identify patients likely to achieve seizure freedom than to exclude those at risk of persistent seizures. These findings support their use as adjunctive tools within presurgical evaluation rather than as stand-alone determinants of surgical candidacy.