Laurini de Souza Meyer, Erika

Background: Differentiated thyroid cancer (DTC) generally carries an excellent prognosis, yet recurrence occurs in up to 8–9% of patients. Artificial intelligence (AI) models are increasingly proposed to predict DTC recurrence; however, their methodological quality and clinical readiness are poorly defined. We conducted a systematic review to evaluate the methodological quality, risk of bias, and applicability of these models using the PROBAST + AI framework. Methods: We searched Ovid MEDLINE, Embase, Cochrane databases, Scopus, and Web of Science from inception to July 13, 2026. Eligible studies included original research developing or validating AI models for recurrence in DTC. Model quality (for development), risk of bias (for evaluation), and applicability were assessed using the PROBAST + AI framework, with a focus on discrimination, calibration, and validation strategies (PROSPERO; registration number [CRD420251004644]). Results: Of 673 records screened, 34 studies were included (21 unique datasets and 13 UCI-derived publications), encompassing 21,267 patients with a recurrence rate of 7.8%. Most were developed in Asian cohorts (17 of 21 datasets) and were structured clinical/tabular variables (82.4%), including demographic, clinical, laboratory, and pathological data, whereas the remaining models incorporated multimodal inputs such as imaging-derived or radiomic features. Notably, nearly one quarter of included studies (13 of 34, 38%) represented reanalysis of a single dataset. Discrimination was reported in 22 of 34 studies (64.7%) (area under the receiver operating characteristic curve [AUC] 0.67–1.00), though only 10 studies assessed calibration, none with standardized metrics. The PROBAST + AI assessment revealed high overall quality concerns in 26 of 34 (76.5%) development studies due to reliance on single-center cohorts, inadequate handling of missing data, and absence of external validation. Only 6 of 34 included studies (17.6%) were judged to have low overall quality concern, whereas 17 of 34 (50.0%) were rated as having low applicability concern. Conclusions: AI models for predicting DTC recurrence show promising discriminatory performance but have substantial methodological shortcomings in development, and a lack of calibration and external validation necessary for clinical use. Given the guidelines’ emphasis on risk-adapted de-escalation of therapy, the development of robust, validated, and clinically calibrated AI tools is a priority. Current evidence remains insufficient to support routine clinical use.