AI-Driven Predictive Models of Early Recurrence of HCC After Surgical Resection: A Systematic Review

Background/Objectives: Early recurrence after curative-intent resection is a major determinant of poor prognosis in hepatocellular carcinoma (HCC). Artificial intelligence (AI)-driven predictive models have emerged to identify patients at high risk of recurrence but remain incompletely synthesized for early recurrence specifically. This review aimed to identify and appraise AI-driven models predicting early recurrence after surgical resection. Methods: PubMed/MEDLINE, Scopus and Web of Science were searched from inception to November 2025. Eligible studies developed and evaluated AI-driven models predicting early recurrence (≤24 months) after curative-intent hepatectomy as first-line treatment for HCC. Risk of bias and applicability were assessed using PROBAST+AI, and findings were synthesized narratively due to methodological heterogeneity. The review was registered in PROSPERO. Results: Thirty-six studies involving 14,716 patients were included. Most studies originated from China (33/36, 91.7%), were single-center (27/36, 75%), and retrospective (35/36, 97.2%). Magnetic resonance imaging (MRI) was the predominant imaging modality (15/36, 41.7%), followed by computed tomography (CT) (11/36, 30.6%) and ultrasound (US)/contrast-enhanced ultrasound (CEUS) (6/36, 16.7%). Three studies developed non-imaging models, and one combined CT and MRI. In within-study comparisons, multimodal models generally showed better discrimination than unimodal approaches. Peritumoral, habitat-based, and multiphasic strategies appeared promising. However, external validation was reported in only 6/36 studies (16.7%), calibration and decision-curve analysis were inconsistently reported, and most studies had high risk of bias. Conclusions: AI-driven models show potential to predict early recurrence of HCC after curative-intent resection. Nevertheless, evidence remains limited by methodological heterogeneity and restricted geographical diversity, while clinical utility remains inconsistently evaluated, and no model has yet been generalized in clinical practice. Prospective multicenter studies with standardized outcomes, transparent reporting, and external validation are needed for clinical implementation.