Artificial intelligence–based predictive models for recurrence and recurrence-free survival in gastrointestinal stromal tumors: A contemporary systematic review.

Abstract

11536 Background: Recurrence after complete resection of GISTs is still a major contributing factor to long-term survival. Traditional stratification models for risk depend on clinicopathologic variables and have limited precision. Recent improvements in artificial intelligence (AI), such as machine learning (ML), deep learning (DL), and multimodal methodologies, might enhance personalized prediction of recurrence and recurrence-free survival (RFS). We performed a contemporary systematic review to evaluate the performance and clinical utility of AI-based prognostic models in resected GIST. Methods: PRISMA 2020 guidelines were followed for conducting a systematic review. PubMed, Scopus, and Web of Science were analyzed to search published studies published during 2019–2025 that evaluated AI-, ML-, or DL-based models predicting recurrence or RFS following complete surgical resection of localized primary GISTs. Eligible studies included retrospective or multicenter cohorts with radiologic, histopathologic, genomic, or multimodal data. Data extracted included model type, input modalities, validation strategy, and performance metrics (area under the curve [AUC] and concordance index [C−index]). Results: Eight studies encompassing approximately 4,000 patients met inclusion criteria. Both deep learning and multimodal fusion models showed the highest prognostic accuracy (C-index values 0.86–0.96, AUC up to 0.995). Radiomics based ML models using CT, MRI, or ultrasound yielded AUCs between 0.85 and 0.92 which were consistently superior to traditional clinicopathologic methodologies. Genomic ML models fine-tuned recurrence risk stratification beyond conventional benchmarks into molecularly distinct prognostic subgroups. DL-based histopathology estimators both predicted for RFS and for key driver mutations (KIT, PDGFRA) by bridging morphologic and molecular features. External validation cohorts showed stable performance (AUC 0.87–0.96), with calibration and decision curve analyses supporting clinical utility. Conclusions: AI-based predictive models demonstrate strong and reproducible performance for recurrence and RFS prediction following GIST resection, consistently outperforming traditional risk stratification systems. Multimodal and deep learning approaches integrating radiologic, pathologic, and genomic data appear most promising for precision prognostication and personalized adjuvant therapy selection. Prospective validation and explainable AI integration are needed prior to routine clinical adoption.