Gastric cancer survival prediction using artificial intelligence models based on electronic health records: a systematic review and meta-analysis.

IMPORTANCE

Artificial intelligence (AI) is increasingly being applied to prognostic modeling in oncology; however, many AI-based survival prediction models rely on complex multimodal data that are not routinely available in clinical practice.

OBJECTIVE

This systematic review and meta-analysis aimed to evaluate the performance of AI models based on routinely collected electronic health record (EHR) data for predicting 5-year overall survival (5-OS) in patients undergoing surgical treatment for gastric cancer.

DATA SOURCES: A systematic literature search was conducted in PubMed, Scopus, Nature, MedRxiv, and bioRxiv databases for studies published between January 2015 and July 2025.

STUDY SELECTION: We included studies reporting area under the receiver operating characteristic curve (AUC) values for AI-based 5-OS prediction. Retrospective studies of adult patients with histologically confirmed gastric cancer who underwent curative-intent surgery were eligible, while studies primarily using non-routine multimodal data or lacking AUC outcomes were excluded.

DATA EXTRACTION AND SYNTHESIS: Risk of bias was assessed using the PROBAST-AI tool. Meta-analyses were performed to compare machine learning-based models with conventional statistical approaches, as well as different AI algorithm classes, including bagging and boosting ensemble methods, neural networks, random forest, support vector machines, and logistic regression. Random or fixed-effects models were applied according to between-study heterogeneity.

MAIN OUTCOMES AND MEASURES: The primary outcome was the pooled mean difference in AUC between machine learning-based and conventional statistical models for 5-OS prediction. Secondary outcomes included comparative performance across different AI algorithm classes and identification of the most frequently selected prognostic features.

RESULTS

Ten retrospective studies comprising 15,643 patients were included. Machine learning-based models demonstrated a modest but statistically significant improvement in predictive performance compared with conventional approaches, with a pooled mean AUC increase of 0.04 (95% CI 0.02-0.07; p = 0.001). Boosting algorithms showed a modest but statistically significant advantage over bagging methods (AUC increase 0.02; p = 0.04). The type of clinical input data, particularly the inclusion of blood-based biomarkers, influenced algorithm performance. The most consistently identified prognostic features across studies were age, T stage, tumor size, serum albumin or prealbumin level, and metastatic-to-examined lymph node ratio.

CONCLUSIONS

and relevance: AI-based prognostic models utilizing routinely available clinical data provide clinically meaningful improvements in 5-year survival prediction after gastric cancer surgery, and selection of the optimal AI algorithm should be guided by the structure and type of input data to maximize both predictive performance and practical applicability in clinical decision support systems.

SYSTEMATIC REVIEW REGISTRATION: https://www.crd.york.ac.uk/PROSPERO/view/CRD420261282797, PROSPERO CRD420261282797.