Machine learning-based models for tumor mutation burden prediction in gastrointestinal cancers: a systematic review and meta-analysis

Abstract

Background Tumor mutation burden (TMB) serves as a key biomarker guiding immunotherapy in gastrointestinal (GI) cancers, yet its measurement via whole-exome sequencing (WES) is costly and invasive. Machine learning (ML)-based models and radiogenomics provide promising non-invasive alternatives, but their diagnostic performance and methodological consistency remain unclear. Objective To systematically evaluate and meta-analyze the diagnostic performance, methodological rigor, and reporting quality of ML-based models developed for TMB prediction in GI cancers. Methods PubMed, Scopus, and Web of Science were searched through January 2025 for studies applying ML or deep learning to predict TMB in human GI cancers. Risk of bias was assessed using the Cochrane QUADAS-2 framework adapted for AI prediction studies. Pooled estimates for area under the curve (AUC) and accuracy were obtained under a restricted maximum-likelihood random-effects model, with heterogeneity quantified by I & sup2; and sensitivity analyses exploring threshold effects. Deeks' funnel asymmetry test was used to assess publication bias. Subgroup analyses examined cancer type and model architecture. Results Ten studies met inclusion criteria. The pooled AUC was 0.89 (95 CI 0.80-0.97; I & sup2; = 92.1), and pooled accuracy was 0.86 (95 CI 0.79-0.94; I & sup2; = 83.9). Graph neural networks achieved the most stable performance (AUC approximate to 0.97), while classical ML models showed consistent results on smaller datasets. Publication bias was significant for AUC (p = 0.007) but not for accuracy (p = 0.15), indicating outcome-specific reporting tendencies. Only 3/10 studies performed external validation, and calibration metrics were rarely reported. Subgroup findings suggested that heterogeneity stemmed more from model architecture and dataset design than from cancer subtype. Conclusion ML-based and radiogenomic models demonstrate high diagnostic potential for predicting TMB in GI cancers, particularly with graph-based architectures. However, scarce external validation, inconsistent TMB definitions, and selective reporting of AUC limit clinical generalizability. Standardized reporting of discrimination and calibration metrics, alongside external validation, is essential to translate ML-driven TMB prediction into reliable precision oncology tools.