From prediction to intervention: Translating machine learning risk models into clinical decision support for gastric cancer surgery

Gastric cancer remains a major global health burden, and radical gastrectomy is the cornerstone of curative treatment.Postoperative complications -particularly delayed wound healing, anastomotic leakage, and surgical site infections -significantly impair patient recovery, prolong hospitalization, and compromise long-term oncological outcomes.Machine learning (ML) models have shown promising discrimination for these complications in some cohorts; however, reported performance is heterogeneous:Area under the receiver operating characteristic curve values most commonly fall in the range of 0.75-0.85,and many estimates are derived from single-center retrospective cohorts, so that estimates above 0.85 should be interpreted with caution.Nevertheless, a substantial gap persists between model development and clinical deployment: The vast majority of high-performing predictive models remain confined to academic publications; only a few are distributed as web calculators or applications or have been piloted in a single centre, and none has yet been evaluated for its effect on surgical decisions or patient outcomes.In this narrative review, we examine the barriers to this translation across four dimensions: Technical (external validation, model interpretability, data heterogeneity, and data and concept drift), clinical (physician trust, workflow integration, and actionability), organizational/regulatory (approval pathways, reimbursement, and liability), and ethical (algorithmic fairness, privacy, and autonomy).We critically evaluate existing clinical decision support system (CDSS) implementations in abdominal surgery, explicitly distinguish direct gastric-cancer evidence from indirect evidence derived from other specialties, extract lessons from both successes and failures, and propose a conceptual five-stage integrative translational framework: (1) Clinical needs definition; (2) Intrinsically explainable model design; (3) Prospective multicenter validation with decision curve analysis; (4) Workflow-embedded user-centered CDSS prototyping; and (5) Post-deployment monitoring with continuous learning.We further highlight how contemporary 4 / 44 gradient-boosted models for predicting delayed wound healing after radical gastrectomy exemplify both the promise and the translation gap.Importantly, direct evidence that ML-based CDSS improves patient outcomes in gastric cancer surgery remains limited or absent, and the feasibility of the proposed framework is currently supported mainly by indirect evidence from other clinical settings.By bridging prediction to intervention, the field can move beyond model-centric performance metrics toward patient-centered outcome improvement in gastric cancer surgery.