Predictive value of multimodal models combining CT body composition and clinical variables for short-term postoperative complications in colorectal cancer: a systematic review and meta-analysis.

Abstract

Background: Malnutrition and sarcopenic obesity are prevalent in colorectal cancer (CRC) and linked to adverse surgical outcomes. Conventional screening often overlooks this occult malnutrition. Preoperative computed tomography (CT)-based body composition analysis (BCA) provides objective biomarkers for muscle depletion and fat distribution. This systematic review evaluates multimodal machine learning (ML) models integrating BCA and clinical variables to predict CRC postoperative complications and inform precision nutrition strategies.

Methods: We systematically searched PubMed, Embase, Web of Science, and Cochrane Library (inception to April 13, 2026) for ML models predicting postoperative complications in CRC using BCA. Quality was assessed using PROBAST. Pooled area under the curve (AUC) was estimated via a random-effects model, alongside subgroup analyses.

Results: Eight studies (n = 7,835) were included. The pooled AUC was 0.84 (95% CI: 0.76-0.89), indicating good discriminative performance, although substantial heterogeneity was present (I ² = 79.7%). Nomogram-based models showed a pooled AUC of 0.87 (95% CI: 0.78-0.92). For anastomotic leakage, the pooled AUC was 0.83 (95% CI: 0.35-0.98), although this estimate was imprecise because only two studies contributed to this subgroup.

Conclusion: Multimodal ML models integrating BCA biomarkers offer a robust, data-driven framework for preoperative risk stratification in CRC. Identifying high-risk body composition phenotypes may facilitate tailored prehabilitation and targeted immunonutrition strategies. Future efforts must prioritize standardized BCA and multi-center external validation.

Systematic review registration: CRD420261364363.