Multimodal ML models using CT body composition to predict colorectal surgery complications

Background

People with colorectal cancer (CRC) often have sarcopenic obesity and malnutrition which are associated with poorer surgical outcomes¹. Moreover, this hidden malnutrition is often missed during standard screening¹. Body composition analysis (BCA) derived from preoperative CT may offer objective markers of muscle loss and fat distribution¹.

Methods

The authors reviewed machine learning (ML) models that combine BCA with clinical variables to predict complications after CRC surgery¹. Embase, PubMed, the Cochrane Library and Web of Science were searched from database inception until April 13, 2026¹. Study quality was appraised with PROBAST, and a random-effects model was used to pool the area under the curve (AUC)¹. Subgroup analyses were also performed¹. The review was registered as CRD420261364363¹.

Results

• Studies: eight, with n = 7,835¹.
• Overall: pooled AUC 0.84 (95% CI: 0.76-0.89), reflecting good discrimination, with considerable heterogeneity (I2 = 79.7%)¹.
• Nomograms: pooled AUC 0.87 (95% CI: 0.78-0.92)¹.
• Anastomotic leakage: pooled AUC 0.83 (95% CI: 0.35-0.98). This figure was imprecise, as only two studies informed it¹.

Authors' conclusions

The authors consider multimodal ML models incorporating BCA biomarkers a robust, data-driven approach to stratifying risk before CRC surgery¹. The recognition of high-risk body composition phenotypes could support individualised prehabilitation and targeted immunonutrition¹. Standardised BCA and external validation across multiple centres is warranted¹.