Letter to editor on “Systematic review and meta-analysis of the role of machine learning in predicting postoperative complications following colorectal surgery: how far has machine learning come?”

To the Editor, Mohamedahmed et al provide a timely systematic review and meta-analysis of machine learning (ML) models for predicting postoperative complications after colorectal surgery, synthesizing 18 studies across multiple outcomes including anastomotic leak (AL), surgical site infection (SSI), ileus, mortality, prolonged length of stay (LOS), readmission and bleeding[1]. In a field where traditional linear scores such as POSSUM, P-POSSUM, and NELA offer only moderate discrimination, their pooled AUCs for AL (0.813), mortality (0.867), prolonged LOS (0.810), and SSI (0.802) suggest that ML has genuine potential to refine risk stratification beyond existing tools[1]. This work has several notable strengths. The authors follow PRISMA, AMSTAR-2, PROBAST, and TITAN 2025 guidance, use a registered PROSPERO protocol, and explicitly assess both risk of bias and applicability[1]. They also move beyond narrative description by pooling area under the receiver operating characteristic curve (AUCs) across outcome domains, providing the clearest quantitative benchmark to date for ML performance in colorectal surgery. Importantly, they show that ML-based mortality prediction outperforms widely used perioperative scores when compared with published benchmarks, underscoring the promise of non-linear models trained on large clinical datasets. Clinically, the review highlights where ML may be most impactful. For AL, models consistently identify tumor height, intraoperative blood loss, and early postoperative laboratory trends (creatinine, C-reactive protein) as key features, aligning with multifactorial leak biology. For SSI and ileus, operative approach, nutritional status, sarcopenia, opioid use, and electrolyte disturbances emerge as modifiable risk factors, supporting the concept of ML-enabled “early warning” for enhanced recovery deviations[1]. At the same time, several limitations rightly emphasized by the authors temper immediate adoption. Many models are single-center, retrospective, and trained on relatively small samples, with optimistic AUCs that may not generalize. Heterogeneity in algorithms, feature sets, outcome definitions, and validation strategies is substantial, and pooling AUCs across such diverse pipelines inevitably masks important differences in calibration and clinical utility[1]. Class imbalance for rare events (e.g. AL, mortality) and limited external validation remain major threats to robustness – challenges echoed across the wider surgical AI literature[2–4]. For colorectal surgeons, the key message is that ML-based tools should currently be viewed as adjuncts to, not replacements for, established scores and multidisciplinary judgment. The next research steps are clear: large, multicenter prospective cohorts; transparent reporting using frameworks such as TRIPOD-AI and TITAN; rigorous external validation; decision-curve and cost-effectiveness analyses; and, crucially, impact studies testing whether ML-guided care pathways actually change management and improve outcomes. Mohamedahmed et al should be congratulated for providing a high-quality evidence map of where ML in colorectal surgery stands today. Their review sets a useful benchmark – and an agenda – for moving from promising AUCs on retrospective data toward trustworthy, workflow-integrated decision support that demonstrably benefits patients. We prepared this commentary in accordance with the TITAN 2025 Guidelines to ensure transparent and responsible reporting[5].