Predictive Analytics in Obstructed Colon Cancer: A Comparative Narrative Review of Clinical and AI-Based Models

This is the authors' abstract. We don't add key points for this paper.

Malignant large-bowel obstruction (LBO) occurs in 8 to 15% of colorectal cancer cases, and it is linked to high rates of complications and death. Identifying patient risk before surgery is essential to choose between emergency surgery and stenting. To predict outcomes in patients with obstructed colon cancer by comparing traditional clinical scoring systems (such as the American Society of Anesthesiologists [ASA] Physical Status Classification System, the Physiological and Operative Severity Score for Enumeration of Mortality and Morbidity [POSSUM], and the Association of Coloproctology of Great Britain & Ireland [ACPGBI] score) with modern artificial intelligence (AI) and machine learning (ML) models (such as Random Forest [RF], K-Nearest Neighbors [KNN], Extreme Gradient Boosting [XGBoost], and nomograms). We conducted a narrative review of peer-reviewed articles on prognostic models for malignant LBO. For each study, we recorded the variables used by the model, how user-friendly and interpretable it was, its performance metrics (such as the area under the receiver operating characteristic [AUROC] curve and accuracy), and its potential for clinical use. We organized the findings into five summary tables, comparing conventional risk scores and AI-based approaches in multiple dimensions. The focus is on practical, clinically-relevant insights rather than detailed algorithmic explanations. Traditional scores rely on a fixed set of clinical and operative variables. While easy to calculate and widely understood, they perform suboptimally in emergent LBO, with the Colorectal POSSUM (CR-POSSUM) showing an AUROC of approximately 0.65, and the ACPGBI reaching an AUROC of approximately 0.80 in elective settings. Conversely, AI models that leverage multiple perioperative inputs achieve higher accuracy: a Random Forest model yielded an AUROC of approximately 0.79 (95%CI: 0.71–0.87) on training and ranging from 0.75 to 0.82 on validation; a KNN model recorded approximately 88% of accuracy (AUROC: ∼ 0.77); and logistic regression nomograms attained AUC values near 0.84 for specific outcomes. Models based on AI generally outperform traditional risk scores in predicting short-term outcomes of malignant bowel obstruction, but they require high-quality electronic data and technological infrastructure. Traditional scores offer ease-of-use and interpretability, but they are less accurate in this setting. Clinicians should consider hybrid strategies, such as using familiar scores (ASA, the Charlson Comorbidity Index [CCI]) for quick triage, and AI/nomogram tools for detailed risk when data is available. Future work must address implementation barriers (such as data integration, model explainability, external validation, and prospective impact studies) to translate these tools into real-world perioperative decision-making.

AI-based predictive models offer superior accuracy over traditional clinical scores for predicting outcomes in obstructed colon cancer, though implementation barriers remain.