PP168 - PREDICTION OF NUTRITIONAL RISK AND POSTOPERATIVE OUTCOMES USING MACHINE LEARNING IN SURGICAL PATIENTS

PP168

PREDICTION OF NUTRITIONAL RISK AND POSTOPERATIVE OUTCOMES USING MACHINE LEARNING IN SURGICAL PATIENTS

A. E. Gungor^1,*, I. Alparslan², P. Tasar³, S. Kilicturgay³

¹Nutrition and Dietetics, European University of Lefke, Lefke, ²Food Metabolism and Clinical Nutrition, Ankara University Institute of Health Sciences, Ankara, ³General Surgery, Uludag University School of Medicine, Bursa, Türkiye

Rationale: Malnutrition screening is recommended in surgical patients; however, its routine implementation remains inconsistent. The study investigated whether nutritional risk and its association with postoperative outcomes can be predicted using routinely available clinical data.

Methods: This retrospective study used electronic health records of 2276 patients admitted to a general surgery department. Demographic, clinical, anthropometric, and laboratory data were included. Nutritional risk was defined as NRS-2002 ≥3. Random Forest, XGBoost, C5.0, and decision tree algorithms were applied under different modelling scenarios. The dataset was randomly split into 70% training and 30% test sets. Model performance was evaluated using accuracy, balanced accuracy, sensitivity, specificity, F1-score, and ROC-AUC.

Results: Among patients classified as at nutritional risk (NRS-2002 ≥3), anthropometric variables, particularly >10% weight loss and low body mass index (BMI), were the strongest predictors across all models. The XGBoost model achieved an accuracy of 0.996, balanced accuracy of 0.986, and ROC-AUC of 0.986. Exclusion of these variables resulted in a marked decline in performance (balanced accuracy ~0.59–0.61). However, these anthropometric variables did not demonstrate similar predictive performance for postoperative outcomes. In contrast, postoperative outcomes were more closely associated with clinical variables reflecting disease severity.

Conclusion: Machine learning models accurately predict NRS-2002–defined nutritional risk primarily through anthropometric variables, whereas postoperative outcomes are better predicted by clinical variables reflecting disease severity. This dissociation suggests that nutritional risk and postoperative outcomes may reflect different underlying dimensions of risk.

Disclosure of Interest: None declared