PD1053 - MACHINE LEARNING–BASED PREDICTION OF POSTOPERATIVE CLINICAL OUTCOMES IN SURGICAL PATIENTS: INTEGRATION OF NUTRITIONAL RISK PARAMETERS IN ACCORDANCE WITH ESPEN GUIDELINES

Linked sessions

PD1053

MACHINE LEARNING–BASED PREDICTION OF POSTOPERATIVE CLINICAL OUTCOMES IN SURGICAL PATIENTS: INTEGRATION OF NUTRITIONAL RISK PARAMETERS IN ACCORDANCE WITH ESPEN GUIDELINES

G. Pasios1, A. Almperti2, D. Karayiannis2,*, F. Malamateniou1, E. Zoulias1

1Laboratory of Health Informatics, Department of Nursing, School of Health Sciences, National and Kapodistrian University of Athens, 2Clinical Nutrition, Evaggelismos General Hospital, Athens, Greece

 

Rationale: Disease-related malnutrition constitutes a significant clinical burden in surgical populations, independently associated with increased postoperative morbidity and prolonged hospitalisation. The potential of supervised machine learning algorithms to enhance early risk stratification by incorporating ESPEN-endorsed nutritional screening parameters warrants systematic investigation.

Methods: A retrospective cohort analysis was performed on 154 consecutive hospitalised surgical patients (80 male; mean age = 57.8 y) at a tertiary referral centre. Variables included demographics (age, sex), clinical data (diagnosis, comorbidities), anthropometrics (weight, height, BMI), biochemical markers (albumin, CRP, creatinine, urea, total proteins, haematocrit, sodium, potassium), nutritional risk indicators (NRI, weight loss, food intake reduction), surgical category, and feeding modality. Three supervised Machine Learning (ML) algorithms were applied: Decision Tree, Random Forest, and XGBoost. Class imbalance was addressed with SMOTE. Performance was evaluated using balanced accuracy, F1-score, and Cohen's kappa, with stratified 80/20 split and k-fold cross-validation.

Results: The Random Forest classifier demonstrated superior predictive performance (Balanced Accuracy: 0.84; F1-score: 0.79; κ: 0.54), followed by XGBoost (Balanced Accuracy: 0.79; F1-score: 0.78). BMI-derived nutritional indices consistently emerged as the most discriminative predictors across models, underscoring the clinical relevance of anthropometric nutritional assessment in perioperative risk stratification.

Conclusion: ML models integrating nutritional risk variables demonstrate robust predictive performance for clinical outcomes in surgical patients. These findings support ESPEN recommendations for systematic nutritional screening and early individualised intervention.

Disclosure of Interest: None declared