Machine learning and the nomogram as the accurate tools for predicting postoperative malnutrition risk in esophageal cancer patients.

Background: Postoperative malnutrition is a prevalent complication following esophageal cancer surgery, significantly impairing clinical recovery and long-term prognosis. This study aimed to develop and validate predictive models using machine learning algorithms and a nomogram to estimate the risk of malnutrition at 1 month after esophagectomy.

Methods: A total of 1,693 patients who underwent curative esophageal cancer surgery were analyzed, with 1,251 patients allocated to the development cohort and 442 to the validation cohort. Feature selection was performed via the least absolute shrinkage and selection operator (LASSO) algorithm. Eight machine learning models were constructed and evaluated, alongside a nomogram developed through multivariable logistic regression.

Results: The incidence of postoperative malnutrition was 45.4% (568/1,251) in the development cohort and 50.7% (224/442) in the validation cohort. Among machine learning models, the Random Forest (RF) model demonstrated optimal performance, achieving area under the receiver operating characteristic curve (AUC) values of 0.820 (95% CI: 0.796-0.845) and 0.805 (95% CI: 0.771-0.839) in the development and validation cohorts, respectively. The nomogram incorporated five clinically interpretable predictors: female gender, advanced age, low preoperative body mass index (BMI), neoadjuvant therapy history, and preoperative sarcopenia. It showed comparable discriminative ability, with AUCs of 0.801 (95% CI: 0.775-0.826) and 0.795 (95% CI: 0.764-0.828) in the respective cohorts ( > 0.05 vs. RF). Calibration curves revealed strong agreement between predicted and observed outcomes, while decision curve analysis (DCA) confirmed substantial clinical utility across risk thresholds.

Conclusion: Both machine learning and the nomogram provide accurate tools for predicting postoperative malnutrition risk in esophageal cancer patients. While RF showed marginally higher predictive performance, the nomogram offers superior clinical interpretability, making it a practical option for individualized risk stratification.