Explainable machine-learning prediction of prolonged postoperative hospitalization after esophagectomy using clinical and CT-derived body-composition features.

Authors

Abstract

Prolonged postoperative hospitalization after esophagectomy reflects delayed recovery, postoperative morbidity, nutritional difficulty, discharge readiness, and healthcare resource utilization. Computed tomography (CT)-derived body-composition features may provide additional information about physiologic reserve, but their incremental value for predicting prolonged length of stay (LOS) after esophagectomy remains uncertain. This study aimed to develop and internally validate machine-learning models for predicting prolonged postoperative hospitalization after esophagectomy using routine clinical variables and CT-derived body-composition features. This retrospective prediction-model development and internal-validation study included 299 patients who underwent esophagectomy between December 2020 and June 2025. The primary outcome was prolonged postoperative LOS, defined as LOS >15 days. Three prespecified feature sets were evaluated: Clinical-only, CT-only, and Clinical + CT. CT-derived variables included average thoracic muscle, average thoracic fat, average lumbar muscle, and average lumbar fat. Logistic regression, least absolute shrinkage and selection operator (LASSO) logistic regression, random forest, and extreme gradient boosting (XGBoost) models were developed. Internal validation was performed using 1,000 bootstrap resampling iterations with out-of-bag evaluation. Model performance was assessed using discrimination, calibration, decision-curve analysis, and supplemental performance metrics. SHapley Additive exPlanations (SHAP) analysis was performed using the XGBoost Clinical + CT model to explain individual-level predictions. Among 299 patients, 70 patients experienced prolonged postoperative LOS. The Clinical + CT LASSO model achieved the highest internally validated discrimination among the LASSO models, with an area under the curve (AUC) of 0.6288, accuracy of 0.6121, sensitivity of 0.5207, specificity of 0.6408, and Brier score of 0.2422. The Clinical-only LASSO model showed similar performance, with an AUC of 0.6194, whereas the CT-only LASSO model showed limited discrimination, with an AUC of 0.5439. The incremental improvement in AUC after adding CT-derived features to clinical variables was small and statistically non-significant [AUC difference +0.0094; 95% confidence interval (CI): -0.1393 to 0.1403; P=0.44]. SHAP analysis identified average lumbar fat, white blood cell count, age, albumin, and average thoracic muscle as the most influential features in the XGBoost Clinical + CT model. Machine-learning models using routine clinical variables and CT-derived body-composition features demonstrated modest ability to predict prolonged postoperative hospitalization after esophagectomy. CT-derived features alone showed limited discrimination and provided only limited incremental value beyond clinical variables in the primary LASSO comparison. However, SHAP analysis of the tree-based model suggested that selected CT-derived features may contribute to nonlinear, individual-level risk estimation. Given the moderate discrimination and relatively high negative predictive value observed, these models should be considered supportive risk-stratification tools rather than replacements for clinical judgment. Their potential role may be in identifying patients less likely to experience prolonged hospitalization and supporting perioperative planning, although external validation and prospective evaluation are required before clinical implementation.

Tags

Topics