Externally Validated Machine Learning Models for 30-/90-/180-Day Unplanned Readmission After Head and Neck Cancer Hospitalizations in the United States.

Abstract

Readmissions after head and neck cancer (HNC) hospitalizations are common and costly. We developed and externally validated machine learning (ML) models to predict unplanned readmissions across short- and longer-term horizons. Using the 2016-2020 Nationwide Readmissions Database, we included adult nonelective admissions with ≥1 malignant HNC diagnosis code. Models were trained on 2016-2019 discharges and externally tested on 2020 (N = 57,201). We engineered 247 discharge time predictors and trained multiple ML models, with thresholds selected by maximizing out-of-fold F1. We evaluated discrimination, calibration, and clinical utility via decision curve analysis and interpretability using Shapley additive explanations (SHAP). In external testing, XGBoost had the best discrimination (area under the receiver operating characteristic curve [AUC] 0.725/0.746/0.756 for 30/90/180-day readmission). Calibration was acceptable, and decision curve analysis showed net benefit over treat all/none across 10%-30% thresholds. Key predictors by SHAP included artificial airway/nutrition openings, discharge timing/disposition, and severity proxies. A ML model, specifically XGBoost trained on a large set of administrative and clinical data, can effectively predict both short- and long-term unplanned readmission risk in patients with HNC and may support targeted discharge planning to improve outcomes and reduce costs.