Abstract
OBJECTIVE: To develop and validate a machine learning model to identify patients at high risk of 30-day mortality and hospital readmission using routinely collected health care data. STUDY DESIGN: Prognostic predictive modeling and retrospective cohort study. The study was conducted in 2024 using data from 2006 to 2018, with at least a 30-day follow-up. SETTING: The 2006 to 2018 National Cancer Database (NCDB). METHODS: The study used deidentified NCDB data on 103,891 head and neck squamous cell carcinoma (HNSCC) patients who underwent surgical resection. Machine learning models were trained on 80% of the data, tested on the remaining 20%, and evaluated using the area under the curve (AUC) and SHapley Additive exPlanations (SHAP) analysis to identify key predictors for 30-day mortality and readmission. RESULTS: Among 103,891 patients, 5838 (5.6%) were readmitted, and 829 (0.8%) died within 30 days. The median age was 62, 69% male, and 89% white. Predictors included demographic and clinical data from the NCDB. Five machine learning models were combined and achieved an AUC of 0.80 (95% CI: 0.77-0.83) for mortality prediction and 0.67 (95% CI: 0.65-0.68) for readmission prediction. SHAP analysis identified sex and urban-rural index as key predictors of mortality and readmission, respectively. CONCLUSION: Machine learning models can accurately predict mortality and readmission risks, offering insights into the most influential factors. With further validation, these models may enhance clinical decision-making in postsurgical care for HNSCC patients.