Machine-learning prediction of 30-, 90-, and 180-day readmissions in head and neck cancer using a national readmission database.

Abstract

e18089 Background: Readmissions impose major burdens on head and neck cancer (HNC) care. We employed the national readmission database(NRD) to develop and validate an XGBoost model for 30-, 90-, and 180-day readmissions. Training data encompassed 2016–2019 (n=229,736) HNC admissions; testing used 2020 data (n=57,201). We hypothesized that a machine learning analysis incorporating a wide range of features would yield superior predictive performance. Methods: We used 249 variables—219 disease-related (ICD-10 codes spanning infections, hematologic, neurologic/psychiatric, cardiovascular, pulmonary, gastrointestinal, musculoskeletal, genitourinary, signs/symptoms, external causes, endocrine/nutritional/metabolic, treatment-related, and various malignancies) and 30 non–disease-related (drawn from NRD administrative data: demographics, admission details, discharge outcomes, hospital characteristics, utilization/cost, classification codes). An XGBoost model was then trained to predict readmission at 30, 90, or 180 days post-discharge, with performance assessed by AUC and Brier score on the 2020 test set. Results: Model performance metrics and the top five most important variables for disease-related and non–disease-related categories are summarized in the table below. Notably, Z93 (artificial openings) emerged as a consistently strong predictor. Y83 (previous surgical procedures) and Z51.11 (chemotherapy) also repeatedly demonstrated strong predictive value. Among non–disease-related factors, elective admission status and diagnosis-related group-based variables (e.g., DRG_NoPOA, APRDRG) consistently emerged as top contributors across all time windows. Conclusions: This machine-learning model, trained on a large NRD cohort, demonstrates improved AUC (0.72–0.75) compared to conventional readmission risk models. Moreover, including variables typically overlooked in traditional models further enhanced accuracy. Future research will include prospective validation and real-time integration to reduce readmissions and enhance patient outcomes. Time Window AUC Brier Score Top 5 Important Variables (Disease-Related) Top 5 Important Variables (Non–Disease-Related) 30-Day 0.7246 0.0714 Z93 (Artificial openings), Y83 (Surgical procedure), Z51.11 (Chemotherapy), Z92 (Personal history of medical treatment), E86 (Volume depletion) ELECTIVE (admission type), DISPUNIFORM (discharge status), DRG_NoPOA (DRG without present-on-admission), APRDRG (All patient refined DRG), MDC (Major diagnostic category) 90-Day 0.7328 0.1038 Z93, Y83, E86, E43 (Severe protein-calorie malnutrition), Z51.11 ELECTIVE, DRG_NoPOA, DMONTH (Month of discharge), MDC_NoPOA (MDC without present-on-admission), APRDRG 180-Day 0.7468 0.1062 Z93, Y83, Z51.11, Metastatic cancer, E43 ELECTIVE, DMONTH, DRG_NoPOA, MDC_NoPOA, APRDRG