Abstract
PURPOSE: Develop a cancer-specific machine learning (ML) model that accurately predicts 30-day unplanned readmissions in patients with solid tumors. METHODS: The initial cohort included patients 18 years or older diagnosed with a solid tumor. Two distinct cohorts were generated: one with and one without detailed social determinants of health (SDOHs) data. For each cohort, data were temporally partitioned in 70% (training), 20% (validation), and 10% (testing). Tree-based ML models were developed and validated on each cohort. The metrics used to evaluate the model's performance were receiver operating characteristic curve (ROC), area under the ROC curve, precision, recall (R), accuracy, and area under the precision-recall curve. RESULTS: We included 13,717 patients in this study in two cohorts (5,059 without SDOH data and 8,658 with SDOH data). Unplanned 30-day readmission occurred in 21.3% of the cases overall. The five main non-SDOH factors most highly associated with an unplanned 30-day readmission (R, 0.74; IQR, 0.58-0.76) were: number of previous unplanned readmissions; higher Charlson comorbidity score; nonelective index admission; discharge to anywhere other than home, hospice, or nursing facility; and higher anion gap during the admission. Neighborhood crime index, neighborhood median home values, annual income, neighborhood median household income, and wealth index were the main five SDOH factors important for predicting a high risk for an unplanned hospital readmission (R, 0.66; IQR, 0.56-0.72). The models were not directly comparable. CONCLUSION: Key drivers of unplanned readmissions in patients with cancer are complex and involve both clinical factors and SDOH. We developed a cancer-specific ML model that with reasonable accuracy identified patients with cancer at high risk for an unplanned hospital readmission.