A Machine Learning Readmission Risk Prediction Model for Classical and Malignant Hematologic Disease

Introduction:

Classical and malignant hematologic diseases represent three of the top 20 highest adult readmission rate diagnoses in the Healthcare Cost and Utilization Project (HCUP)'s Nationwide Readmissions Database (NRD). Sickle cell trait/anemia is the highest on this list at 37.1% and represents over 2.5 million annual 30-day readmissions amongst Medicare and Medicaid patients alone. Additionally, the average cost of readmission for blood-based diseases is higher than that of the index admission at $16,900 vs $11,800 [HCUP Statistical Briefs 304 & 307].

Several machine learning readmission models have been developed successfully for prediction of sickle cell [Patel et al. 2019 - Blood; Patel et al. 2021 - British Journal of Haematology] and hematologic cancer-related readmission [Wong et al. 2020 - Cancer Research], but to date and the best of our knowledge, these models are diagnosis-limited with minimal input features. At least one study [Romero-Brufau et al. 2020 - Applied Clinical Informatics] has conducted a pilot implementation of a general readmission risk prediction model, demonstrating a 25% relative reduction in readmission rate at several community hospitals. Taken together, a high-performing model that predicts readmission risk across the spectrum of hematologic conditions may reduce cost and improve care through early detection of factors that may lead to readmission. The aim of this study was to create a high performing 30-day unplanned hospital readmission/emergency department (ED) visit or death risk prediction model.

Methods:

Three machine learning models were derived from a cohort of acute classical and malignant hematologic inpatient admissions during 2010-2023 within ten integrated Midwestern hospitals. Encounters were identified using HCUP's Clinical Classifications Software Redefined (CCSR) categories, classified into primary categories by organ system and subcategories by underlying physiologic process. Planned readmissions, in-hospital deaths, and discharges to hospice were excluded. A total of 1,280 clinical and non-clinical predictors were included, and the primary composite outcome was 30-day unplanned hospital readmission/ED visit or death. Model performance was evaluated by common metrics such as the area under the receiver operating curve (AUROC) and under the precision-recall curve (PR-AUC), measures of validity such as sensitivity, and a decision analysis curve.

Results:

The primary composite outcome rate was 36.2% amongst 18,096 encounters spanning 11,857 patients, of which the majority were Caucasian (72.2%) with a mean age of 56.6 years old (± 19.8) who had been discharged to home (86.2%). Nutritional anemia, sickle cell trait/anemia, and Non-Hodgkin lymphoma were the most common readmission diagnoses. The extreme gradient boosting (XGBoost) model had the highest overall performance with an AUROC of 0.75 and PR-AUC of 0.65. XGBoost demonstrated a net benefit relative to usual clinical practice by decision curve analysis.

Conclusion:

We successfully built a 30-day classical and malignant hematologic readmission or death risk prediction model that demonstrated a net benefit relative to current practice. Future directions include external model validation and prospective internal implementation with a special focus on analysis of equity performance.

No relevant conflicts of interest to declare.