Building a Time-Series Model to Predict Hospitalization Risks in Home Health Care: Insights Into Development, Accuracy, and Fairness.

OBJECTIVES

Home health care (HHC) serves more than 5 million older adults annually in the United States, aiming to prevent unnecessary hospitalizations and emergency department (ED) visits. Despite efforts, up to 25% of patients in HHC experience these adverse events. The underutilization of clinical notes, aggregated data approaches, and potential demographic biases have limited previous HHC risk prediction models. This study aimed to develop a time-series risk model to predict hospitalizations and ED visits in patients in HHC, examine model performance over various prediction windows, identify top predictive variables and map them to data standards, and assess model fairness across demographic subgroups.

SETTING AND PARTICIPANTS

A total of 27,222 HHC episodes between 2015 and 2017.

METHODS

The study used health care process modeling of electronic health records, including clinical notes processed with natural language processing techniques and Medicare claims data. A Light Gradient Boosting Machine algorithm was used to develop the risk prediction model, with performance evaluated using 5-fold cross-validation. Model fairness was assessed across gender, race/ethnicity, and socioeconomic subgroups.

RESULTS

The model achieved high predictive performance, with an F1 score of 0.84 for a 5-day prediction window. Twenty top predictive variables were identified, including novel indicators such as the length of nurse-patient visits and visit frequency. Eighty-five percent of these variables mapped completely to the US Core Data for Interoperability standard. Fairness assessment revealed performance disparities across demographic and socioeconomic groups, with lower model effectiveness for more historically underserved populations.

CONCLUSIONS AND IMPLICATIONS

This study developed a robust time-series risk model for predicting adverse events in patients in HHC, incorporating diverse data types and demonstrating high predictive accuracy. The findings highlight the importance of considering established and novel risk factors in HHC. Importantly, the observed performance disparities across subgroups emphasize the need for fairness adjustments to ensure equitable risk prediction across all patient populations.