A Machine Learning Approach to Characterizing Long-Term Care Clients Affected by COVID-19 Restrictions - The Case Study of Hospitalization

The COVID-19 pandemic had a high impact on society and healthcare systems. Since healthcare data during COVID-19 accumulated, several studies have reported reductions in hospitalizations and emergency room visits during the period of restrictions. Analyses of individual healthcare utilization could help explain the phenomenon. In this study, we developed a method to training machine learning model for predicting the probability of a decline in hospital use of long-term-care clients due to COVID-19 restrictions. Training data of the model was formed by calculating individual COVID-19 restriction effects by the difference-in-differences method from restricted and control client groups. The characteristics of the clients whose hospital use decreased due to the restrictions were analyzed by model interpretation methods such as Shapley values, partial dependence plots and surrogate models. The discrimination performance of the model was fair (AUROC = 0.763, 95% CI 0.701 − 0.819). Based on the model interpretation, the clients whose hospital visits decreased due to the COVID-19 restrictions had, on average, better functional condition (better ADL score, lower BMI) but higher number of comorbidities, and they used more drugs compared to the clients whose hospital use did not decrease.