Using random forests to model 90-day hometime in people with stroke

Abstract

Abstract Background Ninety-day hometime, the number of days a patient is living in the community in the first 90 after stroke, exhibits a non-normal bucket-shaped distribution, with lower and upper constraints making its analysis difficult. In this proof-of-concept study we evaluated the performance of random forests regression in the analysis of hometime. Methods Using administrative data we identified stroke hospitalizations between 2010 and 2017 in Ontario, Canada. We used random forests regression to predict 90-day hometime using 15 covariates. Model accuracy was determined using the r-squared statistic. Variable importance in prediction and the marginal effects of each covariate were explored. Results We identified 75,745 eligible patients. Median 90-day hometime was 59 days (Q1: 2, Q3: 83). Random forests predicted hometime with reasonable accuracy (adjusted r-squared 0.3462); no implausible values were predicted but extreme values were predicted with low accuracy. Frailty, stroke severity, and age exhibited inverse non-linear relationships with hometime and patients arriving by ambulance had less hometime than those who did not. Conclusions Random forests may be a useful method for analyzing 90-day hometime and capturing the complex non-linear relationships which exist between predictors and hometime. Future work should compare random forests to other models and focus on improving the accuracy of predictions of extreme values of hometime.

Keywords

• Stroke
• Hometime
• Random forests