MACHINE LEARNING TO PREDICT HOMEBOUND STATUS IN OLDER ADULTS USING CANADIAN LONGITUDINAL STUDY ON AGING DATASET

Abstract

Abstract Individuals who are unable to leave their home or with great difficulty are considered homebound or semi-homebound. Homebound status is strongly associated with disability, social isolation, healthcare use and costs, and mortality. Most homebound older adults have multiple chronic conditions and poor health. There is not enough information about homebounded older adults in Canada. The Comprehensive cohort of the Canadian Longitudinal Study on Aging (CLSA) presents an excellent opportunity to study the complex factors associated with homebound status and the interplay between physical, social, psychological, and environmental determinants over time. This is a population-level study which makes use of provincial healthcare registration data to sample older adults across the country. We obtained the first wave of CLSA dataset containing samples from 21667 individuals over 3223 variables. We developed a definition of ‘homeboundedness’ in using life-space index variables present in the CLSA datatset. The dataset contained numerical, categorical and missing values. After preprocessing, we selected 1101 Homebound and 20521 Non-Homebound individuals, and 1771 variables. We showed that Random Forest classifier (with missing values) provided an AUC ROC and PR of 0.89 and 0.49. The missing value imputation did not improve the results significantly. Using feature hashing, we converted the dataset to numerical values; the AUCs improved to 0.96 and 0.71 at the cost of losing interpretation. In future, we will consult clinical experts in choosing the relevant features and use analytical methods to select features and compare. We will also test these predictive models on the next wave of this dataset.