Predicting Home Care Use After Assessment Using Multiple Machine Learning Methods

Abstract

This research is a comparative analysis of applying different machine-learning methods to health care data. The data used is from the interRAI home care assessment instrument, collected in central British Columbia, Canada. The primary dataset used contains more than 100,000 records each with 423 attributes. We built models for predicting home care usage in the three weeks following an assessment by applying different regression and classification machine learning algorithms. The main regression algorithms used in the process were multiple linear regression, lasso, ridge, decision tree and ensemble methods, with the last being the most promising. In the area of classification, KNN, logistic regression, decision tree and ensemble methods were used. Apart from the technical machine learning algorithms, both patient partners and health systems experts participated and provided feedback regarding home care practices and issues. These formed essential element in designing the research question, selecting variables, and improving the models. The highest accuracy achieved was 84.3% which was achieved through a random forest classifier and evaluated using K-fold cross validation.