Machine Learning Methods to Predict Telehealth Utilization

This study develops machine learning methods for predictive modeling of telehealth utilization. We gathered publicly available data from the 2021 National Health Interview Survey. Telehealth utilization was assessed with a binary variable based upon whether the patient had at least one virtual appointment with a clinical provider, either by phone or video, within the previous 12 months. We also selected a range of variables as predictive factors, including demographic variables, health conditions, healthcare utilization, measures of poverty and social needs, and challenges of ability. We considered models based upon logistic regression, support vector machines, extreme gradient boosting, and deep learning techniques. Tuning experiments were used to compare a range of parameters. Then we compared the best model from each category in terms of its predictive accuracy and other associated measures. The data included 27,626 records with complete measurements in all of the selected variables. We divided these data into a training set consisting of 80% of the data and a testing set with the remaining 20%. In the training data, 39.1% of the patients had utilized telehealth within the previous 12 months. In terms of classification accuracy on the testing data based on a 0.5 threshold, extreme gradient boosting (69.6%) and logistic regression (69.3%) had similar rates, with modest improvements over support vector machines (68.5%) and deep learning (68.1%).