Estimating treatment effects of wearable devices using uplift modeling

Abstract

We estimate the individual treatment effect of wearing a wearable device on HbA1c in order to support decision making for health care and health management using a wearable device. In general, since it is not possible to obtain both wearing and non-wearing data for each individual, we use the method of Uplift Modeling in Counterfactual Machine Learning. In the backtesting using Qini curve, we confirmed that our model can extract the population with higher treatment effect more efficiently than the baseline using random values or HbA1c itself. We also confirmed that feature selection and hyper-parameter tuning are effective as in the usual machine learning methods.