A0287
Title: Bayesian multi-task learning for medicine recommendation based on online patient reviews
Authors: Xinlei Wang - University of Texas at Arlington (United States) [presenting]
Yichen Cheng - Georgia State University (United States)
Yusen Xia - Georgia State University (United States)
Abstract: A drug recommendation system that integrates information from both structured data (patient demographic information) and unstructured texts (patient reviews) is proposed. The core of the recommendation system is a multi-task learning model that predicts review ratings of several satisfaction-related measures for a given medicine, where related tasks can learn from each other when predicting. The learned models can then be applied to new patients for drug recommendation. This fundamentally differs from the widely used recommender systems in e-commerce, which do not work well for new customers (referred to as the cold-start problem). Both topic modelling and sentiment analysis are used to extract information from the review texts. The results indicate that the extracted topics help identify each drug's key features and can sometimes (but not always) help predict ratings. A variable selection component is incorporated in the model through Bayesian LASSO, which aims to filter out irrelevant features. The proposed method is effective even with a small sample size and few available features. The method is tested on two sets of drug reviews involving 17 depression or high-blood-pressure-related drugs in total, and the prediction performance is compared with existing benchmark models.
