Learning Individualized Treatment Rules from Multi-Study and Multi-Site

This project integrates cutting-edge techniques from meta-analysis, federated learning, integrative learning, and transfer learning to learn a more reliable and robust individualized treatment rule.

We introduce a novel framework that combines data from multiple randomized controlled trials, each sharing a common treatment arm but differing in alternative treatments. The method adaptively weights information across studies using a regularized misclassification risk. Theoretical properties of the approach are explored, and simulations demonstrate improved estimation of value and benefit functions.

Currently, we are developing methods to integrate electronic health records from multiple hospitals, aiming to learn improved personalized recommendations for patients with type 1 diabetes. This approach has the potential to improve clinical decision-making and patient outcomes by tailoring treatments to individual characteristics.