Learning to Prescribe Interventions for Tuberculosis Patients Using Digital Adherence Data

摘要: Digital Adherence Technologies (DATs) are an increasingly popular method for verifying patient adherence to many medications. We analyze data from one city served by 99DOTS, a phone-call-based DAT deployed Tuberculosis (TB) treatment in India where nearly 3 million people afflicted with the disease each year. The contains 17,000 patients and 2.1M dose records. lay groundwork learning this real-world data, including avoiding effects of unobserved interventions training used machine learning. then construct deep model, demonstrate its interpretability, show how it can be adapted trained three different clinical scenarios better target improve care. In real-time risk prediction setting our model could proactively intervene 21% more before 76% missed doses than current heuristic baselines. For outcome prediction, performs 40% baseline methods, allowing cities resources clinics heavier burden at failure. Finally, we present case study demonstrating end-to-end decision focused achieve 15% solution quality example problem faced health workers.