Learning models for writing better doctor prescriptions

摘要

We develop a data-driven approach for learning and improving the prescription policy physicians use to treat Type 2 diabetes. Our model combines regression, classification and strategy optimization. We use regression algorithms to predict the outcomes of prescriptions, and then adopt a parameterized classification method to learn the physicians' prescription policy. Finally, we improve the prescription policy by optimizing over the parameters in the prescription policy model. Compared with the original prescription policy, patients who shift their treatment according to the recommended policy see significant blood glucose reduction on average. The proposed prescription recommendations offer a better therapeutic effect than the state-of-art deterministic algorithms. Our framework can also be applied to improving the prescription policy for other diseases.