Emulating Human Decision-Making Under Multiple Constraints

Farzin Ahmadi, Tinglong Dai, and Kimia Ghobadi (Johns Hopkins University)

Abstract: In many real-world environments, the details of decision-making processes are not fully known, e.g., how oncologists decide on specific radiation therapy treatment plans for cancer patients, how clinicians decide on medication dosages for different patients, or how hypertension patients choose their diet to control their illness. While conventional machine learning and statistical methods can be used to better understand such processes, they often fail to provide meaningful insights into the unknown parameters when the problem's setting is heavily constrained. Similarly, conventional constrained inference models, such as inverse optimization, are not well equipped for data-driven problems. In this study, we develop a novel methodology (called MLIO) that combines machine learning and inverse optimization techniques to recover the utility functions of a black-box decision-making process. Our method can be applied to settings where different types of data are required to capture the problem. MLIO is specifically developed with data-intensive medical decision-making environments in mind. We evaluate our approach in the context of personalized diet recommendations for patients, building on a large dataset of historical daily food intakes of patients from NHANES. MLIO considers these prior dietary behaviors in addition to complementary data (e.g., demographics and preexisting conditions) to recover the underlying criteria that the patients had in mind when deciding on their food choices. Once the underlying criteria are known, an optimization model can be used to find personalized diet recommendations that adhere to patients' behavior while meeting all required dietary constraints.