Abstract- Web-based appointment systems areemerging in healthcare industry providing patients with convenient anddiversiform services, among which physician recommendation is becoming more andmore popular tool to make assignments of physicians to patients. Motivated by apopular physician recommendation application on a web-based appointment systemin China, this paper gives a pioneer work in modeling and solving the physicianrecommendation problem. The application delivers personalized recommendationsof physician assortments to patients with heterogeneous illness conditions, andthen, patients would select one physician for appointment according to theirpreferences. Capturing patient preferences is essential for physicianrecommendation delivery; however, it is also challenging due to the lack ofdata on patient preferences. In this project, we formulate the physicianrecommendation problem based on which the preference learning algorithm isproposed that optimizes the recommendations and learns patient preferences atthe same time. Since the illness conditions of patients are heterogeneous, thealgorithm aims to make personalized recommendation for each patient. Besidesdemonstrating the effectiveness of algorithm performance in terms of regretbound, we also provide extensive numerical experiments to show the expectedalgorithm performance under heterogeneous reward scenarios and performancecomparison with algorithms in the literature under fixed reward scenarios.
We introduce theflexibility of adjusting preference estimate update interval into our algorithmand conclude that short update interval contributes to short-term performancewhile long
update interval leads to good results in the longrun. Furthermore, we analyze how preference bound helps the algorithm to makeexplorations, which constitute two major contributions of our algorithm.Finally, we discuss the relevance between patient preferences and physicianutilization and present a utilization-balancing approach that is effective innumerical experiments.
Keyword— Dynamicpolicy, patient preference learning, physician recommendation