Modeling Patient Visit Using Electronic Medical Records For Cost Profile Estimation

Estimating health care cost of patients provides promising opportunities for better management and treatment to medical providers and patients. Existing clinical approaches only focus on patient's demographics and historical diagnoses but ignore ample information from clinical records. In this paper, we formulate the problem of patient's cost profile estimation and use Electronic Medical Records (EMRs) to model patient visit for better estimating future health care cost. The performance of traditional learning based methods suffered from the sparseness and high dimensionality of EMR dataset. To address these challenges, we propose Patient Visit Probabilistic Generative Model (PVPGM) to describe a patient's historical visits in EMR. With the help of PVPGM, we can not only learn a latent patient condition in a low dimensional space from sparse and missing data but also hierarchically organize the high dimensional EMR features. The model finally estimates the patient's health care cost through combining the effects learned both from the latent patient condition and the generative process of medical procedure. We evaluate the proposed model on a large collection of real-world EMR dataset with 836,033 medical visits from over 50,000 patients. Experimental results demonstrate the effectiveness of our model.