Description:

Please contact mcccenters@northwestern.edu for the event password

Patients with multiple chronic conditions, also known as multi-morbidity in the clinical literature, have a disproportionate impact on the U.S. healthcare system. According to a 2017 report by RAND, 19% of Americans (around 60 million individuals) had four or more chronic conditions, and they account for more than 50% of total healthcare expenditure. A baseline predictive model for the probabilities of co-occurring conditions is essential for quantifying epidemiological associations between condition groups, resource planning for targeted interventions, and driving decision support for personalized medicine. However, MCC patients exhibit significant heterogeneity in chronic condition combinations, and the number of individuals in a disease dataset is usually small compared to the number of possible disease combinations. Therefore, simple maximum-likelihood estimates of disease co-occurrence will erroneously assign zero probabilities to disease combinations that are missing from the dataset but are likely to occur in the larger population. In this work, we combine maximum-entropy optimization, data mining, and machine learning techniques to create an algorithm, called MaxEnt-MCC, for estimating the prevalence of chronic diseases in a population in the face of sparse data. In a case study using Medical Expenditure Panel Survey (MEPS) data, we show how MaxEnt-MCC can be used to predict previously unobserved but likely disease combinations, quantify associations between groups of chronic conditions, and estimate healthcare costs in a principled manner.