Learning Clinical Workflows to Identify Subgroups of Heart Failure Patients

Heart Failure (HF) is one of the most common indications for readmission to the hospital among elderly patients. This is due to the progressive nature of the disease, as well as its association with complex comorbidities (e.g., anemia, chronic kidney disease, chronic obstructive pulmonary disease, hyper- and hypothyroidism), which contribute to increased morbidity and mortality, as well as a reduced quality of life. Healthcare organizations (HCOs) have established diverse treatment plans for HF patients, but such routines are not always formalized and may, in fact, arise organically as a patient's management evolves over time. This investigation was motivated by the hypothesis that patients associated with a certain subgroup of HF should follow a similar workflow that, once made explicit, could be leveraged by an HCO to more effectively allocate resources and manage HF patients. Thus, in this paper, we introduce a method to identify subgroups of HF through a similarity analysis of event sequences documented in the clinical setting. Specifically, we 1) structure event sequences for HF patients based on the patterns of electronic medical record (EMR) system utilization, 2) identify subgroups of HF patients by applying a k-means clustering algorithm on utilization patterns, 3) learn clinical workflows for each subgroup, and 4) label each subgroup with diagnosis and procedure codes that are distinguishing in the set of all subgroups. To demonstrate its potential, we applied our method to EMR event logs for 785 HF inpatient stays over a 4 month period at a large academic medical center. Our method identified 8 subgroups of HF, each of which was found to associate with a canonical workflow inferred through an inductive mining algorithm. Each subgroup was further confirmed to be affiliated with specific comorbidities, such as hyperthyroidism and hypothyroidism.