2017 HSR&D/QUERI National Conference

4068 — Latent Classes of High Risk Patients Receiving Care from VHA Patient-Centered Medical Home Model

Lead/Presenter: Jean Yoon
All Authors: Wong ES (VA Puget Sound Health Care System) Yoon J (VA Palo Alto Health Care System) Piegari RI (VHA Office of Informatics and Information Governance) Schwartz G (VHA Office of Informatics and Information Governance) Fihn SS (VHA Office of Informatics and Information Governance) Chang ET (VA Greater Los Angeles Health Care System)

Among the goals of the Patient Aligned Care Team initiative is development of population health tools to facilitate more informed delivery of primary care. This study sought to identify and characterize latent classes of high risk patients based on longitudinal patterns of risk scores using machine learning methods.

This was an observational study using VHA and Medicare administrative data. We identified 4,063 age ?65 patients hospitalized through VHA or Fee-For-Service Medicare, discharged over a 1-week period in November 2012 and had a risk score in the top 5th percentile. Weekly risk scores were derived from the validated Care Assessment Needs (CAN) Model. CAN scores represent 90-day hospitalization probability. Using patient CAN score trajectories over a 60-week follow-up period; we applied the non-parametric k-means algorithm to identify latent classes of patients. The k-means algorithm determines the optimal number of classes and a mean trajectory for each latent class. We then conducted descriptive analyses comparing socio-demographic characteristics, comorbidities and health service use of patients in respective latent classes.

At baseline, the mean CAN score was 29.6% (IQR = 18.2% to 38.3%). The best fitting model identified 2 latent classes. The first class is characterized by relatively lower and more variable CAN scores during follow-up with 63.2% of patients in this group. The remaining 36.8% of patients comprised the second class, characterized by persistently high CAN scores and fewer fluctuations over time. Compared to patients in the first latent class, patients with persistently high CAN scores were less likely to be married, had increased hospitalizations, and used more specialty and palliative care services. Furthermore, patients with persistently high CAN scores had higher rates of medical and psychiatric comorbidities at baseline, with the most substantial differences in heart failure, chronic obstructive pulmonary disease and renal failure.

Despite substantial heterogeneity in patient-level CAN score trajectories, we identified two important classes of high risk patients.

The identification of persistently high risk patients in this study may help providers identify potential targets for early intervention. Also, the latent classes identified may provide otherwise unmeasured information, which can help improve prediction of health and health care outcomes.