PS 3-60 ANALYZING POPULATION PROGRESSION DYNAMICS OF DEPRESSION FROM ELECTRONIC HEALTH RECORD USING A FAMILY OF MARKOV MODELS

Tuesday, October 25, 2016
Bayshore Ballroom ABC, Lobby Level (Westin Bayshore Vancouver)
Poster Board # PS 3-60

Jiaqi Huang, MS, University of Washington, Seattle, WA and Shan Liu, PhD, Industrial and Systems Engineering, University of Washington, Seattle, WA

Purpose: Depression has a 10% prevalence and leads to poor health outcomes, high medical resource utilization and costs in the U.S. We aim to model and analyze the dynamics of population-level depression progression, and provide insights on the occurrence and duration of major depression.

Method: Data contain Patient Health Questionnaire (PHQ)-9 scores as measures of depression severity from Electronic Health Record. 610 patients with ongoing treatment within a 20 bi-week time-window were selected and their irregular PHQ-9 scores were transformed into disease trajectories by b-spline fitting. We performed k-means clustering to detect the heterogeneity in the population trajectory patterns and found five subgroups: Severe, Decreasing Severity, Moderate Fluctuation, Increasing Severity and Healthy. We assumed there are five depression states including healthy, mild, moderate, moderately severe and severe. We then applied Multi-State Discrete Time Markov model (MSM), Hidden Markov model (HMM), Semi-Markov model (semi-M) and Hidden Semi-Markov model (HSMM) to each of the five subgroups to model the depression progression dynamics and identify the characteristics of depression severity state transitions.

Result: We estimated subgroup-specific transition probability matrices generated from four Markov models and their corresponding stationary probability, expected first passage time and the proportion of time in a depression state where appropriate. Bootstrap was used in validation. To illustrate, Figure panel a) shows the stationary probabilities across all four Markov models in the Increasing Severity subgroup, where the first column indicates the starting depression severity state proportion in this group. Panel b) shows the long-term proportion of time in a depression state comparing the five subgroups from the semi-M model, and panel c) shows the expected number of transitions in going from a depression state to another state for the first time, defined as the expected first passage time in bi-weeks, from the MSM and HMM in the Increasing Severity Group.

Conclusion: Modeling methods from this study provided an understanding of depression progression in the American adult population undergoing treatment, and shed lights on justifying initiatives to design adaptive depression monitoring guidelines to patients' progression patterns.

             

Fig.1 Long-Term Outcome Measurements

H: Healthy State, Mi: Mild State, Mo: Moderate State, MS: Moderate Severe State, S: Severe State, In: Increasing Severity Group, De: Decreasing Severity Group, H: Healthy Group, MF: Moderate Fluctuation Group, Se: Severe Group