Discovering Engagement Personas in a Digital Diabetes Prevention Program

Abstract

Digital health technologies are shaping the future of preventive health care. We present a quantitative approach for discovering and characterizing engagement personas: longitudinal engagement patterns in a fully digital diabetes prevention program. We used a two-step approach to discovering engagement personas among n = 1613 users: (1) A univariate clustering method using two unsupervised k-means clustering algorithms on app- and program-feature use separately and (2) A bivariate clustering method that involved comparing cluster labels for each member across app- and program-feature univariate clusters. The univariate analyses revealed five app-feature clusters and four program-feature clusters. The bivariate analysis revealed five unique combinations of these clusters, called engagement personas, which represented 76% of users. These engagement personas differed in both member demographics and weight loss. Exploring engagement personas is beneficial to inform strategies for personalizing the program experience and optimizing engagement in a variety of digital health interventions.

Keywords: behavior change; clustering; digital health; mHealth; type 2 diabetes.

Figure 1

(a) Five concatenated univariate app-feature clusters showing log-scaled days per week (y-axis) with activity for coaching exchanges, weigh-ins, and meal logging over 13 weeks in the digital DPP (x-axis). Key (top left panel): Cluster centroids shown in red lines and sample average in blue, dotted lines. (b) Four program-feature clusters showing daily check-ins (y-axis) with educational lessons over the 90 days (x-axis) in the digital DPP. Key (top left panel): Cluster centroids shown in red lines and sample average in blue, dotted lines.

Figure 2

The number of engagement personas (solid, black line; left y-axis) for each cross-referencing threshold (x-axis) and the corresponding average silhouette coefficient (gray, dashed line; right y-axis). We selected 60% based on obtaining the highest silhouette coefficient that captured a large percentage (76%) of members. Lowering to 40%, for example, only increased the percentage of members to 80% but would have added an additional engagement persona with a small n that overfitted the data. In contrast, raising it to 70% would have dropped the percentage of captured members to 52%.

Figure 3

Violin plots of silhouette coefficient distributions with median (interquartile range) for the four program-feature clusters (left) and five app-feature clusters (right). The median silhouette coefficients (longer, dashed lines) for members with an identified engagement persona were higher than the medians for members without an identified engagement persona for all but program-feature cluster 4 and app-feature cluster 4. Mann–Whitney U tests between members with vs. without an identified engagement persona, * p ≤ 0.05; ** p ≤ 0.001; *** p ≤ 0.0001.

Figure 4

Kernel density estimation plot of the bivariate distribution of silhouette coefficients for each engagement persona. Members without an identified engagement persona (brown) had lower silhouette coefficients than those with an identified engagement persona.