Exploring engagement patterns within a mobile health intervention for women at risk of gestational diabetes

Abstract

Background: Gestational diabetes mellitus poses a significant global health concern during pregnancy, with behaviour change interventions offering effective risk reduction.

Objectives: Understanding diverse engagement patterns of pregnant women within mobile health (mHealth) interventions is vital for personalised healthcare. Tailoring interventions based on participant engagement types can enhance program effectiveness. This study aimed to explore engagement patterns among pregnant women at risk of gestational diabetes using the Liva app.

Design: This retrospective study serves as a secondary analysis of a randomised controlled trial, focusing on engagement patterns among participants in the intervention arm who received digital health coaching. The intervention group comprised participants enrolled in the Liva app, receiving mHealth lifestyle coaching. Our analysis concentrated on app usage data from 328 participants within the intervention group during the first phase of the study.

Methods: Principal component analysis reduced data to two dimensions, revealing principal components (PCs). A Gaussian mixture model clustered participants into distinct engagement patterns.

Results: Analysis of data from 328 pregnant women using the Liva app identified 3 distinct engagement clusters: Cluster 1, "Averagers"; Cluster 2, "Goalers"; and Cluster 3, "Immersers." These clusters correlated with two PCs. "Averagers" engaged moderately with both "Coach Features" and "Goal Features." "Goalers" predominantly used "Goal Features," while "Immersers" engaged with both "Coach Features" and "Goal Features." Notably, 82% of participants fell into the "Averagers" category.

Conclusion: This study reveals that individuals, despite similar program participation under uniform conditions, engage with the program differently. Understanding these differences is essential to provide personalised support during pregnancy and has implications for tailored medicine, digital health, and intervention development. Further research is needed to validate these findings across diverse healthcare settings, exploring engagement patterns throughout different pregnancy phases and their impact on health outcomes.

Keywords: algorithms; cluster analysis; gestational diabetes; health behaviour; machine learning; mobile applications; pregnancy; pregnant women; principal component analysis.

Figure 1.

Depiction of what the participant see navigating the Liva app. The first screen shows the health coach booking tool, the middle screen is the personalised health behaviour and goal plan, and the final screen is the goal registration and tracking.

Figure 2.

Histogram of our test data consisting of three normal distributions with depiction of how the parameters p_μ, marked in pink, and p_σ, marked in gray, control our test data and level of similarity between the distributions (and thereby number of clusters). When changing μ the distribution is pulled apart into two groups and when changing σ we change the width of the distribution. Thus, changing these parameters should change the number of clusters in a way that we can predict. When similar, as in this figure, we expect one cluster, but this will be broken up to two clusters when increasing our control parameters. We use this data for testing the optimal clustering algorithm to use for this analysis and to test the validity of our clustering results.

Figure 3.

Left: Plot of test data histogram. Right: GMM clustering results with number of clusters depicted by the AIC/BIC score. (a) testing for similar distributions under the standard setting which should then provide one single cluster and (b) testing for separated distributions which should thus provide two distinct clusters. GMM: Gaussian mixture model; AIC: Akaike information criterion; BIC: Bayesian information criteria.

Figure 4.

Left: Plot of test data histogram. Right: Results of GMM versus K-means clustering. K-means clustering with K = 3 as suggested to be the optimal number of clusters by the silhouette score and elbow method. However, our data consist of three very similar Gaussian distributions why we expect only one cluster. Here, the disadvantage of the Voronoi iteration algorithm underlying K-means becomes very clear as we see the equally sized Voronoi cells. Therefore, K-means is not used as the clustering algorithm in this study. GMM: Gaussian mixture model.

Figure 5.

Representation of our final PCs and their corresponding variable weights. It illustrates the degree to which each variable is associated with the underlying variance of each PC. The maximum possible correlation is denoted as |1|. Red indicates positive correlation and green indicates negative correlation. PCs: principal components.

Figure 6.

Projection of our data (328 × 13) following PCA dimensionality reduction to two PCs, denoted as “Coach Engagement” and “Goal Engagement.” The scatter plot illustrates the varying degrees of interaction exhibited by each participant with these two thematic components. PCA: principal component analysis; PC: principal component.

Figure 7.

The final choice of the clustering model based on our AIC/BIC scores and the likelihood-ratio test. The optimal clustering model is the tied model with K = 3 clusters. AIC: Akaike information criterion; BIC: Bayesian information criterion.

Figure 8.

Our final three clusters (shown in Figure 7) reshaped into one dimension and presented as normal distributions and fitted with a Gaussian function.