Machine learning identifies a COVID-19-specific phenotype in university students using a mental health app

Abstract

Background: Advances in smartphone technology have allowed people to access mental healthcare via digital apps from wherever and whenever they choose. University students experience a high burden of mental health concerns. Although these apps improve mental health symptoms, user engagement has remained low. Studies have shown that users can be subgrouped based on unique characteristics that just-in-time adaptive interventions (JITAIs) can use to improve engagement. To date, however, no studies have examined the effect of the COVID-19 pandemic on these subgroups.

Objective: Here, we sought to examine user subgroup characteristics across three COVID-19-specific timepoints: during lockdown, immediately following lockdown, and three months after lockdown ended.

Methods: To do this, we used a two-step machine learning approach combining unsupervised and supervised machine learning.

Results: We demonstrate that there are three unique subgroups of university students who access mental health apps. Two of these, with either higher or lower mental well-being, were defined by characteristics that were stable across COVID-19 timepoints. The third, situational well-being, had characteristics that were timepoint-dependent, suggesting that they are highly influenced by traumatic stressors and stressful situations. This subgroup also showed feelings and behaviours consistent with burnout.

Conclusions: Overall, our findings clearly suggest that user subgroups are unique: they have different characteristics and therefore likely have different mental healthcare goals. Our findings also highlight the importance of including questions and additional interventions targeting traumatic stress(ors), reason(s) for use, and burnout in JITAI-style mental health apps to improve engagement.

Keywords: Artificial intelligence; Machine learning; Mental health; Mobile applications; University students.

Fig. 1

Summary of the method, statistical, and machine learning approaches used in the present study.

Fig. 2

The optimal number of k-means clusters for each timepoint was determined using a Silhouette score. (A) lockdown, (B) post-lockdown, and (C) normal.

Fig. 3

K-means cluster characterization of mental healthcare app users across three timepoints by principal component analysis (PCA): (A) lockdown, (B) post-lockdown, and (C) normal.

Fig. 4

Plot showing the change in negative predictive value (NPV) based on the results of the recursive feature elimination (RFE). Here, we used RFE to estimate the contribution of features to the classification and regression trees (CART) characterizing cluster 3.

Fig. 5

Analysis of cluster 3 characteristics across the three COVID-19 timepoints. (A) Principal component analysis (PCA) dot plot. (B) Cluster dendrogram.