A Novel Approach to Passively Informed Mental Health Risk Detection in People with Diabetes: Rationale, Methods, and Early Results (Preprint)

BACKGROUND Proactive detection of mental health (MH) needs among people with diabetes mellitus (DM) could facilitate early intervention, improve overall health and quality of life, and reduce individual and societal health and economic burdens. Passive sensing and ecological momentary assessment are newer methods that may be leveraged for proactive detection. OBJECTIVE The primary aim of this study was to conceptualize, develop, and evaluate a novel machine learning approach for predicting MH risk in people with DM. METHODS A retrospective study was designed to develop and evaluate a machine learning model, utilizing data collected on 142,432 individuals with DM enrolled in the Livongo for Diabetes program. First, participants’ MH statuses were verified using prescription and medical and pharmacy claims data. Next, passive sensing signals were extracted from participant behavior in the program. Data sets were then assembled to create participant-period instances and descriptive analyses were conducted to understand the correlation between MH status and passive sensing signals. Passive sensing signals were then entered into the model to train and test its performance. The model was evaluated on seven measures: sensitivity, specificity, precision, area under the curve, F1 score, accuracy, and confusion matrix. Results: In training and 3 subsequent test sets, the model achieved a confidence score of greater than 0.5 in all but three measures. RESULTS In training and 3 subsequent test sets, the model achieved a confidence score of greater than 0.5 in all but three measures. CONCLUSIONS Results demonstrate the utility of a passively informed MH risk algorithm and invite further exploration.