Using Mobile Daily Mood and Anxiety Self-ratings to Predict Depression Symptom Improvement

Abstract

Effective treatment of depression is important for the well-being of individuals and the overall health of the society. The current treatment approach calls for monitoring and assessing depression symptoms using self-administered or clinician-administered questionnaires, which are burdensome, costly, and may suffer from recall and desirability bias. In this study, we explore using daily 5-point Likert-scale mood and anxiety survey in place of burdensome clinical depression questionnaires for monitoring depression treatment. Specifically, we collect daily mood and anxiety surveys using a smartphone app, and use them to predict depression symptom improvement in a clinical depressed population. Using a dataset from 67 participants, we show that both mood and anxiety features obtained from the daily survey have significant correlation with clinical questionnaire scores. We then develop a family of machine learning models that use mood and anxiety features (separately or in combination) to predict symptom improvement status on a weekly basis. The best prediction F1score achieved by these models is 0.65. While this accuracy is lower than what is achieved by clinical questionnaires (best F1score being 0.71), daily survey is much less burdensome, and hence we believe that it provides a promising direction in monitoring depression symptom improvement. We further show that the prediction accuracy is not sensitive to missing data, allowing not very regular responses in practice. Last, we show that adding more historical data beyond the current week does not provide much benefits in improving prediction accuracy, and daily mood/anxiety self-ratings can predict improvement status accurately one or two weeks in the future.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 5 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.