npj Digital Medicine (Dec 2023)

Classifying and clustering mood disorder patients using smartphone data from a feasibility study

  • Carsten Langholm,
  • Scott Breitinger,
  • Lucy Gray,
  • Fernando Goes,
  • Alex Walker,
  • Ashley Xiong,
  • Cindy Stopel,
  • Peter Zandi,
  • Mark A. Frye,
  • John Torous

DOI
https://doi.org/10.1038/s41746-023-00977-7
Journal volume & issue
Vol. 6, no. 1
pp. 1 – 6

Abstract

Read online

Abstract Differentiating between bipolar disorder and major depressive disorder can be challenging for clinicians. The diagnostic process might benefit from new ways of monitoring the phenotypes of these disorders. Smartphone data might offer insight in this regard. Today, smartphones collect dense, multimodal data from which behavioral metrics can be derived. Distinct patterns in these metrics have the potential to differentiate the two conditions. To examine the feasibility of smartphone-based phenotyping, two study sites (Mayo Clinic, Johns Hopkins University) recruited patients with bipolar I disorder (BPI), bipolar II disorder (BPII), major depressive disorder (MDD), and undiagnosed controls for a 12-week observational study. On their smartphones, study participants used a digital phenotyping app (mindLAMP) for data collection. While in use, mindLAMP gathered real-time geolocation, accelerometer, and screen-state (on/off) data. mindLAMP was also used for EMA delivery. MindLAMP data was then used as input variables in binary classification, three-group k-nearest neighbors (KNN) classification, and k-means clustering. The best-performing binary classification model was able to classify patients as control or non-control with an AUC of 0.91 (random forest). The model that performed best at classifying patients as having MDD or bipolar I/II had an AUC of 0.62 (logistic regression). The k-means clustering model had a silhouette score of 0.46 and an ARI of 0.27. Results support the potential for digital phenotyping methods to cluster depression, bipolar disorder, and healthy controls. However, due to inconsistencies in accuracy, more data streams are required before these methods can be applied to clinical practice.