Commentary on: Objective smartphone data as a potential diagnostic marker of bipolar disorder

The lack of accurate, reliable and objective disease-related phenotypes is one of the most persistent challenges for psychiatry. Smartphone-based data have been put forward as tools for addressing this challenge both in research and in clinical care. Smartphones can provide ‘digital phenotypes’ defined as the ‘moment-by-moment quantification of the individual-level human phenotype in situ using data from personal digital devices’ (Torous et al., 2016). The use of smartphones has the advantage of high prevalence in the general population which will likely become ubiquitous in the foreseeable future (Statistica, 2018). Cost efficiency, richness of real-life information, granular temporal resolution and scalability add to the appeal (Torous et al., 2016). Ultimately, the usefulness of digital phenotypes is contingent on empirical assessments of their sensitivity, specificity and predictive value.

Faurholt-Jepsen et al. (2019) assessed the diagnostic usefulness of digital phenotypes in the context of bipolar disorder (BD). Their sample comprised patients with BD (n = 29) attending the Clinic for Affective Disorder at the Psychiatric Centre Copenhagen and psychiatrically healthy blood donors (n = 37) attending the Blood Bank at Rigshospitalet, Copenhagen University Hospital. Interestingly, 43.2% of patients and 46.6% of healthy individuals initially approached declined participation. Over the 12-week study period, all participants completed daily self-evaluations of manic and depressive symptoms while passive data were collected continuously with regard to number of calls and text messages, the duration of phone calls and the duration the smartphone screen was on. Compared to healthy volunteers, patients tended to make longer phone calls regardless of their mental state, texted more when in a manic or mixed state and interacted less with their phone when euthymic. More sophisticated analyses using mixed linear models and machine learning algorithms showed that patients (regardless of their mental state) could be distinguished from healthy volunteers with a sensitivity of 0.92, a specificity of 0.39, a positive predictive value of 0.88 and a negative predictive value (NPV) of 0.52. The results were similar when the patient group was restricted to those who were euthymic. Models that attempted to differentiate between affective states performed less well probably because the number of observations may have been insufficient. The low specificity and the low NPV suggest caution in moving forward. Specificity is the ability of a ‘test’ to correctly identify unaffected individuals (true negative rate), while NPV is the probability that an individual with a negative ‘test’ result is truly unaffected. Because both these indices were low, the smartphone data used here may not be suitable for diagnosing BD in the general population or in unselected clinical populations.

In many ways, this study highlights the key issues we are faced with regarding digital phenotyping. On one hand, active and passive digital data, when appropriately analysed, seem to have the potential to provide novel and useful insights about key scientific and clinical questions. However, the extent to which this potential will be fulfilled remains unclear and will require further studies to identify the most useful digital metrics, to refine study design in terms of duration and larger samples to improve the power and robustness of the results.