Objective smartphone data as a potential diagnostic marker of bipolar disorder

You don’t need a genetic test to tell the difference between a duck and a yak.

David Sackett

Faurholt-Jepsen and colleagues used the MONARCA (Faurholt-Jepsen et al., 2015, 2018, 2019) software to test its potential to assist with the diagnosis of bipolar disorder. They assessed several passive signals from the phones of a small group of people with bipolar disorder (n = 29) compared to a group of healthy controls (n = 37). The passive signals captured several aspects of smartphone use, including the number of outcome and incoming calls and texts, duration of phone calls, the number of times the smartphone screens were turned on and off and the duration the smartphone screen was on (without any active input from the participants). One could refer to these data as ‘digital dust’. To make sense of the digital dust, the investigators used a suite of sophisticated statistical tools including linear mixed effects regression models and machine learning.

They found that a subset of the digital dust distinguished the bipolar from the healthy control group: more incoming texts in the bipolar group for those in a manic or mixed state; longer phone calls in the bipolar group regardless of mood state and the bipolar group had less time screen-on time during euthymia and overall. The group was careful to use the conservative Bonferroni correction to avoid the bias of multiple comparisons.

This study addresses a significant challenge: in making a diagnosis, we currently must rely on patient insight, memory, and clear communication – processes which can be compromised by dysregulated mood states such as mania. The digital dust provides an objective set of dynamic measures which overcome the shortfalls of self-report. However, this digital dust initiative faces the limitations of developing any diagnostic test. Specifically, many diagnostic tests fail to progress through systematic phases of evaluation similar to the phases required to develop a drug (Nierenberg and Feinstein, 1988; Ransohoff and Feinstein, 1978).

As a first step, Faurholt-Jepsen and colleagues compared the phone indicators for individuals with bipolar disorder with healthy control participants. How good was the phone as a marker for bipolar trait (disorder) or mood state? Sensitivity (percentage of patients actually ‘testing positive’) was very good (0.92). The same was true for differentiating euthymic patients with bipolar disorder from healthy controls (0.90). However, specificity for both trait and state (how many healthy controls are incorrectly classified as bipolar or having a certain mood state) was poor to modest. For example, specificity for the diagnosis was 0.39, which indicates that a sizable portion of healthy individuals were misclassified as having bipolar disorder. The same was true for incorrectly classifying euthymic and depressed mood in healthy individuals. MONARCA was most successful at correctly classifying controls as not being manic. Of note, positive and negative predictive values are of limited use in this scenario, as they depend on the base rate (or the proportion of people with and without the disorder of interest included in the study). Imagine having only a few individuals with bipolar in this mixed sample of patients and controls and all patients are misclassified, whereas most controls are not. Positive and negative predictive value would remain relatively high, falsely suggesting that the phone indicators may be a useful marker for diagnosis and mood state at this point in time.

Nevertheless, even with these caveats, the digital dust should eventually yield gold and be an important addition to the limits of our all too human memories.