Predicting the strength of next-day negative emotion states in body dysmorphic disorder using passive smartphone data: An intensive longitudinal assessment study

Body dysmorphic disorder (BDD) is a debilitating and common psychiatric illness associated with high rates of suicide and substance use disorders. Negative emotions – particularly shame and anxiety – are elevated in BDD and correlate with suicide risk and substance use. It is critical to have reliable and valid tools to assess negative emotions in BDD. Retrospective self-reports are subject to recall biases, average one's experiences over broad time frames, and are burdensome to complete. Alternatively, sensor-based digital phenotyping has potential to yield low-burden emotion assessment within acute time frames. This study aimed to use smartphone sensor data (GPS, accelerometer, collected over 3 months) to predict next-day peak shame, anxiety, and general negative emotion states (collected via 28 days of ecological momentary assessment) in 83 adults with BDD. We tested cumulative link mixed models [CLMM]) and random forest [RF] models. RFs outperformed CLMMs across prediction performance metrics and had overall prediction accuracies (i.e., proportion of predicted scores that exactly matched actual scores, out of total predictions) of 42.1–50.0 %, versus 10.9–20.2 % for CLMMs. Binary predictive performance at high levels of negative emotion was moderate. Developing unobtrusive methods for predicting shame, anxiety, and general negative emotion states over acute time frames using smartphone sensor data can enable just-in-time intervention opportunities, as a future step to reduce risk for suicide and substance use in BDD. Models might be strengthened with larger samples, data collected over longer time frames, and incorporation of wearable-based physiological data.

Trial Registration: ClinicalTrials.gov Identifier: NCT04254575.