身体醜形障害におけるスマートフォンセンサーおよび人口統計データを用いた臨床的悪化の低負担検出

Body dysmorphic disorder (BDD) is characterized by distressing preoccupations with perceived appearance flaws, leading to functional impairment and suicidal ideation (SI). Traditional approaches for monitoring clinical deterioration in BDD include self-reports and clinician assessments, which can miss acute changes in risk due to infrequent administration and recall biases. Alternatively, real-time monitoring via smartphones and wearable devices can enable low-burden early detection of deterioration, identifying intervention opportunities before someone's condition critically worsens. This study tests the feasibility of using smartphone sensor and demographic data to predict daily clinical acuity. Eighty-two participants with BDD completed ecological momentary assessments (EMA) over 28 days, reporting levels of SI, BDD-related avoidance, and time spent on BDD-related concerns. Smartphone sensor data were collected for 3 months that overlapped with EMA. Machine learning models were trained to predict same-day levels of SI, avoidance, and time spent on BDD using the Global Positioning System (GPS), accelerometer, and demographic data. We evaluated model performance using mean absolute error, Pearson and Spearman correlations, and permutation tests. Random forest (RF) models using time and random split validation outperformed dummy regressor models across outcomes (maximum SI, mean SI, maximum avoidance, mean avoidance, time spent on BDD-related behaviors). Pearson correlations for RF models showed strong predictive performance for BDD-related time (r = .74-.75) and mean and max SI (r = .70-.73). Mean and max avoidance was moderately well predicted (r = .56-.62). Step count and demographic factors (e.g., education, living situation) were the most consistent and important features. This study provides initial evidence that smartphone sensor and demographic data can be used to monitor real-time clinical worsening in BDD, without burdening the patient. This work has potential for building just-in-time interventions that are delivered as deterioration onsets, to prevent its escalation. Future research should test these models in real-world datasets collected over longer periods and subsequently explore integration into interventions and clinical decision making. Trial Registration: ClinicalTrials.gov Identifier: NCT04254575.