Digital Phenotyping of Individuals with Suicide Attempts: Comparison and Classification with Healthy Controls Using passive-sensing Smartphone Data (Preprint)

<sec> <title>BACKGROUND</title> Suicide is among the leading causes of death and a major burden worldwide. Predicting who is at risk remains challenging and interventions for self-injurious thoughts and behaviors have shown limited effectiveness, with little progress despite extensive research. Digital phenotyping using smartphone-based data offers promise, but passive sensing data alone has shown limited evidence for accurately predicting suicidal thoughts and behaviors. </sec> <sec> <title>OBJECTIVE</title> This study examined whether digital markers derived from longitudinal passive smartphone sensing data can differentiate individuals with a recent history of suicide attempts from healthy controls. The aims were to (1) identify significant behavioral differences between the two groups, (2) classify individuals after suicide attempts versus healthy controls using machine learning (ML) models, and (3) determine the most predictive features for classification. </sec> <sec> <title>METHODS</title> From December 2018 to February 2023, 208 patients witch a recent suicide attempt were approached on the wards of a psychiatric university hospital. Of these 56 consented to data collection, and 15 datasets met the criteria for inclusion in the analysis. These were analyzed alongside 37 sets from non-suicidal control participants. Longitudinal passive sensing data – including movement, location, physical activity, and smartphone interactions – were aggregated into 39 daily features. After pre-processing and feature extraction, group differences were examined, and the features were used in a population-level ML model to classify individuals with a suicide attempt against controls. </sec> <sec> <title>RESULTS</title> Compared to controls, participants in the suicide attempt group showed significantly reduced mobility and engaged less in activities such as running, walking, cycling, and transportation. They also exhibited higher total smartphone usage durations, despite unlocking their phones less frequently than controls. A leave-one-out cross-validation (LOOCV) ML model using the extracted features to classify group membership achieved an accuracy of 0.62 and an ROC-AUC of 0.68. No significant group differences (at α=.01 level) were found for age (M=30.35 ± 11.56), or gender distribution (63.46% female). </sec> <sec> <title>CONCLUSIONS</title> Digital behavioral phenotypes from passive sensing show promise for advanc¬ing the understanding, monitoring, and assessment of suicide risk. While classification performance based solely on sensor-derived features was moderate, we observed marked differences in mobility, physical activity, and smartphone usage between individuals with a recent suicide attempt and healthy controls. Larger samples, more reliable sensor data, and longer observation periods may enhance classification performance in future research. </sec> <sec> <title>CLINICALTRIAL</title> NCT03732300 </sec>