An explainable multimodal artificial intelligence model for classifying suicide attempters with borderline personality disorder: a pilot study.
Authors
Affiliations (8)
Affiliations (8)
- Laboratory of Neuroinformatics, IRCCS Istituto Centro San Giovanni di Dio Fatebenefratelli, Brescia, Italy. [email protected].
- Laboratory of Neuroinformatics, IRCCS Istituto Centro San Giovanni di Dio Fatebenefratelli, Brescia, Italy.
- Unit of Epidemiological Psychiatry and Digital Mental Health, IRCCS Istituto Centro San Giovanni di Dio Fatebenefratelli, Brescia, Italy.
- Unit of Psychiatry, IRCCS Istituto Centro San Giovanni di Dio Fatebenefratelli, Brescia, Italy.
- Department of Medical and Surgical Specialties, Neuroradiology Unit, University of Brescia, 25123, Brescia, Italy.
- Laboratory of Alzheimer's Neuroimaging and Epidemiology (LANE), IRCCS Istituto Centro San Giovanni di Dio Fatebenefratelli, Brescia, Italy.
- Department of Human Science, "Guglielmo Marconi" University, Rome, Italy.
- Third Centre of Cognitive Psychotherapy-Italian School of Cognitive Psychotherapy (SICC), Rome, Italy.
Abstract
Borderline Personality Disorder (BPD) is a severe mental disorder marked by emotional dysregulation. Estimates show that 73% of patients with BPD will have, on average, three suicide attempts in their lifetime, with up to 10% of cases resulting in death. Reliable tools to identify risk factors associated with suicide are lacking. Artificial Intelligence (AI) could fill this gap, supporting the development of effective intervention strategies. This pilot study provides preliminary evidence that a multimodal signature could differentiate suicide attempts in individuals with BPD, paving the way to prospective cohort validation and clinical applications. We developed DRAMA-BPD (Detecting Retrospective suicide Attempts with Machine learning Approaches in Borderline Personality Disorder), an explainable, multimodal, Machine Learning (ML) model based on an ensemble classifier of lifetime suicide attempters among people with BPD. DRAMA-BPD was trained on the sociodemographic, clinical, and MRI data of 104 individuals with BPD recruited from two cohorts. Processing techniques adopted included feature extraction. SHapley Additive exPlanations (SHAP) was used to assess model interpretability. DRAMA-BPD achieved a balanced accuracy of 0.68, sensitivity of 0.58, specificity of 0.77, and AUC of 0.68. SHAP analysis identified cortical volumes and thickness from T1-weighted images and Symptoms Checklist 90 Revised (SCL-90-R) as the main contributors to classification.