A Longitudinal Prediction of Suicide Attempts in Borderline Personality Disorder: A Machine Learning Study

Borderline personality disorder (BPD) is associated with a high risk of suicide. Despite several risk factors being known, identifying vulnerable patients in clinical practice remains a challenge so far. The current study aimed at predicting suicide attempts among BPD patients during disorder-specific psychotherapeutic interventions exploiting machine learning techniques. The study took into account several potential predictors relevant to BPD psychopathology: emotion dysregulation, temperamental and character factors, attachment style, impulsivity, and aggression. The sample included 69 patients with BPD who completed the Temperament and Character Inventory, Attachment Style Questionnaire, Difficulties in Emotion Regulation Scale, Barratt Impulsiveness Scale, and Aggression Questionnaire at baseline and after 6 months of psychotherapy. To detect future suicide attempts, baseline questionnaires were entered as predictors into an elastic net penalized regression, whose predictive performance was assessed through nested fivefold cross-validation. At the same time, 5000 iterations of a non-parametric bootstrap were used to determine predictors' robustness. The elastic net model discriminating BPD suicide attempters from non-attempters reached a balanced accuracy of 64.09% and an area under the receiver operating curve of 70.44%. High preoccupation with relationships, harm avoidance, and reward dependence, along with low motor impulsiveness, verbal aggression, cooperativeness, and self-transcendence were the most contributing predictors. Our findings suggest that interpersonal vulnerability and internalizing factors are the strongest predictors of future suicide attempts in BPD. Machine learning on self-report psychological scales may be helpful to identify individuals at suicidal risk, potentially helping clinical settings to develop individualized preventive strategies.