Machine learning in predicting recurrence of self-inflicted neck trauma in young adults

Self-inflicted injuries, particularly among young adults, are associated with an increased risk of subsequent suicide attempts and suicide. Injuries to the anterior neck present unique challenges in management due to the presence of vital anatomical structures and possibly complicated consequences. Therefore, this study employed machine learning (ML) techniques to predict injury recurrence in patients with self-inflected anterior neck injuries. The cohort comprised 167 individuals aged 18 to 30 years with self-inflicted injury to the anterior neck at the Helsinki University Hospital in 2005–2023. Patient records were systematically analyzed for patient demographics, clinical characteristics, psychiatric comorbidity, and clinical management for building a ML model. Class imbalance and feature selection were handled across three different models; logistic regression, easy ensemble, and extreme gradient boosting-based models. Furthermore, predictions were enhanced using SHapley Additive exPlanations (SHAP) for both global and local interpretability. Our results showed that the most important predictive variables for a recurrence were history of other trauma, patient age at presentation, and mode of arrival (walk-in versus brought in). All three injury recurrence models showed poor performance. The highest balanced accuracy between the three models was 60.0%, Brier score of 0.23, precision-recall area under curve (AUC) value of 0.30, receiver operating characteristic-AUC of 0.66, recall value of 0.75, F1 score of 0.32, and accuracy of 55%. Our models showed poor performance and therefore still require further training and validation and, currently, hold limited clinical usability. However, given the rising rates of self-harm among young adults, identifying patients at risk of injury recurrence may improve clinical management and intervention. Future research is needed with larger, diverse datasets and external validation to improve the predictive capabilities of these ML models, ultimately aiding in the development of effective prediction tools.