Machine Learning Falls Short at Predicting Repeat Self-Inflicted Neck Injuries

Machine learning models trained on nearly two decades of hospital records have failed to reliably predict which young adults will repeat self-inflicted injuries to the front of the neck, according to a study published in Scientific Reports. The finding tempers expectations that algorithmic tools could soon flag patients at highest risk of recurrence after such injuries, which are known to carry an elevated risk of subsequent suicide attempts and suicide. Researchers led by Rasheed Omobolaji Alabi and Riikka E. Mäkitie examined a cohort of 167 individuals aged 18 to 30 years who had been treated for self-inflicted anterior neck injuries at Helsinki University Hospital between 2005 and 2023.

The team systematically reviewed patient records to extract demographic details, clinical characteristics, psychiatric comorbidity and information on clinical management. These variables were used to build three different machine learning models: a logistic regression model, an easy ensemble model, and an extreme gradient boosting model. To address the imbalance between patients who did and did not experience a recurrence, the researchers applied class-imbalance handling and feature selection techniques across all three models.

Interpretability was addressed using SHAP, or SHapley Additive exPlanations, which allowed the researchers to examine both global patterns in the models' predictions and the factors driving predictions for individual patients. According to the SHAP analysis, the most important predictive variables for recurrence were a history of other trauma, the patient's age at presentation, and the mode of arrival at hospital, specifically whether the patient walked in or was brought in by others. Despite these insights, all three models performed poorly.

The best balanced accuracy among the three was 60.0 percent, with a Brier score of 0.23, a precision-recall area under the curve of 0.30, a receiver operating characteristic area under the curve of 0.66, a recall of 0.75, an F1 score of 0.32 and an overall accuracy of 55 percent. These figures indicate that the models could not distinguish reliably between patients who would sustain a repeat injury and those who would not, limiting any clinical application in their current form. The authors are explicit about these limitations.

They state that the models require further training and validation and currently hold limited clinical usability. The small cohort size of 167 patients, drawn from a single hospital, likely constrained the models' ability to learn generalizable patterns, a common challenge when applying machine learning to rare clinical events. Nevertheless, the researchers argue that the effort is worthwhile.

Rates of self-harm among young adults are rising, and identifying patients at risk of injury recurrence could improve clinical management and intervention after an initial self-inflicted neck injury. The study's authors emphasize that future research will need larger and more diverse datasets, along with external validation on patient populations beyond the original cohort, to improve the predictive capabilities of these models. Only then, they suggest, could such tools mature into effective prediction aids that support, rather than mislead, clinical decision-making for young adults at risk of repeating self-inflicted neck trauma.

Subject of Research: Machine learning prediction of recurrence of self-inflicted anterior neck injuries in young adults

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

Article References: Alabi, R. O., & Mäkitie, R. E. (2026). Machine learning in predicting recurrence of self-inflicted neck trauma in young adults. Scientific Reports. https://doi.org/10.1038/s41598-026-75963-6

Image Credits: AI Generated

DOI: 10.1038/s41598-026-75963-6

Keywords: machine learning, self-harm, neck trauma, young adults, suicide risk, predictive modeling, SHAP, logistic regression, gradient boosting, Helsinki University Hospital, clinical prediction, injury recurrence