Can machine learning predict PTSD symptoms from trauma narratives of children and adolescents?

Background: Machine learning approaches are being increasingly tested as a potential means of identifying mental health conditions. Narrative features of trauma memories are proposed to play a significant role in the development of post-traumatic stress disorder (PTSD), meaning that trauma narratives provide an excellent context in which to test machine learning capabilities. The potential for children's trauma narratives to predict post-traumatic stress remains particularly poorly studied. Here, we tested whether the application of machine learning to trauma narrative characteristics can predict PTSD symptoms in young individuals exposed to trauma.

Study methodology: Two pre-trained large language models and two benchmark models were fine-tuned and trained to predict PTSD symptom severity from children's autobiographical narratives of a traumatic event. Data comprised narratives collected one month post-trauma from 400 individuals aged 7-17 years old who experienced a psychological trauma that led to attendance at emergency departments in the United Kingdom (N = 178) and South Africa (N = 222), as well as self-reported PTSD symptoms and trauma memory features.

Findings: Both pre-trained and benchmark models demonstrated poor predictive performance across trauma narratives in the United Kingdom, South Africa, and the combined datasets (e.g. RoBERTa R² = -.05; LASSO R² ≈ 0). However, adding self-reported trauma memory features, disorganisation, and sensory vividness improved the benchmark models' performances, especially in the UK dataset (e.g. LASSO R² = .57; XGBoost R² = .45).

Conclusions: These findings indicate that while trauma narratives alone offer limited predictive value, incorporating self-reported trauma memory characteristics substantially enhances model performance, highlighting the importance of focusing on subjective reports to develop scalable automated tools for PTSD risk prediction in youth.