Multimodal deep learning model using MRS and clinical data classifies auditory hallucinations in bipolar teens

A multimodal deep learning model using MRS-derived vmPFC metabolites and clinical data classified auditory hallucinations in adolescents with bipolar depression with 71.43% accuracy (F1=0.75) in a small retrospective study.

• Retrospective analysis of 47 untreated adolescent bipolar depression patients, split by presence of verbal auditory hallucinations (PANSS P3 score).
• The model, using a Transformer-based framework with bidirectional cross-attention, achieved balanced precision, recall, and F1-score of 0.75 on the fixed test set.
• Limitations include very small sample size, single-institution data, no external validation, and retrospective design; findings are preliminary.

RadPigeon summaries are original and for information only. They are not clinical advice.