Neuropathic Pain Detection Through Embedding Synergies of Deep Language and Image Models

The Global Burden of Disease states that neuropathic pain is suffered by 7–8% of adults worldwide, with severe repercussions on daily life, like drug abuse and psychological disorders. This work introduces a methodology for neuropathic pain classification through embedding synergies of the large language model BERT and image model ResNet50, handling clinical questionaries and EEG records, respectively. The classification task is a three-class problem with low, moderate, and severe pain categories. The embeddings of clinical data learned by BERT and the ResNet50-encoded topo-plots from EEG data feed an SVM classifier, further trained in a GroupKFold scheme from a thirty-six patients dataset. The accuracy obtained of 60%, outperforming single modality approaches, demonstrates the potential of multimodal approaches for enhanced pain diagnosis and treatment.