A Comparison of Two Deep Learning Approaches to Distinguish Functional Dissociative from Epileptic Seizures Using Event Videos

BackgroundDifferentiating between motor functional dissociative seizures (FDS) and motor epileptic seizures (ES) is a common diagnostic challenge, requiring video electroencephalography (vEEG) as gold standard. However, vEEG requires specialized technicians and clinical experts to set up and interpret and oftentimes fails to capture events. We sought to develop machine-learning (ML) tools to carry out this diagnostic task independently of vEEG or human review by a neurologist. MethodsIn this retrospective study, we developed two proof-of-concept ML models to differentiate motor ES from FDS based on video of FDS and ES events in patients who underwent inpatient vEEG monitoring at an academic medical center between 2012 and 2021. The first model employed a pose-estimation approach, using body landmark features that were labeled frame-by-frame by three neurologists. The second utilized an end-to-end 3D convolutional neural network (CNN), thereby learning directly from raw video frames. Using board-certified epileptologist review as a clinical gold standard, we measured model performance by area under the receiver-operating (AUROC) and precision-recall (AUPRC) curves, sensitivity, precision, and accuracy against a held-out test set of videos. FindingsWe included 101 unique patients with 106 total event videos, comprising 61 (60.4%) ES and 45 (44.6%) FDS events. Both ML models distinguished both seizure types better than chance. The pose-estimation-based model achieved an AUROC of 0.71, AUPRC 0.53, sensitivity 0.90, precision 0.50, and accuracy 0.62. The CNN model exhibited superior overall performance, achieving an AUROC 0.78, AUPRC 0.84, balanced sensitivity and precision (both 0.82), and accuracy 0.80. InterpretationOur findings demonstrate the superiority of CNN over pose-estimation models to differentiate between motor ES and FDS using video alone. Although future studies are needed, these models hold potential as adjunct diagnostic tools by enabling rapid, objective seizure evaluations without immediate neurologist involvement. FundingCTSA grant UL1TR004419 (Kummer), NIH grants R01DK133539 and R01HL167050 (Nadkarni), NIMH T32 grant MH122394 and AAN grant 19-1120 (Chan).