2024 American Academy of Neurology Abstract Website

Katherine McFarlane¹, Corinne Allas², Samuel Terman⁴, Sung Hyun Seo⁴, Amir Karimi², Siddhika Sreenivasan², Jena Grauer², Di Sun⁵, Meagan Watson⁶, Elissa Patterson⁴, John Stern², Jamie Feusner³, Laura Strom⁶, William Stacey⁴, Wesley Kerr¹
¹Department of Neurology, University of Pittsburgh, ²Department of Neurology, ³Department of Psychiatry and Biobehavioral Sciences, University of California, Los Angeles, ⁴Department of Neurology, ⁵Department of Radiology, University of Michigan, ⁶Department of Neurology, University of Colorado

Objective:

We evaluated if our machine learning-derived score improved the ability of seizure specialists and non-seizure specialists in accurately identifying functional seizures (FS) as compared to epilepsy based on information available prior to ictal video-electroencephalography.

Background:

The diagnosis of FS (also known as psychogenic nonepileptic seizures [PNES]) without ictal video-electroencephalography is challenging. Delayed and inaccurate diagnosis has been associated with worse long-term treatment outcomes. The Functional Seizures Likelihood Score (FSLS) is a machine learning-based diagnostic that identifies patients with probable FS.

Design/Methods:

117 anonymized cases were constructed using data from patients with documented FS, epilepsy, mixed epileptic and FS, or physiologic seizure-like events. Data were presented sequentially, showing the history of present illness (HPI), followed by FSLS (in 50% of cases), electroencephalogram (EEG), and imaging (MRI). Clinician-participants were asked the most likely diagnosis after viewing each sequential piece of data. We evaluated the impact of level of expertise in seizures and of each piece of data using mixed-effects logistic regression.

Results:

Participants saw 1,057 cases, with a median of 3.5 cases per person [interquartile range: 1-7], with 28% of participants viewing only one case. Based on the HPI alone, seizure-specialists’ accuracy was significantly higher than non-seizure specialists (68.2% vs 51.7%, p=0.02), and not statistically different from the FSLS without human review (78.6%, p=0.06). Non-specialists’ accuracy improved with the FSLS (56.7%, Odds Ratio 1.54, p=0.05), but specialists’ accuracy did not (68.0%, Odds Ratio 0.74, p=0.31). For all participants, diagnostic accuracy increased when viewing more data (OR: 1.33 FSLS, 1.41 EEG, 1.53 MRI, and p=0.09, 0.046, 0.01, respectively).

Conclusions:

The FSLS improved non-seizure specialists’ diagnosis of FS but did not improve seizure specialists’ diagnosis. Therefore, implementation of the FSLS may focus on utilization by non-seizure specialists to reduce delays to referral to seizure specialists and thereby accurate diagnosis.