Feasibility of machine learning to improve the diagnostic odyssey for functional seizures

University Of Pittsburgh At Pittsburgh, Pittsburgh PA

Investigators

Abstract

Project Summary/Abstract As many as 10% of outpatient and emergency visits for recurrent unprovoked seizures are not for epilepsy; they are for functional seizures (FS), also called psychogenic nonepileptic seizures. A fundamental clinical problem is the average delay to FS diagnosis of 8.4 years during which FS transitions to a chronic disability. Latency to diagnosis is a key prognostic indicator for seizure freedom and other outcomes of FS treatment. Computer-aided diagnostic tools (CADT) are designed to address these types of gaps by alerting clinicians to frequently missed diagnoses as early as possible after initial presentation. We developed the Functional Seizures Likelihood Score (FSLS) using data from video-EEG monitoring (VEM) with epileptologists. My long-term career goal is to be an academic epileptologist-statistician developing and implementing data-driven tools to improve the diagnosis and treatment of seizures. The objective of this proposal is to modify the FSLS for use by non-epileptologists, obtain preliminary data for a prospective trial that uses the electronic health record to automatically identify patients with probable FS and triages them to early VEM, and train me to direct that trial. My central hypotheses are that non-expert performance will be enhanced by the FSLS and, in patients selected by the FSLS, the diagnostic yield of early VEM is at least as high as standard-of-care VEM. The rationale of this study is to prepare for a clinical trial to shorten the diagnostic odyssey of patients with FS by enhancing non-epileptologists with a CADT that prompts referral for early VEM. In Aim 1, I propose to validate transfer of the FSLS for use by non-expert providers. Using anonymized online case-vignettes from patients diagnosed with either FS or epilepsy to present sequential information to non-expert providers and evaluate how the FSLS, EEG, neuroimaging, and video information influence their diagnoses. I hypothesize that non-expert plus FSLS accuracy is higher than non-expert accuracy without FSLS. In Aim 2, I will apply the FSLS to outpatient and emergency records as a clinical informatics tool. I hypothesize that this will demonstrate that the positive predictive value of the FSLS is more than 50%. In Aim 3, I will pilot prospective implementation of machine learning tools, like the FSLS, to identify patients who would benefit from early VEM. I hypothesize that, when pre-test probability of FS is elevated by the selection tools supervised by an epileptologist, early VEM is highly diagnostic and changes management. Upon completion of these aims, the expected outcome is to be ready to perform a prospective trial (R01) of early identification of patients with FS in the outpatient and emergency setting. If feasible, this intervention could transform the diagnostic odyssey of FS by greatly reducing time to correct diagnosis. This work will also provide an excellent opportunity to receive training in clinical trial design, analysis, and participant recruitment.