Is Automated Spike Detection and Source Modeling Ready for Routine Use?
Abstract number :
3.297
Submission category :
3. Neurophysiology / 3G. Computational Analysis & Modeling of EEG
Year :
2024
Submission ID :
291
Source :
www.aesnet.org
Presentation date :
12/9/2024 12:00:00 AM
Published date :
Authors :
Presenting Author: Dan Dimitriu, MD – Atrium Health
Yuan Fan, MD, PhD – Atrium Health
Lauren Feldman, DO – Atrium Health
Beverly Williams, R.EEGT, CLTM – Atrium Health
Dmytro Bielushchenko, R.EEGT – Atrium Health
Ashley Hastings, R.EEGT – Atrium Health
John Ebersole, MD – Atrium Health
Rajdeep Singh, MD, MS, FAES – Atrium Health
Rationale: EEG source imaging (ESI) of epileptic spikes in long-term monitoring (LTM) recordings requires considerable effort. Epileptologists today often have little time or expertise for such additional diagnostic activities. One solution may be commercially available services to identify and source model spikes automatically.
Methods: Four epileptologists (MDs) without specialized ESI training at Carolinas Medical Center reviewed, in blinded fashion, the automated source analyses of LTM data from 20 consecutive patients provided by Persyst 14 powered by Epilog. This was compared to a manual review of the raw LTM data by neuro-analyst technologists. Both were asked to determine the number of real spike types versus detected artifacts or normal transients. MDs were additionally asked to if they could confidently interpret the resultant source models. Finally, MDs ranked each automated report as clinically very useful, modestly useful, or not useful.
Results: Spike types automatically detected were often normal transients or artifacts (55/78, 70%). Averaging multiple such transients often produced a confusing spike-like morphology (50/78, 64%). Review of individual spikes was necessary to clarify whether they were real (15/20 patients, 75%). Automated source analysis tended to create duplicate spike types from one likely source (16/78, 21%). MDs felt unsure interpreting the provided source models in 80% (16/20) of patients. Overall, 45% (9/20) of automated analyses were found to be “modestly useful”, but the remainder were thought to be “not useful” clinically in localizing spike foci beyond EEG trace inspection.
Conclusions: Fully-automated ESI is not ready for routine clinical use given high rates of false spike detections and misleading source models of non-spike transients. Epileptologists using commercial services need additional education to interpret ESI results confidently. Automated ESI currently will likely require human sub-selection of real detected spike types to achieve accurate and clinically useful source models.
Funding: None
Neurophysiology