(580) Improvements in a Machine-Learning Algorithm for Detecting Status Epilepticus

Introduction: Point-of-care EEG (POC EEG) systems enable rapid set-up and triage of patients at risk for seizures in acute care settings. Machine learning algorithms can enhance the benefit for bedside care teams by providing continuous monitoring of seizure activity. In this study, we evaluated the accuracy of an algorithm (Clarity, Ceribell Inc.) to detect status epilepticus in a large dataset of POC EEGs.

Methods: We retrospectively analyzed a dataset of 665 POC EEG recordings across 11 different hospitals. The output (maximum seizure burden, max. SzB) of two different versions of Clarity was recorded post-hoc for each file. The POC EEGs were also evaluated post-hoc by at least two epileptologists (three in case of disagreements) who were blinded to the Clarity output. We compared the accuracy of algorithm performance at ≥ 90% max. SzB (indicating suspected status epilepticus) to the majority consensus of the epileptologists in terms of sensitivity, specificity, negative and positive predictive value.

Results: The epileptologists labeled 20 cases as SE. For nineteen of these, the Clarity algorithm generated a status epilepticus alert (≥ 90% max. SzB). Clarity also alerted on 17 additional files, only 1 of which was labeled as slow/non-epileptiform by the epileptologists. Overall, Clarity demonstrated 95% sensitivity, 97% specificity, 53% PPV and 99% NPV for SE. This represents an increase in accuracy over a previous version of the algorithm, which showed 95% sensitivity, 94% specificity, 32% PPV and 98% NPV.

Conclusions: The accuracy of Claritγ for detecting SE events in this large dataset – when compared against evaluations of the EEG by epileptologists – indicates the algorithm’s ability to assist clinicians in quickly identifying and treating, or ruling out, SE in critical care environments. Moreover, the increase in specificity and PPV achieved in the latest version of Claritγ demonstrates continuing improvement of this machine learning algorithm.