Abstracts

This is a Late-Breaking abstract.

Rationale: Artificial intelligence (AI) algorithms are increasingly being integrated into the practice of
medicine as they improve the quality of life of care providers and the precision of their diagnostic
decision making. We recently introduced an FDA-cleared machine learning algorithm (Claritγ, Ceribell
Inc.) that automatically, continuously, and in near real-time, measures the load of seizure activity in a
rolling 5-minute long window of recording every 10 seconds. When the seizure activity is detected to be
present in >90% of the recording (i.e., 4.5 minutes long), it alerts at the bedside warning of possible case
of status epilepticus. Here, we provide new data on the performance accuracy of a new version of this
algorithm after improving its status alert accuracy.

Methods: We retrospectively selected 353 EEG recordings with Rapid EEG (Ceribell Inc) across 6
different hospitals. We used automated measure of seizure burden as the percentage of ten-second
epochs with seizure activity in any 5-min EEG segment. We compared the accuracy of algorithm
performance at ≥ 90% to the majority consensus of at least two expert neurologists.

Results: Majority consensus of neurologists labeled the 353 EEGs as normal or slow activity (n = 249),
highly epileptiform patterns (HEP, n = 87), or seizures [n = 17, nine of which were labeled as status
epilepticus (longer than 5 min]. Clarity algorithm generated a status epilepticus alert (≥ 90% seizure
burden) with 100% sensitivity, 96.2% specificity and 100% negative predictive value for status
epilepticus. However, the positive predictive value of this alert was 40.9% due to overcalling 13 of 353
recordings. All of the 13 overcalled recordings had been labeled by majority consensus of neurologists as
HEP. This is an improvement over the previously published version of the algorithm that had a 27% PPV,
93% specificity and 100% sensitivity and 100% NPV. The algorithm generated a 0% seizure burden for
218 recordings, out of which 181 were labeled by majority consensus of neurologists as normal or slow
activity, 35 as HEP, and 2 as short lasting focal seizures.

Conclusions: While the Claritγ AI algorithm has room for improvement, it detects SE events with high
sensitivity and relatively high specificity. Our data illustrate an important aspect of AI algorithms in
healthcare, namely that their performance will only improve with time and additional cases. The high
negative predictive value of the algorithm at 0% threshold suggests that cases of status epilepticus can
be ruled out relatively accurately in a large proportion of cases within minutes of EEG recordings and
thus can help prevent unnecessary or aggressive over-treatment in critical care settings.

Funding: Not applicable