Abstracts

Rationale: Periodic discharges (PDs, such as lateralized PDs - LPDs, bilateral independent PDs - BIPDs, and generalized PDs - GPDs) are important EEG findings in ICU patients because they are often associated with seizures, acute structural lesions, and CNS infections. PDs have also been associated with poor prognosis (Claassen et al., 2005) and higher mortality. Therefore, early recognition of PDs can be important in diagnostic and therapeutic management. However, continuous review and interpretation of raw EEG data to identify PDs in long-term EEG monitoring is labor intensive and therefore impractical in most clinical settings. Tools aiding in rapid and accurate detection of PDs would greatly improve efficiency of continuous EEG (cEEG) monitoring and enhance the quality of patient care in ICUs. We report a preliminary investigation of a novel automated algorithm for the detection of PDs in cEEG recordings from ICU patients. Methods: One hundred 20-sec EEG epochs were sampled from cEEG recordings in each of 6 ICU patients, three with PDs noted in their video-EEG reports and the others without. To ensure a sufficient sample size of PD epochs in each "PD" patient, an initial sampling was conducted such that 30 epochs clearly exhibit PDs (based on the 2012 ACNS definition of PDs) and the remaining 70 were randomly sampled from segments without clear PDs. For the three "Non-PD" patients, all 100 epochs were randomly sampled from their cEEG recordings and thus include various interictal states as well as artifacts (muscles, chewing, etc.). These 600 EEG epochs were then independently reviewed by four EEG experts and an epoch was considered a PD epoch if it was marked by at least two experts. The PD detection algorithm reported herein first identified, in each channel, EEG discharges (i.e., spikes, sharp waves, or sharply contoured waves) in the epoch based on their peak values, durations, and morphology. A PD event was detected if, in any channel, the identified discharges repeated for more than five times with a deviation among the inter-discharge intervals smaller than a threshold. Sensitivity and specificity of the algorithm were evaluated based on the comparisons of the detection results against the epoch classifications determined by the EEG experts. Results: Based on the experts' review, the proposed algorithm performed a sensitivity of 90.5% with an overall specificity of 99.4% (100% in "PD" subjects and 99.0% in "Non-PD" subjects). In addition, the analysis of the inter-rater variability revealed a moderate (1 out of 6 rater pairs) to substantial agreement among the four EEG experts (Cohen's kappa statistic: range=[0.57,0.91], mean=0.75, SD=0.12), which provided a sufficient confidence level on the assessment of algorithm performance. Conclusions: These findings suggest that the proposed PD detection algorithm has potential to be the basis of clinically useful software that can assist ICU staff in timely identification of PDs. Further studies in a larger sample of patients that include more EEG patterns of PDs and background activities are warranted.