Abstracts

Rationale: There is an increasing awareness that many seizures are without physical signs, so that EEG monitoring is necessary to detect the event. This is particularly true in pediatric patients in the intensive care unit where the percentage of non-convulsive events can be as high is 80% (Abend). One answer is to develop more powerful technologies that expand clinician capacity. Many successful machine learning models apply non-linear dynamic systems analysis tools (Bardia and Goldenholz). This study reports on the between subject correlation structure of multiple forms of entropy for multiple size electrode arrays. Entropy was chosen as a starting point since it is one of the most commonly applied non-linear metrics in this field.

Methods: One signal processing strategy for seizure detection is to combine eigenvalue metrics with entropy metrics, Preliminary results show high detection level performance (86-94%) in children. Initial selection among the more than sixteen methods for calculation of entropy was not based upon empirical comparisons. Hence, we compare the results of three different implementations of entropy (Wentropy toolbox in MATLAB 2020b), across multiple numbers of electrodes, using the recordings in the Boston Children’s/MIT database. Earlier studies in adults suggest optimizing detection rates may not need the full set of the 10/20 electrode locations. The three forms of entropy were compared for six different size electrode sets, in terms of cross correlation measures.

Results: The Shannon, Sure (P value of 5000), and Log Energy entropies equations are shown below. The entropy window size was held constant at 30 sec with a step size of 2 sec. Files of the same patient were concatenated together to produce four-hour files for all 24 patients and filtered using a 2-10 Hertz bandpass filter. The Pearson correlation coefficient compared the baseline 23 channels to the various electrode arrays, for all three entropies, Table 1 shows the averaged correlation coefficients of all 24 patients. The Log Energy metric proved to show the greatest agreement among the entropies. The Shannon and Sure Entropy performed well with the use of 8, 10, and 16 channels. The range of the Log energy values was the smallest of the three forms of entropy and showed the highest absolute values.

Conclusions: There were differences in the performances between the several entropy algorithms, from the standpoint of consistency. Amongst the three tested entropy algorithms the log algorithm was superior compared to Shannon entropy (the most commonly used). It was also clear, from the consistency (correlation) view that more electrodes were more consistent than fewer. This does not answer the question of whether the log entropy metric would be superior from the detection rate or the false alarm measure perspectives. It does answer the question of whether there are differences among the several forms of entropy in the pediatric population, which should be taken into account when developing automated seizure detection methods for this age group.

Funding: Please list any funding that was received in support of this abstract.: Complete funding provided by AGCS.