Authors :
Presenting Author: Daniel Lachner-Piza, PhD – University of Calgary
Margarita Maltseva, MD – University of Calgary
Minette Krisel Manalo, MD – Alberta Children's Hospital
Julia Jacobs, MD, PhD – University of Calgary, Alberta Children's Hospital, Calgary, AB, Canada
Rationale: High Frequency Oscillations (HFO) are an electro-encephalographic (EEG) biomarker of both epilepsy and physiological processes. The recording of HFO on the scalp has been proven to be possible, as long as the signal to noise ratio is not too low
1. The process of visually marking HFO is a major challenge, primarily because of the required expertise, time-demand and subjectivity of the process. We have tackled the HFO marking problem by developing a two-stage automatic detector of scalp HFO.
Methods: Our raw data was comprised by the EEG from 27 patients with epilepsy, during slow-wave sleep and with a duration of 30 minutes. The 1
st stage of the HFO detection consisted of 11 HFO detectors that have been published and validated for either intracranial or scalp EEG. The unpruned-HFO events that came from these 11 detectors were then validated by a human expert-scorer (Margarita Maltseva, MM) using a software package developed for this task
2(elpi). The 2
nd stage consisted of a pruning classifier (gradient boosted tree), which discerned between expert validated and invalidated HFO. The features used to characterize the-HFO were engineered and procured to capture the characteristics of authentic HFO (Fig.1):
- 6 Oscillations or more
- Amplitude stands out from surrounding background
- Spectrogram shows an encapsulated power increase (blob-like)
Results: The 1
st stage detection produced a total of 170k unpruned HFO. The expert scorer marked only 5% of the unpruned HFO as valid (Fig.2). The data partitioning for the 2
nd stage pruning classifier was 64% training, :16% validation and :20% test. These ratios where kept in each patient. The pruning stage rejected on average 98% of the HFO (SD.=1.82%). The performance was measured for each patient using Cohen's kappa coefficient and was on average 0.63 (SD=0.2).
Conclusions: Our approach first produces an over-complete set of candidate HFO by leveraging the previous achievements from several research groups with expertise in HFO detection. The over-complete set of candidate HFO is then pruned to procure selecting only authentic HFO. The agreement with the expert scorer, measured by the kappa score and achieved after the pruning stage, is considered substantial. The 2-stage detection of scalp-HFO produced occurrence rates that were in the same range as visually marked HFO. The developed detector will allow to analyse the scalp HFO activity not only on longer EEG durations but also on larger patient cohorts.
Funding: This work was financially supported by the Canadian Institutes of Health Research.