Authors :
Presenting Author: Franz Furbass, PhD – Beacon Biosignals
David Josephs, MS – Beacon Biosignals; Dave kleinschmidt, PhD – Beacon Biosignals; Alexander Chan, PhD – Beacon Biosignals; Sydney Cash, MD, PhD – Beacon Biosignals; M. Brandon Westover, MD, PhD – Beacon Biosignals; Jay Pathmanathan, MD, PhD – Beacon Biosignals; Jacob Donoghue, MD, PhD – Beacon Biosignals
Rationale:
Machine learning algorithms are transforming the interpretation of EEG, are increasingly valuable for screening high volumes of critical care monitoring datasets and generating neurobiomarkers for the development of therapies. While classifiers have traditionally focused on classical ictal patterns, it is increasingly understood that periodic and rhythmic patterns also pose clinical risk even when they are inter-ictal and subclinical. Here we describe a machine learning model capable of identifying classically ictal, rhythmic, periodic, and sporadic patterns. Methods:
We propose a unified classifier model capable of characterizing a continuum of epileptic and nonepileptic abnormalities. Raw EEG time samples from a 10-20 recording are taken as input, and for every one second window in the recording, the model outputs the probability that it contains seizure, periodic discharges (PD), rhythmic delta activity (RDA), or any other pattern (pathological, normal, non-biological noise, etc.). Our model is a 12-layer U-Net with roughly 3.4 million parameters. Model training is set up to accept sparse, missing, and multiple expert labels for any given segment, allowing for the utilization of datasets with diverse labeling methodologies. The model was trained and evaluated using studies drawn from 6428 subjects ( >40,000 labels from up to 10 human experts) selected from the Beacon Datastore™ and the Temple EEG dataset.Results:
The classifier model was able to achieve performance in line with consensus human expert opinion, as measured by true positive rate vs false positive rate. Model results are listed in Table 1. Figure 1 shows model performance relative to human experts, in addition to characteristics of model performance at varying sensitivity thresholds (ROC).
Conclusions: A multiclass classifier capable of identifying ictal and interictal patterns is feasible. We demonstrate a model trained on highly diverse labeled data, capable of classifying a continuum of pathological EEG patterns with performance matching human experts. The incorporation of ictal-interictal continuum patterns is of critical importance given their broad association with neurological risk – making automated identification of these diverse patterns a necessity in critical care medicine and a valuable tool for the computation of neurobiomarkers in development of therapies.
Funding:
This work was supported by Beacon Biosignals.