A Toolbox and Crowdsourcing Platform for Automatic Labeling of Independent Components in Electroencephalography

Independent Component Analysis (ICA) is a conventional approach to exclude non-brain signals such as eye movements and muscle artifacts from electroencephalography (EEG). A rejection of independent components (ICs) is usually performed in semiautomatic mode and requires experts’ involvement. As also revealed by our study, experts’ opinions about the nature of a component often disagree, highlighting the need to develop a robust and sustainable automatic system for EEG ICs classification. The current article presents a toolbox and crowdsourcing platform for Automatic Labeling of Independent Components in Electroencephalography (ALICE) available via link http://alice.adase.org/. The ALICE toolbox aims to build a sustainable algorithm to remove artifacts and find specific patterns in EEG signals using ICA decomposition based on accumulated experts’ knowledge. The difference from previous toolboxes is that the ALICE project will accumulate different benchmarks based on crowdsourced visual labeling of ICs collected from publicly available and in-house EEG recordings. The choice of labeling is based on the estimation of IC time-series, IC amplitude topography, and spectral power distribution. The platform allows supervised machine learning (ML) model training and re-training on available data subsamples for better performance in specific tasks (i.e., movement artifact detection in healthy or autistic children). Also, current research implements the novel strategy for consentient labeling of ICs by several experts. The provided baseline model could detect noisy IC and components related to the functional brain oscillations such as alpha and mu rhythm. The ALICE project implies the creation and constant replenishment of the IC database, which will improve ML algorithms for automatic labeling and extraction of non-brain signals from EEG. The toolbox and current dataset are open-source and freely available to the researcher community.

A supervised machine learning semantic segmentation approach for detecting artifacts in plethysmography signals from wearables

Objective. Wearable devices equipped with plethysmography (PPG) sensors provided a low-cost, long-term solution to early diagnosis and continuous screening of heart conditions. However PPG signals collected from such devices often suffer from corruption caused by artifacts. The objective of this study is to develop an effective supervised algorithm to locate the regions of artifacts within PPG signals. Approach. We treat artifact detection as a 1D segmentation problem. We solve it via a novel combination of an active-contour-based loss and an adapted U-Net architecture. The proposed algorithm was trained on the PPG DaLiA training set, and further evaluated on the PPG DaLiA testing set, WESAD dataset and TROIKA dataset. Main results. We evaluated with the DICE score, a well-established metric for segmentation accuracy evaluation in the field of computer vision. The proposed method outperforms baseline methods on all three datasets by a large margin (≈ 7 percentage points above the next best method). On the PPG DaLiA testing set, WESAD dataset and TROIKA dataset, the proposed method achieved 0.8734±0.0018, 0.9114±0.0033 and 0.8050±0.0116 respectively. The next best method only achieved 0.8068±0.0014, 0.8446±0.0013 and 0.7247±0.0050. Significance. The proposed method is able to pinpoint exact locations of artifacts with high precision; in the past, we had only a binary classification of whether a PPG signal has good or poor quality. This more nuanced information will be critical to further inform the design of algorithms to detect cardiac arrhythmia.

P.128 Quantitative EEG Detects REM Sleep Microstructure

Background: Rapid eye movement sleep (REM) is divided into phasic and tonic microstates. Phasic REM is defined by presence of REMs with reportedly greater antiepileptic effect. We assessed whether quantitative EEG (QEEG) software can detect REM microstates. Methods: We applied artifact reduction and detection trends from QEEG software (Persyst 14) on 18 patients undergoing 30 day-night high density EEG recordings in the epilepsy monitoring unit. We identified phasic REM as 10-second epochs of previously human-scored REM that demonstrated presence of either vertical or horizontal eye movements on the QEEG artifact detection panel. Remaining epochs were identified as tonic REM. Results: Out of 91.2 average minutes of REM (range 24.5-167.5) per recording, a mean of 2.5% (range 0-18.9%) demonstrated eye movements intensive enough for QEEG artifact detection to be identified as phasic REM. On average, only 40% (range 0-500%) of eye movements per recording was flagged as vertical. Conclusions: These findings provide proof-of-concept that QEEG can automatically assess REM microstructure by readily detecting phasic and tonic REM. These findings also confirm that most REMs are horizontal. Having the ability to easily and automatically detect phasic versus tonic REM can help further future studies examining the antiepileptic effect of REM sleep.

P.021 Electdetect: An Artificial Neural Network for the Detection of Artifacts in intracranial EEG Recordings From Patients with Epilepsy

Background: Intracranial electroencephalography (iEEG) recordings are obtained from the sampling of sub-cortical structures and provide extraordinary insight into the spatiotemporal dynamics of the brain. As these recordings are increasingly obtained at higher channel counts and greater sampling frequencies, preprocessing through visual inspection is becoming untenable. Consequently, artificial neural networks (ANNs) are now being leveraged for this task. Methods: One-hour recordings from six patients diagnosed with drug-resistant epilepsy at Toronto Western Hospital were obtained alongside fiduciary ECG and EOG activity. R-wave peaks and local maxima were identified in the ECG and EOG recordings, respectively, and were time-mapped onto the iEEG recordings to delimit one-second epochs around 1.6 million cardiac and 600 thousand ocular artifacts. Epochs were then split into train-test-evaluation sets and fed into an ANN as one-second spectrograms (0 - 1,000 Hz) over 30-time steps. Results: The ANN model achieved formidable classification results on the evaluation set with an F1, positive predictive value, and sensitivity scores of 0.93. Furthermore, model architecture computed the classification probability at each time-step and enabled insight into the spatiotemporal features driving classification. Conclusions: We expect this research to promote the public sharing of new ANN from multiple institutions and enable novel automated algorithms for artifact detection in iEEG recordings.

An Unsupervised Machine Learning Paradigm for Artifact Removal from Electrodermal Activity in an Uncontrolled Clinical Setting

Objective: Electrodermal activity (EDA) reflects sympathetic nervous system activity through sweating-related changes in skin conductance. To enable EDA data to be used robustly in clinical settings, we need to develop artifact detection and removal frameworks that can handle the types of interference experienced in clinical settings. Methods: We collected EDA data from 69 subjects while they were undergoing surgery in the operating room. We then built an artifact removal framework using unsupervised learning methods and informed features to remove the heavy artifact that resulted from the use of surgical electrocautery during the surgery and compared it to other existing methods for artifact removal from EDA data. Results: Our framework was able to remove the vast majority of artifact from the EDA data across all subjects with high sensitivity (94%) and specificity (90%). In contrast, existing methods used for comparison struggled to be sufficiently sensitive and specific, and none effectively removed artifact even if it was identifiable. In addition, the use of unsupervised learning methods in our framework removes the need for manually labeled datasets for training. Conclusion: Our framework allows for robust removal of heavy artifact from EDA data in clinical settings such as surgery. Since this framework only relies on a small set of informed features, it can be expanded to other modalities such as ECG and EEG. Significance: Robust artifact removal from EDA data is the first step to enable clinical integration of EDA as part of standard monitoring in settings such as the operating room.