Unsupervised machine-learning classification of electrophysiologically active electrodes during human cognitive task performance

Krishnakant Saboo, Yogatheesan Varatharajah, Brent Berry, Václav Křemen, Michael R. Sperling, Kathryn A. Davis, Barbara C. Jobst, Robert E. Gross, Bradley Lega, Sameer A. Sheth, Gregory A. Worrell, Ravishankar K. Iyer, Michał T. KucewiczScientific 2019 · DOI

TL;DR

This paper introduces an unsupervised, fully automated machine learning framework to identify "active" electrodes in intracranial EEG (iEEG) data that capture task-relevant brain activity. By developing novel spectral metrics and employing clustering, the method achieves high accuracy in distinguishing active electrodes, paving the way for more objective and efficient analysis of brain function during cognitive tasks.

What Problem Does This Paper Address?

Understanding brain function relies heavily on analyzing electrophysiological signals, particularly from intracranial EEG (iEEG) recordings, which offer high spatial resolution. However, not all implanted electrodes capture meaningful activity related to a specific cognitive task. Identifying these "active" electrodes is crucial for both clinical applications (e.g., localizing seizure onset zones) and research into cognitive processes like memory. Traditional methods often involve manual or semi-automatic selection, which can be time-consuming, subjective, and prone to inconsistencies between experts. This paper aims to bridge this gap by developing an automated, unsupervised approach that objectively identifies active electrodes based on their spectral characteristics during a cognitive task.

Methods

The core of this work lies in a three-step unsupervised machine learning approach:

  1. Normalization: The iEEG data from multiple electrodes and subjects are normalized. This step is crucial for comparing signals across different individuals and electrodes.
  2. Metric Computation: Novel, interpretable metrics are derived from the normalized iEEG signal's spectrum. These metrics quantify:
  3. Induced Power (IP): The sum of absolute mean band power changes across time in specific frequency bands (e.g., theta, gamma). This captures the magnitude of task-related spectral changes.
  4. Smoothness Score (SS): The 1-lag autocorrelation of the mean band power change in low- and high-theta bands. This metric assesses the temporal smoothness of spectral changes, hypothesizing that active electrodes will show more consistent patterns.
  5. Gamma Consistency (GC): Pairwise absolute cross-correlation between the mean band power changes in different gamma frequency ranges (low gamma, high gamma 1, high gamma 2). This metric captures the coordination of gamma activity across frequencies, expected to be higher in active electrodes.
  6. Unsupervised Clustering: A Gaussian Mixture Model (GMM) with two components is used to cluster electrodes based on the computed metric values. The smaller cluster is identified as the "active" electrode group. This approach avoids the need for pre-labeled "ground truth" data for training.

Datasets

The study utilized a large public dataset of intracranial recordings from 115 patients who underwent iEEG monitoring as part of clinical treatment for drug-resistant epilepsy. These patients performed a delayed free-recall verbal memory task. A total of 11,869 electrodes were analyzed. The "ground truth" for evaluating the method was established through a semi-automatic approach involving manual inspection by experts on a subset of these recordings.

Key Insights and Ideas

  • Spectral Signatures of Activity: The research highlights that active electrodes exhibit distinct spectral signatures during cognitive tasks, characterized by increased high-gamma power and suppression of low-frequency power. These signatures are quantifiable and can be used to differentiate active from inactive electrodes.
  • Interpretable Metrics: The development of metrics like Induced Power, Smoothness Score, and Gamma Consistency provides interpretable ways to characterize electrode activity. This interpretability is key for understanding why an electrode is classified as active, linking spectral features to physiological processes.
  • Unsupervised Power: The unsupervised nature of the clustering approach is a significant advantage, reducing reliance on subjective manual labeling and enabling objective classification across diverse subjects and electrodes.
  • Anatomical Validation: The finding that classified active electrodes are not randomly distributed but are concentrated in brain regions known to support verbal memory processing (e.g., hippocampus, prefrontal cortex) provides strong anatomical validation for the method.

Key Results

  • The proposed unsupervised method achieved 97% sensitivity and 92.9% specificity with the most efficient metric (Induced Power in high-gamma band).
  • The Induced Power (IP) metric, particularly for high-gamma frequencies, demonstrated the highest specificity (92.9%), while gamma consistency (GC) and IP also showed high sensitivity.
  • Anatomical localization revealed that active electrodes were significantly more prevalent in brain regions associated with verbal memory processing, such as the left ventrolateral prefrontal cortex and the left occipito-temporal cortex junction.
  • The method outperformed standard machine learning classifiers like LDA and SVM when using mean band power change as input, achieving higher sensitivity and comparable AUC scores in leave-one-subject-out cross-validation.

Limitations and Future Work

  • Sampling Frequency Dependence: The Induced Power and Smoothness Score metrics are sensitive to sampling frequency. While the data was resampled to 500 Hz, the authors acknowledge that careful consideration of sampling frequency is necessary when applying the framework.
  • Generalizability: The current method was developed and tested on a specific cognitive task (verbal memory). Future work should explore its generalizability to other cognitive tasks, motor tasks, and perceptual tasks.
  • Subjectivity in "Ground Truth": While the method is unsupervised, the "ground truth" used for validation was derived from a semi-automatic, expert-driven process, which itself has inherent limitations.
  • Boundary Cases: The authors note that their method is particularly good at detecting electrodes that are clearly active or inactive, and may be better at handling electrodes at the boundary between clusters compared to other methods.
  • Future Directions: The authors suggest incorporating other features, such as information-theoretic criteria, and exploring the impact of active electrode selection on downstream classification tasks (e.g., classifying successful vs. unsuccessful memory trials). They also propose investigating the clinical utility of the method for determining language dominance.

Summary updated

Abstract

Identification of active electrodes that record task-relevant neurophysiological activity is needed for clinical and industrial applications as well as for investigating brain functions. We developed an unsupervised, fully automated approach to classify active electrodes showing event-related intracranial EEG (iEEG) responses from 115 patients performing a free recall verbal memory task. Our approach employed new interpretable metrics that quantify spectral characteristics of the normalized iEEG signal based on power-in-band and synchrony measures. Unsupervised clustering of the metrics identified distinct sets of active electrodes across different subjects. In the total population of 11,869 electrodes, our method achieved 97% sensitivity and 92.9% specificity with the most efficient metric. We validated our results with anatomical localization revealing significantly greater distribution of active electrodes in brain regions that support verbal memory processing. We propose our machine-learning framework for objective and efficient classification and interpretation of electrophysiological signals of brain activities supporting memory and cognition.
BibTeX
@inproceedings{saboo2019unsupervised,
  author = {Krishnakant Saboo and Yogatheesan Varatharajah and Brent Berry and Václav Křemen and Michael R. Sperling and Kathryn A. Davis and Barbara C. Jobst and Robert E. Gross and Bradley Lega and Sameer A. Sheth and Gregory A. Worrell and Ravishankar K. Iyer and Michał T. Kucewicz},
  title = {{Unsupervised machine-learning classification of electrophysiologically active electrodes during human cognitive task performance}},
  booktitle = {Scientific},
  year = {2019},
  doi = {10.1038/s41598-019-53925-5},
}
Figure