Predicting task performance from biomarkers of mental fatigue in global brain activity

Lin Yao, Jonathan Baker, Nicholas D. Schiff, Keith P. Purpura, Mahsa ShoaranJournal of Neural Engineering 2020 · DOI

TL;DR

This paper demonstrates that electrocorticography (ECoG) signals from non-human primates can be used to predict mental fatigue and task performance decrements before they occur. By analyzing novel spectrotemporal and connectivity biomarkers, researchers achieved high accuracy in predicting incorrect trials, paving the way for adaptive neuromodulation therapies for conditions involving chronic fatigue.

What Problem Does This Paper Address?

Mental fatigue, characterized by a decline in vigilance and attention, significantly impairs performance on cognitive tasks. This is particularly problematic for individuals with brain injuries, who often experience chronic fatigue and impaired cognitive function. Current treatments lack the ability to dynamically adapt to these fluctuating states. The core challenge is to reliably detect the onset of mental fatigue in real-time, using neural signals, to enable timely interventions. This paper aims to bridge this gap by identifying predictive neural biomarkers from ECoG data.

Methods

The researchers employed a machine learning approach using electrocorticography (ECoG) signals recorded from two non-human primates (NHPs) performing a sustained attention task. They extracted a range of spectrotemporal and connectivity features from the ECoG data, including:

  • Conventional spectral power features: Standard measures of brain activity in different frequency bands.
  • Higher-Order Spectral-Temporal (HOST) features: Novel biomarkers such as wavelet entropy (measuring signal order/disorder), Hjorth parameters (time-domain statistical properties), phase-amplitude coupling (PAC), and instantaneous amplitude and frequency (IA/IF ratio).
  • Connectivity measures: Partial directed coherence (PDC) and Phase Locking Index (PLI) to assess inter-areal neural interactions.

These features were then fed into a gradient boosting decision tree ensemble (XGBoost) classifier, trained to distinguish between correct and incorrect trials. A nested 5-fold cross-validation was used for robust evaluation, and a wrapper-based approach was employed for feature selection to identify the most discriminative biomarkers. Independent Component Analysis (ICA) was also used to mitigate the influence of eye movement artifacts.

Datasets

The study utilized ECoG data collected from two healthy adult male macaque monkeys (Macaca mulatta) performing a visuomotor reaction-time vigilance task. A total of 99 experimental sessions were analyzed (55 for NHP1 and 44 for NHP2). The data was collected during non-deep brain stimulation (DBS) periods. No new benchmarks were introduced; rather, the study aimed to establish a predictive model using existing experimental paradigms.

Key Insights and Ideas

  • Beyond Spectral Power: The study strongly suggests that "higher-order" spectrotemporal features (HOST) are significantly more effective than traditional spectral power measures for predicting mental fatigue and task performance. This highlights the importance of looking at more complex signal dynamics.
  • Early Prediction is Feasible: The ability to predict incorrect trials several seconds before the behavioral response is a critical finding. This temporal lead time is crucial for any closed-loop intervention system.
  • Global Brain Activity Matters: Analyzing ECoG signals from across the cortical surface, rather than just localized regions, provided a more comprehensive picture and improved predictive power.
  • Eye Movements as a Signal (with caveats): While eye movement artifacts can contaminate ECoG signals, the study found that these signals, after ICA processing, could still contribute to predicting performance. This suggests that even "noise" can sometimes carry relevant information about cognitive state.
  • Feature Selection is Key: The use of a wrapper-based approach identified a small subset of optimal features (5-10) that were sufficient for high predictive accuracy, demonstrating efficiency and potentially reducing computational load for real-time applications.

Key Results

  • High Predictive Accuracy: The XGBoost classifier achieved high F1 scores and accuracy in predicting incorrect trials.
  • NHP1: Average F1 score of 77.5% ± 8.2%, Accuracy of 79.5% ± 8.9%.
  • NHP2: Average F1 score of 91.2% ± 3.6%, Accuracy of 87.6% ± 3.9%.
  • HOST Features Outperform: HOST features significantly outperformed conventional spectral power features for both animals (p<0.001 for accuracy).
  • Optimal Time Window: The delay period of the task provided the most predictive samples of brain activity.
  • Few Features Needed: A small set of 5-10 features was sufficient to achieve near-optimal classification performance.
  • Wavelet Entropy and IA/IF: Wavelet entropy and instantaneous amplitude/frequency (IA/IF) were identified as particularly discriminative single-channel features. Connectivity measures like PDC and PLI were also frequently selected.

Limitations and Future Work

  • Subject-Specific Models: The current study focused on subject-specific prediction. Generalizing the approach to a larger population of animals and eventually humans is a key future direction.
  • Translation to Humans: The authors acknowledge that translating these findings to human patients will require further studies to identify biomarkers relevant to impaired cognition and executive function in clinical populations.
  • Real-time Application: While the study demonstrates feasibility in an offline setting, the ultimate goal is real-time application for closed-loop control of neuromodulation strategies.
  • Artifact Handling: The study highlights the general challenge of dealing with "noisy" ECoG/EEG data, particularly movement artifacts, and suggests that further refinement of artifact removal techniques is important.

Summary updated

Abstract

<i>Objective.</i>Detection and early prediction of mental fatigue (i.e. shifts in vigilance), could be used to adapt neuromodulation strategies to effectively treat patients suffering from brain injury and other indications with prominent chronic mental fatigue.<i>Approach.</i>In this study, we analyzed electrocorticography (ECoG) signals chronically recorded from two healthy non-human primates (NHP) as they performed a sustained attention task over extended periods of time. We employed a set of spectrotemporal and connectivity biomarkers of the ECoG signals to identify periods of mental fatigue and a gradient boosting classifier to predict performance, up to several seconds prior to the behavioral response.<i>Main results.</i>Wavelet entropy and the instantaneous amplitude and frequency were among the best single features across sessions in both NHPs. The classification performance using higher order spectral-temporal (HOST) features was significantly higher than that of conventional spectral power features in both NHPs. Across the 99 sessions analyzed, average F1 scores of 77.5% ± 8.2% and 91.2% ± 3.6%, and accuracy of 79.5% ± 8.9% and 87.6% ± 3.9% for the classifier were obtained for each animal, respectively.<i>Significance.</i>Our results here demonstrate the feasibility of predicting performance and detecting periods of mental fatigue by analyzing ECoG signals, and that this general approach, in principle, could be used for closed-loop control of neuromodulation strategies.
BibTeX
@article{yao2020predicting,
  author = {Lin Yao and Jonathan Baker and Nicholas D. Schiff and Keith P. Purpura and Mahsa Shoaran},
  title = {{Predicting task performance from biomarkers of mental fatigue in global brain activity}},
  journal = {Journal of Neural Engineering},
  year = {2020},
  doi = {10.1088/1741-2552/abc529},
}
Figure