fMRI-based detection of alertness predicts behavioral response variability

Sarah E. Goodale, Nafis Ahmed, Chong Zhao, Jacco A. de Zwart, Pınar S. Özbay, Dante Picchioni, Jeff H. Duyn, Dario J. Englot, Victoria L. Morgan, Catie ChangeLife 2021 · DOI

TL;DR

This paper introduces a novel method to extract a continuous measure of alertness directly from fMRI data, eliminating the need for simultaneous EEG or pupillometry. This "fMRI alertness index" accurately predicts trial-to-trial behavioral responses and can even improve the detection of task-related brain activity, opening new avenues for analyzing existing fMRI datasets.

What Problem Does This Paper Address?

Human behavior and cognition are heavily influenced by our level of alertness, which naturally fluctuates. While fMRI is a powerful tool for studying brain activity during tasks, it typically lacks concurrent measures of alertness like EEG or pupillometry. This makes it challenging to account for how these internal state changes impact behavior and neural responses, especially in large existing datasets. The authors aim to bridge this gap by developing a method to infer alertness solely from fMRI data, enabling a more comprehensive understanding of brain-behavior relationships.

Methods

The core of the paper is the development of an "fMRI alertness index." This is achieved by:

  1. Template Creation: A spatial "template map" is created from resting-state fMRI data that is simultaneously recorded with EEG. This map identifies brain regions whose activity patterns are consistently correlated with an EEG-derived measure of alertness (specifically, the alpha/theta ratio).
  2. Alertness Index Extraction: This template map is then projected onto the fMRI data of new scans (even those without EEG). The spatial correlation at each time point (TR) yields a continuous time series representing the estimated "fMRI alertness index."
  3. Validation: The fMRI alertness index is validated by comparing it to simultaneously recorded EEG alertness measures and by examining its correlation with behavioral responses (hits/misses, reaction times) during an auditory task.
  4. Voxel Reduction: The authors explore the impact of using only a small fraction of the voxels from the template map, demonstrating that a highly reduced template (top 1% of positive and negative voxels) still retains significant predictive power.
  5. GLM Analysis: The fMRI alertness index is incorporated as a covariate in standard General Linear Models (GLMs) for task fMRI analysis to assess its impact on detecting task-activated brain areas.

Datasets

  • Training Data: Resting-state fMRI data with simultaneous EEG recordings were used to generate the "vigilance template."
  • Evaluation Data: Auditory task fMRI data (also with simultaneous EEG and behavioral responses) were used to evaluate the fMRI alertness index.
  • Cross-validation: The study employed a cross-validation strategy, splitting subjects into two groups to train the template on one group's resting-state data and test it on the other group's task data, and vice-versa, to assess generalizability.

Key Insights and Ideas

  • fMRI as a Window into Alertness: The most significant insight is that fMRI data alone, when analyzed with a specific spatial pattern, can reliably track dynamic changes in a subject's alertness. This bypasses the need for cumbersome physiological recordings during scans.
  • Predictive Power of Pre-Stimulus Alertness: Alertness levels measured just before a stimulus (within a 5-second window) are highly predictive of how a subject will respond to that stimulus, including whether they will detect it or how quickly they will react. This highlights the crucial role of internal state in sensory processing.
  • Low-Dimensionality of Alertness Signals: The finding that a very small subset of voxels (top 1%) is sufficient to capture most of the predictive power suggests that the neural basis of alertness is remarkably concentrated and not spread diffusely across the entire brain.
  • Enhancing Task Activation Detection: Including the fMRI alertness index as a covariate in standard fMRI analyses can improve the statistical detection of task-related brain activity, particularly in sensory and motor regions, and can also reduce widespread deactivations. This implies that some observed "deactivations" might be artifacts of fluctuating arousal rather than true neural suppression.

Key Results

  • The fMRI alertness index showed a strong temporal correlation (r ≈ 0.6-0.7) with simultaneously recorded EEG alertness measures.
  • Pre-stimulus fMRI alertness was significantly higher for "hit" trials (correct responses) compared to "miss" trials.
  • Fast hit trials were associated with higher pre-stimulus fMRI alertness compared to slow hit trials (p=0.001, Cohen's d=1.43).
  • Using a reduced template (top 1% positive and negative voxels) resulted in only a 5.7% decrease in performance (correlation with EEG alertness) compared to the full template.
  • Incorporating the fMRI alertness index as a covariate in GLM analyses increased t-scores in bilateral auditory cortex and other key regions, and reduced negative activation patterns.

Limitations and Future Work

  • Small Sample Size: The authors acknowledge that the study used a relatively small sample size (14 subjects), which may lead to overestimation of effect sizes and limit generalizability. Future work should involve larger cohorts and diverse experimental conditions.
  • Relative vs. Absolute Alertness: The current method primarily detects relative shifts in alertness within a scan, not absolute levels. The index is bounded between -1 and 1.
  • Scanner Artifacts: The detrending of fMRI signals to remove scanner artifacts might obscure slower baseline shifts in alertness.
  • Generalizability of Template: While cross-validation was performed, further investigation into the generalizability of the template across different tasks and populations is warranted.
  • Alternative Models: The paper explored a template-based approach and a lasso regression model. The authors suggest that other linear models and parcellation atlases could be explored for vigilance prediction.
  • Impact on Deactivations: The authors note that while including the alertness covariate reduced negative activation, it's unclear if this is solely due to accounting for arousal or if it might also remove genuine local neural effects. Future work should investigate this in various task contexts.

Summary updated

Abstract

Levels of alertness are closely linked with human behavior and cognition. However, while functional magnetic resonance imaging (fMRI) allows for investigating whole-brain dynamics during behavior and task engagement, concurrent measures of alertness (such as EEG or pupillometry) are often unavailable. Here, we extract a continuous, time-resolved marker of alertness from fMRI data alone. We demonstrate that this fMRI alertness marker, calculated in a short pre-stimulus interval, captures trial-to-trial behavioral responses to incoming sensory stimuli. In addition, we find that the prediction of both EEG and behavioral responses during the task may be accomplished using only a small fraction of fMRI voxels. Furthermore, we observe that accounting for alertness appears to increase the statistical detection of task-activated brain areas. These findings have broad implications for augmenting a large body of existing datasets with information about ongoing arousal states, enriching fMRI studies of neural variability in health and disease.
BibTeX
@inproceedings{goodale2021fmri,
  author = {Sarah E. Goodale and Nafis Ahmed and Chong Zhao and Jacco A. de Zwart and Pınar S. Özbay and Dante Picchioni and Jeff H. Duyn and Dario J. Englot and Victoria L. Morgan and Catie Chang},
  title = {{fMRI-based detection of alertness predicts behavioral response variability}},
  booktitle = {eLife},
  year = {2021},
  doi = {10.7554/elife.62376},
}
Figure