Closed-loop fMRI neurofeedback to reduce negative attentional bias in depression

Anne C. Mennen, Nicholas B. Turk‐Browne, Grant Wallace, Darsol Seok, Adna Jaganjac, Janet Stock, Megan T. deBettencourt, Jonathan D. Cohen, Kenneth A. Norman, Yvette I. ShelinebioRxiv 2020 · DOI

TL;DR

This paper introduces a novel closed-loop fMRI neurofeedback system designed to retrain attentional biases in individuals with depression. By using real-time fMRI data and machine learning, the system helps participants learn to disengage from negative stimuli, leading to reduced negative attentional bias and decreased depression severity.

What Problem Does This Paper Address?

Depression is often characterized by an attentional bias towards negative stimuli and thoughts, a phenomenon known as rumination. While traditional cognitive training methods like Attention Bias Modification (ABM) have shown promise, their transfer to clinical outcomes has been inconsistent. A key limitation of previous ABM studies is the use of pre-programmed training schedules, which may not adapt to individual responses. This paper addresses the need for more adaptive and sensitive interventions that can directly target the neural underpinnings of this negative attentional bias in depression.

Methods

The core of this study is a closed-loop real-time fMRI (rt-fMRI) neurofeedback system. Here's a breakdown of the key methods:

  • Real-time fMRI Data Acquisition and Processing: fMRI data was acquired during a go/no-go task where participants had to attend to one category (scenes) and ignore another (faces). Machine learning classifiers were trained in real-time to decode the participant's attentional state based on their brain activity.
  • Closed-Loop Neurofeedback: The decoded attentional state was "externalized" by visually manipulating the opacity of stimuli on the screen. If the system detected the participant attending to negative faces, those faces would become more visible, making the task harder and prompting the participant to shift their attention. This adaptive feedback mechanism is the "closed-loop" aspect.
  • Machine Learning Classification: A multi-voxel pattern classifier (specifically, logistic regression from scikit-learn) was used to decode the extent to which attention was directed towards scenes versus faces. This classifier was retrained for each neurofeedback run to adapt to individual brain activity patterns.
  • Attentional State Persistence Measure: To quantify the difficulty in disengaging from negative states, the researchers developed a measure of "attentional state persistence." This calculated the conditional probability that an attentional state (e.g., being in the "most negative state") would persist over time.
  • Cloud-Based Processing: A significant methodological innovation is the use of cloud-based processing for real-time analysis of fMRI data. This approach aims to overcome the technical bottlenecks and hardware requirements typically associated with rt-fMRI, making it more accessible.

Datasets

The study utilized a dataset of grayscale face and scene images adapted from existing databases, including the Chicago Face Database, Karolinska directed emotional faces, and NimStim. No new benchmarks were introduced; rather, existing stimuli were adapted for the task.

Key Insights and Ideas

  • Neural Measures are More Sensitive: The study found that while behavioral performance on the go/no-go task showed no group differences between depressed and healthy individuals, neural measures did reveal differences in attentional bias. This highlights the value of rt-fMRI for capturing subtle cognitive differences.
  • "Stuck" in Negative States: Depressed individuals showed a significantly greater tendency to remain "stuck" in negative attentional states, particularly at the beginning of training. This provides a neural correlate for rumination.
  • Adaptive Feedback Works: The closed-loop neurofeedback system, which dynamically adjusted stimulus visibility based on real-time brain activity, was effective in reducing this tendency to get stuck in negative states.
  • Cloud Computing for Accessibility: The successful implementation of real-time fMRI analysis via cloud processing is a key innovation, potentially democratizing access to this advanced neurofeedback technology.

Key Results

  • Reduced Negative Attentional Bias: Depressed participants showed a significant decrease in their probability of staying in the most negative attentional state after neurofeedback training, relative to healthy controls.
  • Decreased Depression Severity: Depression severity, as measured by the MADRS scale, significantly decreased from pre-training to post-training and at follow-up assessments in the depressed group.
  • Correlation between Bias Reduction and Symptom Improvement: A marginally significant positive correlation was found between the reduction in getting stuck in negative attentional states and the reduction in depression severity within the MDD group.
  • No Initial Behavioral Differences: Before training, there were no significant differences in behavioral performance on the go/no-go task between the depressed and healthy control groups.

Limitations and Future Work

  • Generalizability of Cloud Processing: While the cloud processing approach is promising, the paper notes that it was implemented on a secure network. Future work could involve extending this framework to a more widely accessible cloud-based web server.
  • Specificity of Individualized Feedback: The authors suggest future work to verify that the observed clinical benefits are specific to the individualized nature of the neurofeedback. This could involve control groups receiving yoked feedback or feedback from irrelevant ROIs.
  • Efficacy of Negative vs. Positive Training: The study focused on training away from negative stimuli. Future research could compare the relative efficacy of training away from negative stimuli versus training towards positive stimuli.
  • ROI-Based vs. Cognitive State Training: The study trained a decoded cognitive state, differing from other rt-fMRI studies that focus on mean ROI activity. Further research is needed to compare the efficacy of these different approaches.

Summary updated

Abstract

Abstract Depressed individuals show an attentional bias toward negatively valenced stimuli and thoughts. Here we present a novel closed-loop neurofeedback procedure that seeks to remediate this bias. Internal attentional states were detected by applying machine learning techniques to fMRI data in real-time, and externalized using a visually presented stimulus that the participant could learn to control. We trained 15 depressed and 12 healthy control participants over three fMRI sessions, preceded and followed by behavioral and clinical assessments. Initially, depressed participants were more likely than non-depressed participants to get “stuck” in negative attentional states, but this diminished with neurofeedback training relative to controls. Depression severity also decreased from pre- to post-training. These results demonstrate that our method is sensitive to the negative attentional bias in depressed individuals, and its reduction after training showcases the potential of this method as a treatment in the future.
BibTeX
@misc{mennen2020closed,
  author = {Anne C. Mennen and Nicholas B. Turk‐Browne and Grant Wallace and Darsol Seok and Adna Jaganjac and Janet Stock and Megan T. deBettencourt and Jonathan D. Cohen and Kenneth A. Norman and Yvette I. Sheline},
  title = {{Closed-loop fMRI neurofeedback to reduce negative attentional bias in depression}},
  year = {2020},
  doi = {10.1101/2020.06.07.137943},
  howpublished = {bioRxiv},
}

Loading related content...

Figure