TL;DR
This study explores how closed-loop brain-machine interfaces (BMIs) can be used to modulate pain perception. They found that while participants could learn to influence the BMI to reduce pain, this process involved complex endogenous brain responses, including a paradoxical decrease in pain signal decodability in some brain regions and an increased role for the pregenual anterior cingulate cortex (pgACC).
What Problem Does This Paper Address?
The ultimate goal of closed-loop BMIs for pain management is to dynamically adjust interventions based on real-time brain activity related to pain. However, pain is notoriously susceptible to subjective modulation by the brain itself. This raises a crucial question: could the very act of using a BMI to decode pain signals inadvertently alter those signals, either by enhancing or interfering with them? This paper investigates this "endogenous controllability" by examining how participants' attempts to influence a real-time fMRI-based pain modulation system affect both the neural signals and their subjective pain experience.
Methods
The researchers developed an adaptive brain-machine interface system that used real-time fMRI (rtfMRI) to decode pain intensity. This decoded signal then controlled an electrical pain stimulator.
- Decoder Construction (Day 1): Participants received high and low intensity electrical stimuli. fMRI data was collected to train a multivariate pattern analysis (MVPA) decoder to distinguish between these pain levels, primarily using the insula cortex.
- Neurofeedback Adaptive Control (Day 2): The trained decoder was used in real-time. A reinforcement learning (RL) algorithm controlled the pain stimulator, aiming to minimize the decoded pain signal. Participants were aware that their brain activity influenced the system, creating an incentive to modulate their neural signals.
- Whole-Brain Analysis: Beyond the primary ROI (insula), searchlight analyses were used to examine brain-wide changes in pain representation.
- Frequency Learning Model: A Bayesian learning model was employed to assess how participants learned the probability of receiving high or low pain and the associated uncertainty (entropy).
Datasets
The study used fMRI data collected from 19 healthy participants over two days of experimental sessions. No external datasets were used for training or evaluation.
Key Insights and Ideas
- Paradoxical Decodability: Despite the incentive to "help" the BMI by making pain signals clearer, decodability of pain intensity actually decreased or remained unchanged in key pain-processing regions like the insula and somatosensory cortex. This suggests the brain isn't simply "boosting" the signal.
- pgACC as a Hub: The pregenual anterior cingulate cortex (pgACC) was the only region showing enhanced decodability of pain signals. This region is implicated in pain modulation and endogenous control, suggesting it plays a critical role in this adaptive process.
- Uncertainty and Pain: As participants learned the success rate of the BMI, a signal of computational uncertainty (entropy) emerged. This uncertainty was positively correlated with subjective pain ratings, indicating that when the system's performance was less predictable, pain perception was higher.
- Co-adaptive Learning: The results highlight a "co-adaptive" process where both the brain and the BMI learn and adjust in response to each other. The brain learns to modulate its signals, and the BMI learns to interpret them and adjust its output.
Key Results
- Decoder Performance: The MVPA decoder achieved reasonable classification accuracy (around 65% on Day 1, 56% on Day 2). While accuracy slightly decreased on Day 2, the number of voxels contributing to decoding increased.
- Pain Reduction: The RL-controlled system successfully learned to deliver significantly fewer high-pain stimuli on Day 2 compared to Day 1.
- Brain Region Changes:
- Decreased BOLD responses were observed in the amygdala and putamen on Day 2 compared to Day 1.
- Increased responses were seen in the periaqueductal gray (PAG) on Day 2.
- Searchlight analysis revealed decreased decoding accuracy in the left insula and DLPFC/MFG, and increased accuracy in the pgACC on Day 2.
- Subjective Reports: 17 out of 19 participants believed the machine was successful in reading their pain and trying to reduce it, and 15 believed they were successful in influencing the system using mental strategies.
Limitations and Future Work
- Novelty/Anxiety Confound: The sequential nature of the experiment (Day 1 followed by Day 2) introduces a potential order confound related to novelty and anxiety. While analyses attempted to mitigate this, it remains a consideration.
- ROI Choice: The study arbitrarily used a single ROI (bilateral insula) for decoding, which might not be optimal. Future work could explore multi-ROI classifiers for higher accuracy.
- Simplicity of RL: The RL algorithm used was relatively simple. More complex policies could potentially lead to more sophisticated value learning.
- Extended Periods: Brain representations might change over longer periods of neurofeedback, a phenomenon not fully explored in this study.
- Future Directions: The authors suggest that the RL approach has broad applicability for therapeutic interventions where optimal parameters are unknown. They also highlight the potential for "neural conditioning" over extended periods of feedback. Future research could explore more complex learning models and longer-term interventions.
Summary updated