Adversarial AI uncovers hidden brain mechanisms and potential treatments for disorders of consciousness

Adversarial AI reveals mechanisms and treatments for disorders of consciousness - Nature Neuroscience

Understanding disorders of consciousness (DOC) remains one of the most challenging problems in neuroscience, hindered by the lack of experimental models for probing mechanisms or testing interventions. Here, to address this, we introduce a generative adversarial artificial intelligence (AI) framework that pits deep neural networks -- trained to detect consciousness across more than 680,000 ten-second neuroelectrophysiology samples and validated on 565 patients, healthy volunteers and animals -- against interpretable, machine learning-driven neural field models. This adversarial architecture produces biologically realistic simulations of both conscious and comatose brains that recapitulate empirical neurophysiological features across humans, monkeys, rats and bats. Without explicit programming, the AI model retrodicts known DOC responses to brain stimulation and generates testable predictions about the mechanisms of unconsciousness. Two such predictions are validated here: selective disruption of the basal ganglia indirect pathway, supported by diffusion magnetic resonance imaging in 51 patients with DOC, and increased cortical inhibitory-to-inhibitory synaptic coupling, supported by RNA sequencing of resected brain tissue from 6 human patients with coma and a rat stroke model. The model also identifies high-frequency stimulation of the subthalamic nucleus as a promising intervention for DOC, supported by electrophysiological data from human patients. This work introduces an AI framework for causal inference and therapeutic discovery in consciousness research, as well as in complex systems more broadly.

Dueling AI agents could reveal keys to restoring consciousness

Can a "friendly" rivalry between two artificial intelligence (AI) agents help reveal how the brain supports consciousness? That's the suggestion coma researcher Martin Monti and his colleagues at the University of California, Los Angeles make in a paper published today in Nature Neuroscience. One of their two AI models generated realistic imitations of electrical patterns seen in conscious and unconscious brain states, from wakefulness to deep comas. Its counterpart had to identify these states. The results largely support established ideas about how the brain behaves during comas, vegetative states, and other disorders of consciousness. But they also suggest roles for a brain structure and a pattern of cell signaling not previously known to be involved in such disorders -- predictions the scientists were able to test. Monti spoke with Science about how the paper's two models, which he calls the "black box" and the "glass brain," could reveal new ways to restore consciousness after brain injury. This interview has been edited for clarity and length. Q: You built two AI models, with one designed to interrogate the other. Can you explain how they talk to each other? A: So here's the game: We have two friends. One -- let's call it the black box -- knows how to tell consciousness from unconsciousness. It's been trained on 680,000 snippets of EEG [electroencephalography] data from animals and people in different states of consciousness. The other -- think of it as a glass brain -- is a real, biologically plausible simulation of the human brain. We tell it, "Your job is to move all of your knobs, every single parameter you've got, to trick the other guy -- the black box -- to think that you're creating a real EEG of a conscious or unconscious state." Now, we ask the glass brain, "Which brain parameters made the box think the EEG was unconscious?" Q: And what did that game reveal? A: The simulated brain created authentically unconscious-looking EEGs. I work a lot with clinicians. I see a lot of clinical data. They look like things that we knew and that we expected. And then, of course, the question is, how did you tweak your knobs to get these things? And we got a lot of parameters that we'd expected. But there were two novel things that we did not expect, and those were just not on anybody's bingo card. One was the role of a part of the basal ganglia called the globus pallidus externa. The less connectivity between this and [another basal ganglia structure called] the striatum, the more likely the EEG produced was an unconsciousness EEG. We confirmed this with imaging data from people with disorders of consciousness. The second finding is that there should be, in unconsciousness, more coupling between inhibitory neurons -- neurons that restrain the firing of other neurons. We were lucky enough to be able to validate this using data from another group comparing human brain tissue from individuals who died in a coma versus individuals who did not. Q: Your model also predicted that stimulating one brain region -- called the subthalamic nucleus -- might trigger awakening from unconsciousness. Is there evidence for that in people? A: Nobody has ever done deep-brain stimulation [DBS] in disorders of consciousness to the subthalamic nucleus. But Daniel Toker, the first author on the study, found a study of people with an implanted DBS device for cervical dystonia, a type of neck spasm. Some patients had stimulation to the subthalamic nucleus. We fed their EEG data to our neural network and we said, "Hey, give us a consciousness score for these EEGs." They were conscious to begin with, but the neural network scored them higher after stimulation. In patients who are not conscious, is stimulation going to wake them up? I don't know. Q: There's only one way to find out. A: There's only one way. And we're trying to set up a clinical trial for that. Q: What other novel things could this method tell us about consciousness? A: Maybe this is sci-fi, but we can look at other species now. Before, we could look at their EEGs and wonder about their degree of consciousness, but now we can also look at what mechanistically is going on. Q: You also suggest that these findings are relevant for other neurological and psychiatric disorders. Could you create the same kind of opposing AI models to learn more about what the brain is doing in those cases? A: Absolutely. It would be exactly the same. You would first train a network to recognize the features, let's say EEGs from individuals experiencing depression. And healthy volunteers. And EEGs from other mood states. I sincerely hope that people will apply this method to something else and discover something that was not already known about some other condition. I would be so proud if that happened.