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

Dueling AI Agents Simulate Brain States to Decode Consciousness

A team led by coma researcher Martin Monti at UCLA has introduced a generative adversarial artificial intelligence framework that uses dueling AI agents to investigate disorders of consciousness, one of neuroscience's most challenging problems 1

1

. Published in Nature Neuroscience, the research trained deep neural networks on more than 680,000 ten-second neuroelectrophysiology samples from 565 patients, healthy volunteers, and animals to detect conscious and unconscious brain states 1

1

.

The adversarial architecture features two AI models working in opposition. The first, dubbed the "black box," learned to distinguish consciousness from unconsciousness using EEG data from animals and people in different states 2

2

. The second, called the "glass brain," generates biologically realistic simulations of brain activity by adjusting parameters to trick its counterpart into identifying the generated patterns as real conscious or comatose states 2

2

. This approach produces authentic-looking unconscious EEGs that recapitulate empirical neurophysiological features across humans, monkeys, rats, and bats 1

1

.

Novel Predictions Validated Through Patient Data

Without explicit programming, the adversarial AI retrodicted known responses to brain stimulation in patients with disorders of consciousness and generated testable predictions about mechanisms of unconsciousness 1

1

. Two unexpected findings emerged from analyzing which parameters the glass brain adjusted to create unconscious-looking patterns. The first involved the basal ganglia, specifically showing that reduced connectivity between the globus pallidus externa and the striatum correlated with unconscious EEG patterns 2

2

. This prediction was validated using diffusion magnetic resonance imaging data from 51 patients with disorders of consciousness 1

1

.

The second discovery revealed increased coupling between inhibitory neurons during unconscious states 1

1

. These inhibitory neurons normally restrain the firing of other neurons, and their enhanced interaction appears to play a role in maintaining coma states 2

2

. Researchers confirmed this mechanism through RNA sequencing of resected brain tissue from 6 human patients who died in coma and a rat stroke model 1

1

.

Brain Stimulation Emerges as Promising Intervention

The model identified high-frequency stimulation of the subthalamic nucleus as a potential intervention to treat disorders of consciousness 1

1

. While no one has performed deep-brain stimulation targeting the subthalamic nucleus specifically for restoring consciousness, first author Daniel Toker found supporting evidence in an unexpected place 2

2

. Patients with cervical dystonia who received deep-brain stimulation to the subthalamic nucleus showed higher consciousness scores on EEG data when analyzed by the neural network, even though they were already conscious 2

2

. Whether this stimulation could wake unconscious patients remains unknown, but Monti's team is working to establish a clinical trial to test this hypothesis 2

2

.

Framework Opens Doors for Therapeutic Discovery

This work introduces a framework for causal inference and therapeutic discovery in consciousness research that extends beyond treating coma patients 1

1

. The method could investigate consciousness across different species by examining not just their EEG patterns but also the underlying mechanisms 2

2

. Monti suggests the same adversarial approach could apply to other neurological disorders and psychiatric conditions like depression, training networks to recognize disease-specific EEG features and identify potential interventions 2

2

. The ability to generate and test hypotheses about complex brain states positions this adversarial AI system as a tool for understanding consciousness and developing treatments where experimental models have been lacking.