Under general anesthesia, the brain still processes language and predicts the next word

Neurons in the hippocampus of patients under general anesthesia tracked the grammar of short stories and carried information that could predict upcoming words, according to a study from Baylor College of Medicine published in the journal Nature. The recordings came from seven people undergoing epilepsy surgery, and none of them remembered the sounds afterward.

Tech news worth reading: Get Techpresso, the daily email with the most interesting tech news and insights, read by 400,000+ professionals from Google, Apple and OpenAI.

Get Techpresso →

We may earn a commission if you sign up.

Sameer Sheth, a professor of neurosurgery at Baylor, is a senior author of the work, titled “Plasticity and language in the anaesthetized human hippocampus.” He said that even when patients are fully anesthetized, their brains continue to analyze the world around them.

Neuropixels recordings in seven epilepsy patients

The paper, listed in Nature with Kalman Katlowitz as first author, rests on a rare opportunity. Seven people were scheduled for an anterior temporal lobectomy to treat severe epilepsy, and surgeons temporarily placed Neuropixels probes in the hippocampus, a structure tied to memory and learning, according to Scientific American. The probes recorded the activity of hundreds of individual neurons at once, while the patients were under intravenous general anesthesia, mainly propofol.

The researchers ran two experiments. In the first, patients heard a repeating tone that was occasionally interrupted by a different sound. In the second, they heard short stories, including episodes of the Moth Radio Hour podcast. The Baylor summary on ScienceDaily reports that the neurons responded differently to the unexpected tones and became better at telling them apart as the session went on, a pattern the authors take as evidence that the hippocampus kept learning while the patients were unconscious, which the authors read as neural plasticity. Scientific American says that learning built up over about 10 minutes and was not detectable in the first few minutes.

Nouns, verbs, and the next word

With the stories, the activity patterns carried information about word categories, with some neurons responding more to nouns and others to verbs and adjectives. Scientific American adds that the neurons tracked semantic similarity, placing a word like cat near dog and far from pen. The patterns also carried information that could be used to predict the upcoming word, which the authors connect to predictive coding, the idea that the brain constantly anticipates what comes next.

Benjamin Hayden, a Baylor professor of neurosurgery, told The Debrief that this kind of predictive coding is normally associated with being awake and attentive. Sheth said the brain appears to anticipate what comes next in a story even without conscious awareness, and he compared the process to the way large language models predict text. The team also suggested the results could eventually inform speech prosthetics and brain-computer interfaces.

The tone experiment had a second part. According to The Debrief, the team built a recurrent neural network model trained on the same repeating-tone task, a way of checking whether the pattern seen in the neurons resembled what a simple learning system would produce. Kalman Katlowitz, a Baylor neurosurgery resident and study author, raised the possibility of using such signals to run a speech prosthetic. That application is a stated hope rather than a demonstrated result, since the work involved recording from patients who were not trying to speak or communicate.

The authors also drew a conclusion about consciousness itself. Sheth said the work pushes researchers to rethink what it means to be conscious, and the team suggests consciousness may depend on how information is coordinated across larger brain networks, not only on whether one region is active. That is an interpretation offered alongside the data, and the recordings themselves came from a single structure.

What the anesthesia study does not show

The caveats are specific. All seven patients were under propofol-based anesthesia, and Sheth told Scientific American the team wants to be careful to say the findings were obtained under a particular anesthetic regimen. Technology Networks notes that the results may not apply to other anesthetics or to other unconscious states such as sleep or coma, and that nothing is known yet about whether similar processing happens in those conditions.

Only the hippocampus was recorded, so how widespread the processing is across the brain is unknown, and how other regions interact with the hippocampus in this state is a question the authors leave for later work. The sample is also very small, which limits how far any single patient’s pattern can be generalized. The findings also do not show that anesthetized patients consciously understood the stories. What was measured was neural activity that carried information about grammar and upcoming words, while the participants themselves reported no recollection of what they heard.

Janna Helfrich, an anesthesiologist at Yale who was not involved in the study, told Scientific American the results fit with earlier reports of patients recognizing words at above-chance levels under anesthesia without explicit memory. She also raised a practical question the study does not answer: how much of the auditory environment patients process during surgery, and whether what they hear in the operating room should be managed more deliberately, a concern that applies to the conversations and sounds of an operating room during any procedure under general anesthesia.

This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.

Tech news worth reading: Get Techpresso →

We may earn a commission if you sign up.