SAN DIEGO — Artificial intelligence (AI) is transforming the field of neuroscience, enabling scientific breakthroughs previously out of reach. New applications of AI and machine learning techniques will be presented at Neuroscience 2025, the annual meeting of the Society for Neuroscience and the world’s largest source of emerging news about brain science and health.
The promise of AI is immense. In the realm of neuroscience, AI models can analyze enormous datasets quickly and effectively, from subtle body movements to internal word processing. AI could eventually even simulate some brain functions, allowing researchers to test hypotheses without the need for a human or animal subject. This analysis can accelerate experiments and lead to faster insights, which in turn will drive diagnosis and treatment options.
New findings show that:
- Using recordings from inside the brain, scientists applied machine learning to decode word categories from corresponding brain activity with up to 77% accuracy — the highest reported for this kind of process. This decoding could help nonspeaking patients better communicate. (Matthew Nelson, University of Alabama at Birmingham)
- AI algorithms correctly quantified gait patterns and disturbances from smartphone videos of individuals walking, matching the output of expert rehabilitation clinicians. AI could be used to augment clinical judgement of gait problems. (Trisha Kesar, Emory University)
- AI was able to predict freezing of gait — a symptom of Parkinson’s disease that causes a temporary inability to take a step — before it happened, allowing researchers to characterize its subtle neural signature and eventually improve the use of deep brain stimulation techniques to mitigate it. (Paul Cantlay, Cleveland Clinic Foundation)
- An AI system can predict a simulated neuron's ion channel composition, which could help scientists create “digital twins” of real neurons and use them to answer questions about brain function and dysfunction. (Roy Ben-Shalom, University of California Davis)
- Artificial neural networks (ANNs) that capture key properties from real networks in the brain are leading to new understandings of complex brain functions. (Marcel Oberlaender, Max Planck Institute for Neurobiology of Behavior; Center for Neurogenomics and Cognitive Research at VU Amsterdam)
“AI is no longer just a tool borrowed from computer science,” said Christopher Rozell, PhD, executive director of the Institute for Neuroscience, Neurotechnology, and Society at the Georgia Institute of Technology and moderator of the press conference. “It’s deeply integrated into neuroscience, enabling new discoveries and therapies by allowing us to identify patterns and mechanisms that were invisible before. At the same time, AI was inspired by biological intelligence, so the more we learn about the brain, the more we can improve AI.”
For complete access to Neuroscience 2025 in-person and online, request media credentials. This research was supported by national funding agencies including the National Institutes of Health and private funding organizations.
Monday, November 17, 2025
11 a.m.–noon PST
San Diego Convention Center, Room 15A, and online for registered media
AI Press Conference Summary
- These studies utilized AI and machine learning to help clinical diagnosis, improve brain-computer interfaces, and analyze distinct cell and brain network functions.
- One showed AI’s potential in improving brain-computer interfaces by decoding word categories from brain activity alone instead of relying on physical cues.
- One study showed how algorithms can quickly and effectively offer gait diagnoses in human subjects just as well as a clinician; another used deep learning to predict gait freezing in Parkinson’s patients before it happened.
- One study showed how quickly an AI system can identify the ion channel composition in a neuron, a process that used to take months; another study used neuroanatomical reconstructions of the human cortex to develop ANNs for greater sophistication.
Decoding and characterizing the intracranial representation of semantic information
Matthew Nelson, matthewnelson@uabmc.edu, Abstract PSTR249.12
- Brain-computer interfaces (BCIs) decode language from brain activity and could allow severely disabled patients who cannot speak but whose mental abilities are still intact to communicate. Usually, BCIs work by decoding sounds or by analyzing lip and tongue movement.
- Researchers analyzed rare recordings taken from inside the brain of epilepsy patients to decode semantic content. Patients thought about words from 15 different categories such as tools or animals, while a machine learning tool decoded the category of each word from the patients’ brain activity.
- The tool achieved up to 77% accuracy in predicting a word's category when random guessing would yield a correct answer only 7% of the time. It managed up to 97% accuracy when distinguishing between a living and non-living object; guessing would be only 50% accurate.
- This is the highest reported accuracy for this kind of semantic decoding and suggests that brain activity can reveal the meaning of words and potentially improve communication techniques.
Integrating clinician insights into markerless gait analysis: Toward AI-driven, interpretable gait assessment
Trisha Kesar, trisha.m.kesar@emory.edu, Abstract PSTR253.20
- Clinicians identify walking impairments — which can be signs of simple aging or of neurodegenerative disease — by observing their patients' gait, but their assessments require years of training and can still be subjective. 3D motion capture systems can achieve a similar analysis but require specialized equipment. Now, computer vision technologies can extract similar information from a simple smartphone video.
- The researchers used AI algorithms to analyze 743 smartphone videos of normal and impaired walking patterns and showed that the model's outputs matched those of expert rehabilitation clinicians.
- Clinicians also reported enthusiasm for using AI tools in clinical decision-making. AI analysis could therefore help improve accuracy and reliability around gait impairment diagnosis
Early Detection of Freezing of Gait episodes in Parkinson’s disease using a Deep Learning Approach
Paul Cantlay, cantlap@ccf.org, Abstract PSTR478.10
- Deep brain stimulation (DBS), a promising treatment for certain Parkinson’s symptoms, is limited in treating freezing of gait because onset is often unknown.
- Researchers trained a deep learning model on data from Parkinson’s patients to identify the unique neural signature of freezing of gait and predictors of freezing onset.
- The model’s ability to detect freezing of gait before its onset shows potential for real-time DBS intervention before freezing occurs.
Accurate inference of single-neuron biophysics from voltage responses with deep learning
Roy Ben-Shalom, rbenshalom@health.ucdavis.edu, Abstract PSTR145.09
- Malfunctioning ion channels can cause disorders including autism and epilepsy. But determining which ion channels exist in a single neuron can take months of work.
- Researchers developed an AI system called NeuroInverter that can analyze a neuron's electrical signals and thus reverse-engineer the ion channel composition. After training on millions of computer-simulated "brain cells” from the Blue Brain Project, the model successfully predicted ion channel properties for 170 different neuron types — even accurately predicting ion channels for cells it had not been trained on.
- Once the approach is validated with real neurons, it could help scientists quickly characterize cells and generate “digital twins” of the brain cells to test hypotheses around brain function and dysfunction at the ion channel level.
Predicting Structure-Function Relationships in Cortex via Artificial Neural Networks
Marcel Oberlaender, m.oberlaender@vu.nl, Abstract PSTR478.11
- Artificial neural networks (ANNs) lack many of the mechanisms that real networks use in the brain.
- Researchers developed ANNs that capture some of these biological mechanisms to better understand their relevance for complex brain functions.
- This model revealed that ANNs with fewer connections, just like in the brain, can learn faster and process information more efficiently.
- ANNs mimicking neuroanatomical architectures also showed a greater ability to combine inputs from different information streams, in particular in “noisy” environments.
###
The Society for Neuroscience (SfN) is an organization of nearly 30,000 basic scientists and clinicians who study the brain and the nervous system.