Summary: Researchers show that implant stimulation settings in animal brains can be fine-tuned autonomously by learning algorithms, without requiring continuous human adjustment. This advance points to more adaptive, personalized neuroprostheses that could improve outcomes for people living with spinal cord injuries, stroke-related motor deficits, and other movement disorders.
Source: University of Montreal
For years, scientists have explored neurostimulation as a way to restore movement and sensory function after strokes and spinal cord injuries—conditions that affect hundreds of thousands of people in Canada and millions worldwide.
A recent study published in Cell Reports Medicine describes an autonomous approach to optimizing the stimulation parameters of brain and spinal implants in animals. Using adaptive learning algorithms, the research team demonstrated that these devices can discover effective stimulation patterns on their own, reducing the need for continuous human intervention during tuning.
The work was carried out at Université de Montréal by neuroscience professors Marco Bonizzato, Numa Dancause and Marina Martinez, in collaboration with mathematics professor and Mila researcher Guillaume Lajoie. The project united neuroscience and artificial intelligence expertise to build adaptive neurostimulation systems that learn in real time.
A promising phase for neuroprosthetics
“Neuroprostheses—devices designed to reconnect neural circuits after a loss of motor function—are entering a very promising phase of development,” said Guillaume Lajoie. “Our results illustrate how autonomous optimization can enhance their performance.”
According to Marco Bonizzato, the improved outcomes stem from autonomous learning algorithms that optimize stimulation protocols more efficiently than manual or exhaustive search methods. These algorithms can quickly identify stimulation patterns that produce meaningful improvements in motor outputs.
“Optimization algorithms enable the design of refined stimulation protocols and support personalization according to each patient’s neural state,” Bonizzato added. Numa Dancause emphasized that artificial intelligence is instrumental for extracting maximal value from neural data and for anticipating changing biological conditions that standard approaches may not handle well.
These autonomous methods bring researchers closer to practical neuroprosthetic tools capable of alleviating motor deficits caused by injury or disease and of adapting over time to shifts in physiology or interface performance.
About this neurotech research news
Author: Press Office
Source: University of Montreal
Contact: Press Office – University of Montreal
Image: The image is in the public domain
Original Research: Open access. “Autonomous optimization of neuroprosthetic stimulation parameters that drive the motor cortex and spinal cord outputs in rats and monkeys” by Marco Bonizzato et al., Cell Reports Medicine. DOI: 10.1016/j.xcrm.2023.101008
Abstract
Autonomous optimization of neuroprosthetic stimulation parameters that drive the motor cortex and spinal cord outputs in rats and monkeys
Highlights
- An autonomous learning algorithm was used to optimize complex neuromodulation patterns in vivo across diverse experimental settings.
- The approach supports “intelligent” neuroprostheses that can immediately reduce motor deficits after appropriate stimulation is found.
- The method is robust to physiological changes such as plasticity, and to variations in the stimulation interface.
- Knowledge transfer to clinicians and researchers is facilitated through an open-source framework that documents algorithmic and experimental procedures.
Summary
Neural stimulation can reduce paralysis and sensory deficits. Advances in high-density neural interfaces make it possible to apply finely tuned, multi-site stimulation patterns, but they also expand the parameter space that must be searched to find effective protocols. To address this challenge, the researchers adopted an algorithmic framework based on Gaussian-process (GP) Bayesian optimization (BO).
GP-BO is an efficient strategy for exploring large, continuous parameter spaces. The team showed that GP-BO finds high-performing stimulation combinations far more quickly than exhaustive or naive search strategies, often after evaluating only a small fraction of possible parameter sets.
Through real-time, multidimensional neurostimulation experiments, the study validated optimization across multiple biological targets (motor cortex and spinal cord), animal models (rats and non-human primates), and experimental conditions including healthy subjects and subjects receiving neuroprosthetic interventions after injury. The algorithm supported both immediate optimization within a session and continual learning across multiple sessions, adapting as the biological system or interface changed.
Importantly, the GP-BO framework can incorporate prior knowledge from experts and clinicians, improving performance by seeding the search with clinically informed expectations. This capability makes the learning agent more sample-efficient and aligns algorithmic search with existing therapeutic experience.
Overall, the results support treating adaptive learning agents as a core component of future neuroprosthetic systems. By enabling personalization and continuous optimization, these agents can help maximize therapeutic benefit and accelerate translation of neurostimulation technologies to clinical applications for spinal cord injury, stroke rehabilitation, and movement disorders.