Stanford's AI predicts neurostimulation with 90.6% accuracy from minutes of data

Multi-compartment Hodgkin-Huxley (HH) models are the gold standard for predicting neural dynamics under electrical stimulation, but fitting them typically requires invasive intracellular recordings that limit throughput. Multi-electrode arrays (MEAs) offer high-density extracellular data from full neural populations, but HH model complexity has made biophysical inference from such data unreliable. Researchers from Stanford—led by Amrith Lotlikar, Ian Christopher Tanoh, and collaborators from Stanford, UC Santa Cruz, and Salk Institute—now present a framework leveraging differentiable biophysical simulation and simulation-based inference to rapidly infer HH parameters from designed features of extracellular MEA measurements.

Validated on hundreds of hours of stimulation and recording data from isolated macaque retina using a 512-electrode array (30 μm pitch), the model predicted responses to previously unseen multi-electrode stimulation patterns with 90.6% accuracy. Crucially, HH models were fit from only a few minutes of recording, replacing hours of stimulus testing. This work, accepted at ICML 2026, directly addresses a central challenge in translational neuroengineering: enabling precise, personalized neurostimulation for applications like retinal prosthetics, without the need for invasive intracellular recordings or lengthy calibration sessions.