Predict neuromuscular performance in human epidural electrical stimulation: phase 1 trial interim results

Epidural electrical stimulation (EES) has emerged as a promising therapy for restoring motor function in patients with paralysis. A primary challenge in this therapy lies in identifying feasible stimulation parameters in huge selection space for different movements, given the limited understanding of the precise alignment between stimulation and corresponding neuromuscular performance. We aimed to develop a computational framework that predicts neuromuscular performance under EES and thereby reduces the need for extensive in-clinic parameter searches. We implanted purpose-designed 32-contact epidural interfaces in two individuals with motor-complete spinal cord injury and reconstructed personalized spinal anatomies from medical imaging. Finite element simulations and axonal recruitment modeling were integrated with machine learning to establish a predictive mapping between stimulation parameters and muscle responses. A dimensionality-reduction Bayesian optimization algorithm was subsequently applied to identify compact sets of stimulation parameters targeting specific motor objectives, and selected configurations were validated through clinical testing. This study is an interim report of an ongoing registered clinical trial (Closed-loop Functional Spinal Cord Stimulation in Patients with Spinal Cord Injury, ClinicalTrials.gov Identifier: NCT04969042), sponsored by Beijing PINS Medical Co., Ltd. Here we present purpose-designed 32-contact epidural neural interfaces that enabled two individuals with spinal cord injury to regain lower-limb motor function. The hybrid predictive model demonstrated strong quantitative agreement with experimentally measured muscle responses (mean squared error = 0.0096). Algorithm-guided parameter recommendations comprising 180 configurations outperformed both the historical dataset (1,602 configurations) and conventional bipolar settings across four functional objectives. Clinical validation further confirmed that the recorded muscle activations were in close agreement with model predictions. The AI-aided computational framework can serve as a reliable and feasible agent for evaluating and recommending effective EES parameters. By bridging anatomical modeling with functional outcomes, this approach offers a practical pathway toward optimizing neuromodulation therapies and advancing the development of personalized treatment strategies for individuals with spinal cord injury. People with spinal cord injury often lose the ability to move their legs because the connection between the brain and spinal cord is damaged. Over recent decades, epidural electrical stimulation—delivering pulses of electricity to the spinal cord—has been shown to help restore leg movement. However, finding the proper stimulation settings for each person is usually time-consuming and depends on trial and error. In this study, we designed and manufactured a 32-contact spinal implant and implanted it in two people with spinal cord injury. We also developed a personalized hybrid model that combines computer simulations and neural-network predictions to estimate how different stimulation settings affect muscle activity. An optimization algorithm then recommends parameter sets for specific movements, and these recommendations were tested in the clinic. By automating and personalizing parameter selection, this approach can reduce the testing burden on patients and clinicians and help enable more tailored neuromodulation treatments. Li et al. integrate computational modeling and AI-based optimization to personalize epidural electrical stimulation for individuals with spinal cord injury. The approach accurately predicts muscle responses and guides stimulation settings that enhance lower-limb motor recovery.