Machine learning-based surrogate models to predict the outcome of neuromodulation protocols

Neuromodulation offers significant potential in restoring lost functions in patients with neurological disorders, including those affecting the peripheral, spinal cord, and central nervous systems. Previous works have shown that the optimization of neuromodulation protocols can be approached through model-driven methods. However, achieving clinically viable solutions requires fast, patient-specific simulations. The computational framework commonly used to simulate neuromodulation, known as hybrid modelling (HM), is biophysically accurate but encounters significant computational bottlenecks. This work, conducted within the context of a clinical trial at San Raffaele Hospital on spinal cord injured (SCI) patients, narrows the gap toward the development of a rapid, patient-specific pipeline. After reconstructing a biophysical model for spinal cord stimulation based on hybrid modelling, we employ neural networks and operator learning to create surrogate models of the key components of HM, reaching a significant speed-up of the simulation process.