Organizers: Tay Netoff, PhD, University of Minnesota & Babak Mahmoudi, PhD, Emory University
Neuromodulation devices offer billions to quadrillions of possible stimulation settings due to the combinatorial selection of electrodes and waveform parameters. Identifying effective configurations within this vast parameter space presents a major clinical and experimental challenge. Bayesian Optimization provides a principled and efficient framework for navigating this space by balancing exploration of novel parameter settings with exploitation of settings known to provide therapeutic benefit.
This workshop will begin with motivating examples drawn from optimization of deep brain stimulation for Parkinson’s disease and spinal cord stimulation for spinal cord injury. We will then introduce the core concepts and workflow of Bayesian Optimization, emphasizing intuitive and conceptual understanding rather than mathematical formalism. Finally, participants will engage in hands-on exercises implementing Bayesian Optimization in Python using Google Colab. By the end of the workshop, attendees will be equipped to apply and adapt these methods within their own research or clinical workflows.
| 1:25 PM | 1. Introduction (10 min) What is Bayesian Optimization? | Why Bayesian Optimization is well-suited for neuromodulation | Practical motivation: large parameter spaces, noisy measurements, limited evaluations |
|---|---|
| 1:35 PM | 2. Application Examples (10 min) Deep Brain Stimulation for Parkinson’s Disease: Babak Mahmoudi, PhD — Optimization of DBS programming using kinematic biomarkers Spinal Cord Stimulation for Spinal Cord Injury: Tay Netoff, PhD — Preference- and probit-based optimization frameworks |
| 1:45 PM | 3. Core Concepts and Theory (30 min) Problem formulation and choice of initial settings | Gaussian Process Regression (GPR) | Mean functions and kernels | Uncertainty estimation | Acquisition functions | Exploration vs. exploitation | Sampling strategies (e.g., Thompson sampling) | Extensions and Practical Considerations | Multi-objective optimization | Safety constraints and safe exploration | Limitations and challenges (Non-stationarity, Drift and adaptation over time, Hierarchical and patient-specific modeling) | Key terminology |
| 3:15 PM | Break (5 min) |
| 3:20 PM | 4. Hands-On Implementation (35 min) Environment setup (Google Colab, Python libraries GPflow) | Simulating an Optimization Problem (Defining a “true” objective surface, Sampling and estimation, Generative models, Measurement noise and drift) | Constructing objective functions (Single vs. multi-objective, Probit / preference-based models, Safety constraints) | Building the Bayesian Optimization Pipeline (Gaussian Process Regression, Kernel selection, Length-scale interpretation, Data weighting and uncertainty, Acquisition functions, Sample selection functions) |