SYNAPSE: Simulation Benchmark of Neuro-Adaptive Patient-Specific Evaluation for Episodic Decision-Making

Ge Gao, Qitong Gao, Miroslav Pajic, Emma Brunskill

Keywords: Simulator and benchmark, Deep brain stimulation for Parkinson’s disease treatment, healthcare reinforcement learning

Abstract: Recent advances in time-series analysis, treatment outcome prediction, and reinforcement learning (RL) have demonstrated great potential to automate decision-making in healthcare. However, the high stakes nature complicates the deployment of such frameworks in practice, clinically, or in the long term. A major challenge is the absence of realistic benchmark environments that capture the sequential, patient-specific nature of various therapies, which could enable extensive offline testing, evaluation, and model selection prior to clinical adoption. To address this, we introduce the SImulation Benchmark of Neuro-Adaptive Patient-Specific Evaluation (SYNAPSE), in the context of adaptive deep brain stimulation (DBS), a treatment for managing the motor symptoms of Parkinson’s disease (PD). Specifically, SYNAPSE is constructed using real-world data collected from both clinical and at-home studies involving participants undergoing DBS therapy. It enables offline training and evaluation of different treatment strategies, reflecting both short- and long-term effects, as well as treatment outcome prediction capturing participants’ responses to a range of temporal dynamics. Additionally, it allows for the assessment of safety-critical constraints inherent to neurostimulation decision-making. By rigorously validating its realism against clinical data and supporting both short- and long-term decision-making, SYNAPSE offers clear guidance for future DBS policy development, as well as helps identify and address key challenges in advancing truly personalized neurostimulation therapies.

Primary Area: datasets and benchmarks

Submission Number: 23609