Research map / Probabilistic and scientific modeling / Dynamical systems
OMiSO: Adaptive optimization of state-dependent brain stimulation to shape neural population states
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes OMiSO, an online microstimulation framework that conditions electrode-pattern selection on pre-stimulation PFC population state via a behavioral-cloned inverse model, adaptively refined with PPO, and tests it in one monkey.
Problem
Closed-loop brain stimulation aims to drive high-dimensional neural population activity to target states, but existing methods ignore the pre-stimulation brain state when choosing stimulation parameters and do not leverage it for high-dimensional population activity.
Approach
Combine a pre-stimulation-state-aware stimulation-response model with a behavioral-cloning inverse model that outputs electrode suitabilities, and adaptively update the inverse model online with PPO using observed post-stimulation errors relative to the target latent state.
Key results
- OMiSO achieved significantly smaller error to target than Static with pre-stim (p=0.05) and MiSO (p=0.01) across 14 test sessions (Wilcoxon signed-rank).
- Pre-stim-aware stimulation-response models (CNN x LSTM with 5 pre-uStim bins) outperformed the uStim-pattern-only model in post-uStim state prediction (p<0.05).
- BC inverse model inversion error saturated around 0.13 a.u., versus trial-by-trial std of 0.7 a.u. in no-uStim trials.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add pre-stimulation state information to the uStim pattern-only response model (ablation) | Post-uStim state prediction · Monkey PFC uStim, 6 sessions | post-uStim state prediction performance: p<0.05 for all four models | Helped |
| Use five pre-stimulation time bins with an LSTM alongside the uStim pattern (ablation) | Post-uStim state prediction · Monkey PFC uStim, 6 sessions | post-uStim state prediction performance: p<0.05 | Helped |
| Use uStim pattern and five pre-stimulation bins with an LSTM instead of pre-stimulation information alone (ablation) | Post-uStim state prediction · Monkey PFC uStim, 6 sessions | post-uStim state prediction performance: p<0.05 | Helped |
| Train the inverse model for more epochs (ablation) | OMiSO behavioral-cloning inverse model · Monkey PFC uStim inversion validation | Manhattan distance: around 100,000 epochs; around 0.13 a.u. | Helped |
| Use a pre-stimulation-state-conditioned inverse model rather than a static pattern-only lookup (ablation) | uStim pattern selection · Monkey PFC target-state experiments, 14 sessions | absolute error to target state: N=14 sessions, p=0.3 | No clear effect |
| Adaptively update the inverse model using newly observed stimulation responses (ablation) | OMiSO · Monkey PFC target-state experiments, 14 sessions | absolute error to target state: N=14 sessions, p=0.05 vs Static with pre-stim; p=0.01 vs MiSO | Helped |
Surprising findings
- Static with pre-stim was not significantly better than the Static table-based method (p=0.3), despite the inverse model being only approximately invertible.
Limitations
- Single monkey, limited sessions (14 test sessions), with differences only reaching significance at p=0.05 or 0.01.
- Benefit of pre-stimulation state may be small under strong stimulation or attractor dynamics, and under distribution shift between training and use.
- Uses only on-policy samples within a session without replay buffer, limiting sample efficiency.
Experimental setup
Datasets: uStim recordings from 10 training sessions in a macaque PFC (96-electrode Utah array) 7 offline analysis sessions
Benchmarks: in-house monkey uStim experiments (14 test sessions)
Metrics: absolute error to target latent state (L1 distance in 2-d la Manhattan distance inversion error one-tailed t-test Wilcoxon signed-rank test
Baselines: MiSO Static Static with pre-stim
Builds on: Proximal Policy Optimization Behavioral cloning MiSO Factor analysis
Closest papers by meaning
- Coprocessor Actor Critic: A Model-Based Reinforcement Learning Approach For Adaptive Brain Stimulation (ICML 2024 · similarity 0.84)
- Active learning of neural population dynamics using two-photon holographic optogenetics (NeurIPS 2024 · similarity 0.84)
- BLEND: Behavior-guided Neural Population Dynamics Modeling via Privileged Knowledge Distillation (ICLR 2025 · similarity 0.83)
- NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces (NeurIPS 2026 · similarity 0.83)
- A Scalable, Causal, and Energy Efficient Framework for Neural Decoding with Spiking Neural Networks (NeurIPS 2025 · similarity 0.83)
- MindSimulator: Exploring Brain Concept Localization via Synthetic fMRI (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.