Disentangling Shared and Private Neural Dynamics with SPIRE: A Latent Modeling Framework for Deep Brain Stimulation

AI summary

Overview

Research area: Computational neuroscience and machine learning for neural signal analysis, specifically latent-variable modeling of multi-region intracranial recordings during deep brain stimulation (DBS).

Technical level: Advanced. The paper assumes familiarity with autoencoders, recurrent networks (GRUs), variational and non-variational latent models, canonical correlation analysis, and neural recording modalities such as local field potentials (LFPs).

Scope: The paper introduces and validates SPIRE, a deep multi-encoder autoencoder that separates shared (cross-regional) from private (region-specific) neural dynamics, and applies it to synthetic benchmarks and pediatric intracranial DBS recordings.

What This Paper Is About

Brain regions coordinate with one another, and interventions like deep brain stimulation change that coordination in ways that local measurements cannot capture. Existing latent-variable models either assume linear relationships or treat all latent dimensions as one unified space, so they cannot cleanly separate what is shared between regions from what is unique to each region. The authors build SPIRE to perform this separation nonlinearly, train it only on baseline (stimulation-off) data, and then use it to show how stimulation reorganizes cross-regional dynamics in human recordings.

Key Contributions

1. A nonlinear disentangling model. SPIRE is presented as the first deep nonlinear framework that explicitly factorizes multi-region intracranial recordings into shared and private latent subspaces, using new alignment and disentanglement losses tailored to LFPs.

2. Synthetic validation against ground truth. On three synthetic datasets (D0, D1, D2) with known shared and private latent sources, SPIRE is compared statistically to DLAG across four random seeds, outperforming it in the nonlinear (D1) and time-varying delay (D2) regimes for shared latents and significantly outperforming it for private latents.

3. First application of disentangled latent modeling to pediatric DBS. The paper reports what it describes as the first demonstration of this kind of modeling on pediatric DBS recordings, showing stimulation-specific reorganization of shared dynamics across globus pallidus internus (GPi) and subthalamic nucleus (STN).

4. Reproducible release. Code, configs, and scripts are released at https://github.com/Rahil-Soroush/spire-iclr2026.

Main Findings

• SPIRE recovers shared structure better than DLAG under nonlinear distortion. On synthetic dataset D1, CCA-aligned shared latents of region 1 reached correlations of 0.92, 0.91, and 0.71 for SPIRE versus 0.86, 0.79, and 0.60 for DLAG.

• Private latents are recovered significantly better by SPIRE. Averaged across four seeds and both regions on datasets D0–D2, SPIRE was statistically significantly better than DLAG at retrieving ground-truth private latents. For shared latents, SPIRE outperformed DLAG in the nonlinear (D1) and time-varying delay (D2) regimes, but not with statistical significance. The authors note D0 is a linear, DLAG-friendly regime and emphasize D1 and D2 as more realistic.

• Shared GPT/STN subspaces align almost perfectly in patient data. After alignment, top-k CCA between shared GPi and STN spaces showed near-unit canonical correlation (median around 1.0), while CCA between shared and private subspaces within each region was markedly lower (medians 0.55–0.65), consistent with successful disentanglement.

• Shared latents carry most of the recoverable signal. Reconstruction error on held-out test data was near zero using the full latent space (medians 0.00211 for GPi and 0.000983 for STN). Private-only reconstructions were much worse (medians 0.544 for GPi and 0.391 for STN). Same-region shared latents reconstructed substantially better (medians 0.0462 for GPi and 0.0178 for STN), and cross-region shared latents performed worse than same-region shared latents.

• SPIRE beats SharedAE and MMVAE on reconstruction. SPIRE consistently achieved lower MSE across both GPi and STN regions. DLAG could not be applied to the intracranial recordings because it consistently failed to converge, producing numerical instabilities during Gaussian process optimization.

• Shared latents decode stimulation frequency better than private latents. Random Forest classifiers trained on raw timepoint-level embeddings decoded four GPi stimulation conditions (Off, 85, 185, 250 Hz; n = 17 hemispheres) and three STN conditions (Off, 85, 185 Hz; n = 13 hemispheres). Both latent types performed above chance, but shared latents were significantly more accurate than private latents in both datasets (linear mixed-effects model with Tukey correction, p < 0.001), with no significant difference between GPi- and STN-derived shared spaces. Maximum Mean Discrepancy analysis also confirmed frequency-dependent divergence in both latent types.

• The decomposition adapts to the subject. The authors report two contrasting cases: S3_R (balanced), where shared and private clusters are distinct and paired shared dimensions co-fluctuate with small lags; and S8_R (private-dominant), where shared GPi/STN traces remain phase-aligned but have very low amplitude and slow baseline co-modulation, while private latents carry larger region-specific dynamics.

Methodology in Plain English

SPIRE gives each brain region its own encoder-decoder pair. A GRU encoder turns each region's multichannel signal into a hidden state, which is then linearly projected into two separate latent sequences: a shared one and a private one. The decoder reconstructs the original signal from the concatenation of both.

To make shared latents comparable across regions, SPIRE passes them through a lightweight alignment step: a depthwise 1D convolution over time (ConvAlign) followed by a linear mapper, both initialized to identity so they start as do-nothing operations. The pairing is directional, so the alignment from region A to region B need not equal the reverse.

Training uses a weighted sum of nine loss terms: reconstruction, cross-reconstruction (reconstructing one region from another region's shared latents), self-reconstruction (from a region's own shared latents), alignment (VICReg applied to shared latents across regions), orthogonality (penalizing cross-covariance between standardized shared and private latents), two variance guards, and two regularizers that keep the mappers and ConvAlign filters close to identity. Loss weights follow a three-phase schedule (pre, ramp, post) aligned to a private gate that gradually opens between epochs 80 and 140; shared projections are frozen during epochs 90–110.

Inputs are lag-augmented (stacked time-lagged copies of each channel) to tolerate temporal misalignment. The model is trained in PyTorch with Adam, up to 500 epochs, GRU hidden size 64, dropout 0.3, learning rate initialized at 10⁻³ and halved on plateaus, gradients clipped to L2 norm 1.0, and early stopping.

For the synthetic benchmark, the authors generated three datasets with known ground-truth latents: D0 (linear mixing with Gaussian noise), D1 (region-mismatched nonlinear warps, bilinear mixing, 1/f noise, AR(1) latents), and D2 (D1 plus a sinusoidal time-varying inter-regional lag of amplitude 3 samples). Each contained 100 trials of 250 timepoints at 500 Hz, with three shared and three private latent dimensions per region.

For the human data, they analyzed LFPs from ten pediatric patients (ages 5–23) with dystonia implanted with DBS leads targeting GPi and STN. Recordings were sampled at 24,414 Hz, bipolar re-referenced, downsampled to 500 Hz, notch filtered, and low-pass filtered at 50 Hz, then segmented into non-overlapping 0.5 s windows. Because subjects varied, SPIRE was trained across a grid of shared (3–5) and private (2–4) latent dimensions, with the best configuration selected per case based on latent variance and validation loss.

Why This Matters

Impact on research. The paper offers a way to treat external perturbation as a probe of network reorganization rather than as a local effect. By training on baseline data and then measuring how shared latents shift under stimulation, it provides a reference-frame approach that could be transferred to other multi-view dynamical systems. It also addresses a practical barrier: the authors note reproducibility is often difficult in neuroengineering, and SPIRE's small set of interpretable losses is designed to be easy to reproduce and adapt.

Real-world applications:

• Understanding and potentially refining DBS targeting and parameter selection for movement disorders such as dystonia and Parkinson's disease.
• Building biomarkers of circuit-level dysfunction from cross-regional coordination rather than single-electrode features.
• Monitoring how stimulation reorganizes brain networks over clinically relevant frequency settings.
• Providing a general analysis tool for any multi-region recording setup where shared versus region-specific signals must be separated.

Industry relevance. The framework is relevant to companies and clinical groups developing closed-loop neuromodulation systems, DBS device programming tools, and neural data analysis pipelines, since it produces compact latent representations that could feed into decoding or control loops. The released code lowers the barrier to adoption.

Future Directions

• Extending SPIRE to longer stimulation paradigms, since the present work is restricted to relatively short-timescale stimulation.
• Integrating spiking data alongside field potentials, and adding probabilistic objectives for uncertainty quantification.
• Testing generalization to more than two regions (for example cortex and thalamus), to other etiologies, and to chronic timescales.
• Assigning biophysical meaning to individual latent dimensions, which the authors state will require complementary experiments and multimodal validation; they describe the latents as statistical abstractions for now.

Target Audience

Computational neuroscientists and machine learning researchers working on latent-variable models for neural data; neuroengineers and clinicians involved in DBS research who need tools for analyzing multi-region recordings under stimulation; and methods-focused readers interested in disentangled representation learning applied to real, nonlinear, misaligned biological time series. Readers without background in either latent modeling or intracranial electrophysiology will find the methods sections dense.

Authors’ abstract

Disentangling shared network-level dynamics from region-specific activity is a central challenge in modeling multi-region neural data. We introduce SPIRE (Shared-Private Inter-Regional Encoder), a deep multi-encoder autoencoder that factorizes recordings into shared and private latent subspaces with novel alignment and disentanglement losses. Trained solely on baseline data, SPIRE robustly recovers cross-regional structure and reveals how external perturbations reorganize it. On synthetic benchmarks with ground-truth latents, SPIRE outperforms classical probabilistic models under nonlinear distortions and temporal misalignments. Applied to intracranial deep brain stimulation (DBS) recordings, SPIRE shows that shared latents reliably encode stimulation-specific signatures that generalize across sites and frequencies. These results establish SPIRE as a practical, reproducible tool for analyzing multi-region neural dynamics under stimulation.