Learning Generalizable Reconstruction of High-Dimensional Neural Dynamics

Accurately reconstructing long-duration, high-dimensional neural recordings—such as multichannel local field potentials (LFPs) containing transient sharp wave-ripples—is difficult due to high temporal resolution, channel complexity, and inter-subject variability. Traditional dynamic mode decomposition (DMD) methods often suffer from poor numerical conditioning or lose fine-scale temporal structures over long observation horizons, whereas deep learning models demand substantial computational resources and lack dynamical interpretability.

To address these limitations, the authors propose PCA-DMD. This framework segments multichannel LFP recordings into overlapping temporal windows, projects them into a compact principal component analysis (PCA) latent space, learns a linear approximation of Koopman evolution within that latent space, and reconstructs the continuous signals via inverse projection and overlap-add aggregation.

The framework is evaluated through a series of increasingly demanding experiments:

• Reconstruction Benchmark: On 200,000-sample hippocampal recordings, PCA-DMD outperforms classical DMD and specialized variants like SpDMD, MrDMD, and HODMD, yielding lower Kullback-Leibler divergence (KLD) and Hellinger distance (HD).
• Cross-Subject Zero-Shot Generalization: In an all-pair transfer experiment across 300,000 samples, models trained on one subject generalize to unseen subjects without fine-tuning, achieving high correlations (0.9504–0.9800) and minimal distributional discrepancies.
• Temporal Prediction: Out-of-sample temporal prediction on held-out LFP segments demonstrates accurate one-step prediction across unseen intervals, multiple channels, and different temporal regions.
• Scalability and External Validation: Scalability tests from 400,000 to 900,000 samples show stable reconstruction performance with predictable runtime growth. External validation on a 93-channel Allen Neuropixels recording further confirms the method's applicability across different acquisition setups.