Machine learning-based high-frequency neuronal spike reconstruction from low-frequency and low-sampling-rate recordings

Nari Hong, B. F. Kim, Jaewon Lee, Han Kyoung Choe, Kyong Hwan Jin, Hongki KangNature Communications 2024 · DOI

TL;DR

High-density neural recordings are essential for modern brain-machine interfaces (BMIs) but generate massive data volumes that cause hardware bottlenecks and tissue-damaging heat. Researchers have developed Spk-Recon, a Transformer-based model that reconstructs high-frequency neuronal spikes from low-sampling-rate data, enabling up to a 25x reduction in data volume while preserving critical spike timing and network connectivity information.

What Problem Does This Paper Address?

To understand the brain, researchers need to record from as many neurons as possible. However, individual "spikes" (action potentials) occur on a millisecond timescale, requiring sampling rates of at least 10 kHz. When you scale this to hundreds or thousands of electrodes, you run into three major walls:

  1. Data Storage/Bandwidth: The sheer volume of data is difficult to store or transmit wirelessly.
  2. Power Consumption: Processing and transmitting this data drains batteries rapidly in untethered devices.
  3. Thermal Damage: High-speed electronics generate heat. In an implantable device, even a few degrees of temperature rise can cause permanent damage to surrounding brain tissue.

Current solutions often involve "on-chip" spike detection, which throws away the low-frequency signals (Local Field Potentials or LFPs) that contain vital information about brain states. This paper seeks a way to record at low speeds (saving power and data) while "hallucinating" the high-speed spikes back into existence using AI.

Methods

The authors treat the problem as a 1D version of image super-resolution. Just as AI can turn a blurry, low-res photo into a sharp one by "learning" what edges and textures look like, the authors' model, Spk-Recon, learns to reconstruct sharp neuronal spikes from smooth, low-frequency signals.

  • The Architecture: They adapted the SwinIR (Swin Transformer for Image Restoration) architecture. Transformers are particularly good at capturing long-range dependencies in data, which helps in identifying the context of a spike.
  • Pre-interpolation: Unlike standard image models, they use a Fourier-based interpolation step to upsample the low-frequency signal to the target resolution before it enters the Transformer.
  • Spike-Focused Training: Because neurons spend most of their time "silent," a random sample of brain activity is mostly noise. They implemented Spike-Focused Window Selection (SFWS), ensuring that half of the training data contained actual spikes to help the model learn the specific geometry of action potentials.
  • The "Filter" Secret: Crucially, they discovered the model relies on the "residue" of high-frequency information that survives standard low-pass filtering (using a Butterworth filter).

Datasets

The model was validated across several biological contexts:

  • In vitro: Rat hippocampal neurons grown on microelectrode arrays (MEAs).
  • In vivo: Recordings from the cortex and hippocampus of anesthetized male mice using penetrating depth probes.
  • Pathological: A "seizure" model where mice were treated with kainic acid to induce intense, abnormal neuronal firing.
  • Benchmarks: They compared Spk-Recon against existing CNN-based models like TCN (Temporal Convolutional Network) and EDSR (Enhanced Deep Super-Resolution).

Key Insights and Ideas

  • The "Aha" Moment on Filtering: The researchers found that the model performs significantly worse if an "ideal" filter (which perfectly cuts off high frequencies) is used. The model is essentially "de-aliasing" the subtle high-frequency signatures that leak through standard, non-ideal hardware filters.
  • Preserving the Network: It’s not just about making the signal look right. The reconstructed data was accurate enough to perform spike sorting (assigning spikes to specific neurons) and functional connectivity analysis (mapping how neurons talk to each other) with over 96% accuracy.
  • Generalization: A model trained on in vitro (lab-grown) neurons could still reconstruct spikes from in vivo (living brain) data, though performance improved when the model was fine-tuned on the noisier, more complex signals of a living brain.

Key Results

  • Massive Compression: The model successfully reconstructed spikes even when the data was downsampled by a factor of 25x.
  • High Fidelity: At 8x downsampling, the "hit rate" (correctly identifying a spike occurred) was between 0.8 and 0.9.
  • Timing Precision: The error in spike timing was often less than the sampling period of the low-res signal, meaning the AI "placed" the spike more accurately than the raw low-speed clock could.
  • Superiority: Spk-Recon consistently outperformed CNN-based architectures in both waveform accuracy (measured by NRMSE) and spike detection rates.

Limitations and Future Work

  • Hardware Dependency: Because the model learns to exploit the specific "leakage" of a filter, a model trained for one type of recording hardware might need retraining or calibration for another.
  • Computational Cost: While the model reduces recording and transmission costs, the reconstruction currently requires a powerful GPU (NVIDIA RTX 3090). For real-time BMI applications, this processing would need to be optimized for edge computing.
  • Cross-Modality Potential: The authors suggest this framework could eventually be used to reconstruct cellular-level spikes from even less invasive recordings, such as ECoG (electrodes on the brain surface) or even EEG (electrodes on the scalp).

Summary updated

Abstract

Recording neuronal activity using multiple electrodes has been widely used to understand the functional mechanisms of the brain. Increasing the number of electrodes allows us to decode more variety of functionalities. However, handling massive amounts of multichannel electrophysiological data is still challenging due to the limited hardware resources and unavoidable thermal tissue damage. Here, we present machine learning (ML)-based reconstruction of high-frequency neuronal spikes from subsampled low-frequency band signals. Inspired by the equivalence between high-frequency restoration and super-resolution in image processing, we applied a transformer ML model to neuronal data recorded from both in vitro cultures and in vivo male mouse brains. Even with the x8 downsampled datasets, our trained model reasonably estimated high-frequency information of spiking activity, including spike timing, waveform, and network connectivity. With our ML-based data reduction applicable to existing multichannel recording hardware while achieving neuronal signals of broad bandwidths, we expect to enable more comprehensive analysis and control of brain functions.
BibTeX
@inproceedings{hong2024machine,
  author = {Nari Hong and B. F. Kim and Jaewon Lee and Han Kyoung Choe and Kyong Hwan Jin and Hongki Kang},
  title = {{Machine learning-based high-frequency neuronal spike reconstruction from low-frequency and low-sampling-rate recordings}},
  booktitle = {Nature Communications},
  year = {2024},
  doi = {10.1038/s41467-024-44794-2},
}
Figure