In a fascinating intersection of neuroscience and audio technology, researchers have discovered that neural audio codecs, typically used for compressing and transmitting audio data, can be repurposed to handle EEG (electroencephalogram) signals. This breakthrough, led by Ard Kastrati, Luca Lanzendörfer, Riccardo Rigoni, John Staib Matilla, and Roger Wattenhofer, opens up new possibilities for efficient EEG data compression and transmission, which is crucial for applications in medical diagnostics and brain-computer interfaces.
EEG and audio signals are fundamentally different. EEG signals measure electrical activity in the brain, characterized by low sampling rates, multiple channels, and small-scale voltage fluctuations. In contrast, audio signals capture sound waves, which have higher sampling rates, fewer channels, and larger-scale amplitude variations. Despite these differences, the researchers found that with appropriate preprocessing, EEG data can be adapted to fit the input constraints of neural audio codecs.
The team used DAC, a state-of-the-art neural audio codec, as their foundation. They demonstrated that raw EEG signals can be mapped into the codec’s stride-based framing, allowing the reuse of the audio-pretrained encoder-decoder. Even without any modifications, this setup produced stable EEG reconstructions. However, fine-tuning the codec on EEG data significantly improved the fidelity and generalization of the reconstructions compared to starting from scratch.
To capture the spatial dependencies across EEG electrodes, the researchers proposed DAC-MC, a multi-channel extension of the DAC codec. This extension incorporates attention-based cross-channel aggregation and channel-specific decoding, while still benefiting from the audio-pretrained initialization. The evaluations on the TUH Abnormal and Epilepsy datasets showed that the adapted codecs preserve clinically relevant information. This was reflected in both spectrogram-based reconstruction loss and downstream classification accuracy.
The researchers systematically explored the compression-quality trade-offs by varying parameters such as residual codebook depth, codebook (vocabulary) size, and input sampling rate. Their findings suggest that neural audio codecs, with some adjustments, can be a powerful tool for EEG signal compression. This could lead to more efficient storage and transmission of EEG data, ultimately benefiting medical professionals and researchers who rely on this information for diagnostics and further study.
This research not only highlights the versatility of neural audio codecs but also bridges the gap between audio technology and neuroscience. It challenges the conventional boundaries of these fields and paves the way for innovative applications in medical technology and brain-computer interfaces. As we continue to explore the capabilities of neural networks, such interdisciplinary approaches could unlock even more potential for advancements in both fields.
