Want smarter insights in your inbox? Sign up for our weekly newsletters to get only what matters to enterprise AI, data, and security leaders. Subscribe Now
Electroencephalography, or EEG, measures electrical activity in the brain using electrodes placed on the scalp. It’s used by sleep specialists to diagnose and evaluate neurological disorders, which can be a tedious undertaking — annotating dips and spikes in hours of recorded brain activity requires specialized training and plenty of patience.
Researchers at Stanford and Université Paris-Saclay in France recently proposed an alternative: artificial intelligence that predicts the location, duration, and type of events in EEG charts. It’s described in a new paper (“A Deep Learning Architecture to Detect Events in EEG Signals During Sleep“) published on the preprint server Arxiv.org.
EEG pattern-detecting algorithms have been around a while, but the researchers note that most are event-specific; they’re hardwired to recognize known electrical patterns. By contrast, machine learning systems have the potential to learn events, like K-complexes (EEG waveforms that occur during stage 2 of NREM sleep) and sleep spindles (bursts of brain activity from the thalamus that occur during light sleep), as they’re trained on new data.
“We propose a deep learning method that … predicts locations, durations and types of events in EEG time series,” they wrote. “Detecting such events is meaningful to better understand sleep physiology and relevant to the physiopathology of some sleep disorders.”
AI Scaling Hits Its Limits
Power caps, rising token costs, and inference delays are reshaping enterprise AI. Join our exclusive salon to discover how top teams are:
- Turning energy into a strategic advantage
- Architecting efficient inference for real throughput gains
- Unlocking competitive ROI with sustainable AI systems
Secure your spot to stay ahead: https://bit.ly/4mwGngO
The team leveraged computer vision — specifically a convolutional network, a type of neural network with an architecture resembling that of the human visual cortex — to detect EEG signals. It extracted features from a dataset of 19 records from 19 subjects and used two modules, a localization module and classification module, to predict events’ start and end times and their labels.
Their results showed that after training on just 10 records and 2 validation records, the processing pipeline was able to consistently identify spindles and K-complexes in EEG readouts. Additionally, it was able to detect jointly multiple types of events, making it much more efficient than conventional, serialized algorithms.
For the estimated 50-70 million adults in the U.S. who suffer from a sleep disorder, it’s encouraging progress.
“The proposed approach seems to perform quite well with respect to different gold standards,” the team wrote. “Yet it remains to study how the method performs compared to the inter-scorer agreement. This shall be also addressed in future works.”
The researchers aren’t the only ones applying machine learning to the task of sleep analysis. Google subsidiary Verily in July announced a partnership with sleep company ResMed to form a new venture focused on developing sleep apnea treatments and connected health products.
