Byte-Pair Encoding for Classifying Routine Clinical Electroencephalograms in Adults Over the Lifespan.

Area of Science:

• Neuroscience
• Computational Biology
• Medical Informatics

Background:

• Routine clinical electroencephalogram (EEG) interpretation is time-consuming and subject to inter-reader variability.
• Automated decision support tools are needed to enhance the efficiency and consistency of EEG classification.
• Challenges in EEG classification include interpretability, variable recording durations, and diverse recording devices.

Purpose of the Study:

• To validate a novel framework for automated EEG classification by transforming EEG signals into unstructured text.
• To assess the framework's ability to handle heterogeneous EEG data and satisfy interpretability requirements.
• To demonstrate the framework's utility in predicting patient biological age and identifying clinically relevant EEG patterns.

Main Methods:

• Transformed multichannel EEG time series into symbolic representations.
• Applied a byte-pair encoding (BPE) algorithm to extract frequent waveform patterns (tokens) from the symbolic EEG data.
• Utilized a Random Forest regression model to predict biological age using the extracted EEG features.

Main Results:

• The EEG age prediction model achieved a mean absolute error of 15.7 years.
• Token occurrence frequencies showed significant correlations with patient age, particularly at frontal and occipital channels.
• The framework demonstrated feasibility in classifying routine clinical EEG with minimal preprocessing.

Conclusions:

• An NLP-based approach is feasible for classifying routine clinical EEG recordings.
• The proposed framework can aid in automated EEG classification and the identification of short, clinically relevant events like epileptic spikes.
• This method offers a promising avenue for improving the efficiency and accuracy of neurological evaluations using EEG data.