From Scalp to Ear-EEG: A Generalizable Transfer Learning Model for Automatic Sleep Scoring in Older People

Abstract

Objective: Sleep monitoring has extensively utilized electroencephalogram (EEG) data collected from the scalp, yielding very large data repositories and well-trained analysis models. Yet, this wealth of data is lacking for emerging, less intrusive modalities, such as ear-EEG.

Methods and procedures: The current study seeks to harness the abundance of open-source scalp EEG datasets by applying models pre-trained on data, either directly or with minimal fine-tuning; this is achieved in the context of effective sleep analysis from ear-EEG data that was recorded using a single in-ear electrode, referenced to the ipsilateral mastoid, and developed in-house as described in our previous work. Unlike previous studies, our research uniquely focuses on an older cohort (17 subjects aged 65-83, mean age 71.8 years, some with health conditions), and employs LightGBM for transfer learning, diverging from previous deep learning approaches.

Results: Results show that the initial accuracy of the pre-trained model on ear-EEG was 70.1%, but fine-tuning the model with ear-EEG data improved its classification accuracy to 73.7%. The fine-tuned model exhibited a statistically significant improvement (p < 0.05, dependent t-test) for 10 out of the 13 participants, as reflected by an enhanced average Cohen's kappa score (a statistical measure of inter-rater agreement for categorical items) of 0.639, indicating a stronger agreement between automated and expert classifications of sleep stages. Comparative SHAP value analysis revealed a shift in feature importance for the N3 sleep stage, underscoring the effectiveness of the fine-tuning process.

Conclusion: Our findings underscore the potential of fine-tuning pre-trained scalp EEG models on ear-EEG data to enhance classification accuracy, particularly within an older population and using feature-based methods for transfer learning. This approach presents a promising avenue for ear-EEG analysis in sleep studies, offering new insights into the applicability of transfer learning across different populations and computational techniques.

Clinical impact: An enhanced ear-EEG method could be pivotal in remote monitoring settings, allowing for continuous, non-invasive sleep quality assessment in elderly patients with conditions like dementia or sleep apnea.

Keywords: Automatic sleep scoring; ear-EEG; hearables; machine learning; wearable EEG.

FIGURE 1.

The experimental EEG setup, with the scalp electrodes and the generic Ear-EEG sensor visible.

FIGURE 2.

Sleep data for a representative participant. (A) Time-frequency spectrogram of ear-EEG recorded during one night of sleep, showing the frequency distribution over time. (B) Consensus hypnogram, illustrating the ground-truth sleep stages throughout the night. Red lines indicate periods of REM sleep. (C) Predicted hypnogram using the ear-EEG, generated by the pre-trained model.

FIGURE 3.

Comparative analysis of the Higuchi fractal dimension feature across EEG modalities and its relation to sleep stages. (A) Boxplot distributions showcasing the spread of the Higuchi fractal dimension feature derived from ear-EEG across all sleep stages and participants. (B) Corresponding boxplot distributions for the same feature, but derived from scalp EEG data. (C) Scatterplot visualizing mutual information scores between feature values and sleep stages, with higher scores indicating greater similarity between features and sleep stages. Features from scalp EEG (x-axis) had higher mutual information with sleep stages than features from ear-EEG (y-axis).

FIGURE 4.

Comparison of sleep stage predictions by the pre-trained model across EEG modalities. (A) Confusion matrix illustrating the classification results of the pre-trained model using ear-EEG data over all sleep epochs across participants. (B) Confusion matrix, but based on scalp EEG data, highlighting the performance differences between the two modalities. SE/PR denote respectively sensitivity (above) and precision (below).

FIGURE 5.

Enhanced sleep stage classification results following continued training on the ear-EEG dataset. (A) Confusion matrix depicting the classification outcomes of the fine-tuned model using ear-EEG data over all sleep epochs across all participants, showcasing improved accuracy and sensitivity. SE/PR denote respectively sensitivity (above) and precision (below). (B) Scatterplot presenting a direct comparison of kappa scores per subject: those derived from the pre-trained model on the x-axis versus the kappa scores from the fine-tuned model on the y-axis, highlighting the enhancement in model performance post fine-tuning.

FIGURE 6.

Results of continued training on the scalp EEG dataset, illustrating the contrasting effect compared to ear-EEG fine-tuning. (A) Confusion matrix presenting the classification outcomes of the fine-tuned model using scalp EEG data over all sleep epochs across participants, underscoring the limited improvement in performance. SE/PR denote respectively sensitivity (above) and precision (below). (B) Scatterplot offering a direct juxtaposition of kappa scores: values derived from the pre-trained model on the x-axis against those from the fine-tuned model on the y-axis, indicating the minimal variance in model efficacy despite the continued training approach.

FIGURE 7.

Comparative analysis of the highest average SHAP values associated with predicting the N3 sleep stage using ear-EEG features. (A) Bar chart depicting the top 5 SHAP values and their corresponding features as determined by the pre-trained model, offering insights into the most influential features before fine-tuning. (B) Corresponding bar chart for the top 5 SHAP values and their associated features for the fine-tuned model, highlighting the shift in feature importance and relevance post model refinement.