Deep transfer learning for improving single-EEG arousal detection.

Area of Science:

• Sleep Science
• Machine Learning
• Artificial Intelligence

Background:

• Sleep science datasets pose challenges for machine learning due to variations in recording equipment across different clinics.
• Inconsistent setups lead to a 'channel mismatch' problem, degrading the performance of single-electroencephalography (EEG) models.

Purpose of the Study:

• To investigate deep transfer learning strategies for overcoming the channel mismatch problem in sleep science.
• To improve the performance of machine learning models when dealing with heterogeneous sleep data sources.

Main Methods:

• A baseline model was trained on multivariate polysomnography data.
• The first two layers of the baseline model were replaced to adapt the architecture for single-channel EEG data.
• A fine-tuning strategy was employed to train the adapted model.

Main Results:

• The fine-tuned model achieved performance comparable to the baseline multivariate model (F1 scores of 0.682 and 0.694).
• The proposed transfer learning approach significantly outperformed a comparable single-channel model.
• Results indicate the efficacy of transfer learning for handling channel mismatches.

Conclusions:

• Deep transfer learning offers a promising solution for sleep science researchers facing data heterogeneity and small database limitations.
• This method enables the effective use of pre-trained deep learning models on larger datasets for analysis of smaller, potentially differently configured, datasets.