A multi-task and multi-channel convolutional neural network for semi-supervised neonatal artefact detection.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Automated artefact detection in neonatal electroencephalogram (EEG) is vital for reliable analysis.
• Limited expert annotations hinder deep learning model development for EEG artefact detection.

Purpose of the Study:

• Propose a semi-supervised deep learning approach for neonatal EEG artefact detection using minimal labelled data.
• Train a multi-task convolutional neural network (CNN) for improved artefact detection.

Main Methods:

• Developed a multi-output model combining an autoencoder and an artefact classifier for joint unsupervised and supervised optimization.
• Processed multi-channel EEG inputs using a novel semi-supervised multi-task training strategy.
• Compared the proposed method against supervised strategies and state-of-the-art models on two neonatal EEG datasets.

Main Results:

• The proposed multi-task, multi-channel CNN achieved high F1 scores (86.2% and 95.7%) on separate datasets, outperforming existing methods.
• The semi-supervised multi-task strategy proved superior to supervised training when labelled data was scarce.
• A correlation was observed between the error in functional brain age (FBA) prediction and the quantity of automatically detected artefacts.

Conclusions:

• The semi-supervised multi-task training strategy effectively trains CNNs with limited labelled data, offering a promising solution for artefact detection in neonatal EEG.
• Artefact detection is crucial for robust automated EEG analysis, as indicated by its impact on FBA prediction accuracy.