Amplifying Pathological Detection in EEG Signaling Pathways through Cross-Dataset Transfer Learning

TL;DR

This study investigates cross-dataset transfer learning for EEG pathology detection using TUAB as the source and NMT as the target, addressing data scarcity and distribution shifts inherent in real-world clinical data. It evaluates four CNN architectures, applies diverse data augmentations, and employs a cropped training regime with AdamW optimization, alongside a concatenated-data strategy and discriminative fine-tuning guided by representation similarity via CK A. The findings show that pretraining on TUAB improves NMT performance in low-data settings, but can cause catastrophic forgetting on the source; discriminative fine-tuning and CK A-informed transfer points help mitigate this. Overall, the work demonstrates the viability of transfer learning to boost EEG pathology detection, provides insights into representation transfer across datasets, and highlights practical considerations such as label noise and data scaling for robust clinical deployment.

Abstract

Pathology diagnosis based on EEG signals and decoding brain activity holds immense importance in understanding neurological disorders. With the advancement of artificial intelligence methods and machine learning techniques, the potential for accurate data-driven diagnoses and effective treatments has grown significantly. However, applying machine learning algorithms to real-world datasets presents diverse challenges at multiple levels. The scarcity of labelled data, especially in low regime scenarios with limited availability of real patient cohorts due to high costs of recruitment, underscores the vital deployment of scaling and transfer learning techniques. In this study, we explore a real-world pathology classification task to highlight the effectiveness of data and model scaling and cross-dataset knowledge transfer. As such, we observe varying performance improvements through data scaling, indicating the need for careful evaluation and labelling. Additionally, we identify the challenges of possible negative transfer and emphasize the significance of some key components to overcome distribution shifts and potential spurious correlations and achieve positive transfer. We see improvement in the performance of the target model on the target (NMT) datasets by using the knowledge from the source dataset (TUAB) when a low amount of labelled data was available. Our findings indicate a small and generic model (e.g. ShallowNet) performs well on a single dataset, however, a larger model (e.g. TCN) performs better on transfer and learning from a larger and diverse dataset.

Figures (6)

• Figure 1: A comparison of the data distribution with respect to age in two EEG datasets: TUAB and NMT. The figure shows that the statistics between splits, gender and datasets vary drastically. The age distribution differs in data splits for gender and pathological states. The number of samples in different conditions is not balanced, which may affect the performance of the models.
• Figure 2: The effect of pretraining on pathology detection balanced accuracy in the NMT dataset. The x-axis shows the number of training samples and the y-axis shows the test balanced-accuracy. The orange line represents the performance of fine-tuning a pretrained model and the blue line represents the baseline model trained from scratch. The error bars show the standard error of the mean across five random seeds.
• Figure 3: The performance on the source dataset (TUAB) while finetuning or training on the target dataset (NMT). The x-axis shows the number of training samples from the target dataset and the y-axis shows the balanced accuracy on the source dataset. The orange line represents the pretrained model that is finetuned on the target dataset and the blue line represents the baseline model that is trained from scratch on the target dataset.
• Figure 4: The similarity between different layers of networks. From the left, the first and second are within a model and the last one is cross-model similarity.
• Figure 5: The effect of discriminative fine-tuning on the balanced accuracy (BAC) of the model for the NMT dataset. The x-axis shows Discriminative Learning Rates, and the y-axis shows the BAC score. The figure shows that the model with discriminative fine-tuning achieves a higher BAC score than the baseline model, indicating that it can better adapt to the new domain and task.