Amplifying pathological detection in EEG signaling pathways through cross-dataset transfer learning.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Electroencephalogram (EEG) signal analysis is crucial for diagnosing neurological disorders.
• Machine learning (ML) offers potential for data-driven diagnoses but faces challenges with real-world data.
• Scarcity of labeled data, especially in low-data regimes, necessitates advanced techniques like transfer learning.

Purpose of the Study:

• To investigate the effectiveness of data and model scaling and cross-dataset knowledge transfer in EEG pathology classification.
• To highlight challenges such as negative transfer and distribution shifts in ML applications.
• To demonstrate the benefits of transfer learning in low-data scenarios for neurological disorder diagnosis.

Main Methods:

• Exploration of data scaling techniques and their impact on classification performance.
• Application of transfer learning by leveraging knowledge from a source dataset (TUAB) to a target dataset (NMT).
• Comparison of model performance, including small (ShallowNet) and large (TCN) models, in single-dataset and transfer learning settings.

Main Results:

• Data scaling yielded varying performance improvements, emphasizing the need for careful data evaluation and labeling.
• Successful knowledge transfer from the TUAB dataset improved target model performance on NMT datasets with limited labeled data.
• Larger models (TCN) demonstrated superior performance in transfer learning compared to smaller models (ShallowNet) on diverse datasets.

Conclusions:

• Transfer learning is effective for EEG pathology classification in low-data scenarios.
• Addressing distribution shifts and spurious correlations is key to achieving positive transfer.
• Model size and dataset diversity significantly influence transfer learning effectiveness, with larger models benefiting more from diverse data.