Related Experiment Video
Updated: Jun 10, 2025
08:22
Author Spotlight: Advancing the Study of Brain-Heart Interplay with a Comprehensive EEGLAB Plugin for Multimodal Signal Analysis
Published on: April 26, 2024
1.7K
Concept-based AI interpretability in physiological time-series data: Example of abnormality detection in
Alexander Brenner1, Felix Knispel2, Florian P Fischer3
1Institute of Medical Informatics, University of Münster, Münster, Germany.
Computer Methods and Programs in Biomedicine
|October 12, 2024
Summary
Testing with Concept Activation Vectors (TCAV) enhances deep learning interpretability for electroencephalography (EEG) data. This method shows promise in explaining model behavior and detecting data biases in physiological time series.
Area of Science:
- Artificial Intelligence
- Biomedical Engineering
- Computational Neuroscience
Background:
- Deep learning models face clinical adoption barriers due to their opaque nature.
- Feature-based explanations have limitations, driving interest in concept-based interpretability.
- Testing with Concept Activation Vectors (TCAV) uses human-understandable concepts to explain model behavior.
Purpose of the Study:
- To explore TCAV for building interpretable deep learning models on physiological time series.
- To apply TCAV to abnormality detection in electroencephalography (EEG) data.
- To investigate concept definition strategies using metadata and signal characteristics.
Main Methods:
- Utilized the XceptionTime model for multi-channel EEG data analysis.
- Developed concept definition strategies through metadata mining and frequency extraction.
- Evaluated approach robustness using hospital-collected EEG data.
Main Results:
- TCAV scores aligned with clinical expectations, with known EEG pathology concepts scoring higher than neutral concepts.
- Concept generation strategies yielded consistent results.
- Demonstrated TCAV's applicability to EEG abnormality detection.
Conclusions:
- TCAV can enhance the interpretability of deep learning models for multi-channel physiological signals.
- The method shows potential for identifying data biases.
- Further research is needed to refine concept definition and validation for clinical relevance.
