EEG Pathology Detection via Clinical Concept Bottleneck Models

Deep learning models have been developed for detecting EEG pathology, achieving high accuracy; however, they lack interpretability. This paper presents concept bottleneck models (CBMs) for EEG pathology detection, which utilize clinically computed EEG biomarkers as a bottleneck to aid in prediction. These concepts encompass band power, spectral ratios, alpha peak frequency and amplitude, hemispheric asymmetry, and phase locking values.

The backbone of the CBM is a shallow convolutional neural network (CNN). The paper compares a black box baseline model with two CBM variants: hybrid and pure.

The hybrid variant receives a residual projection of the backbone features, while the pure CBM version uses only the predicted concepts as input to the classifier. The models were evaluated on two tasks: abnormality detection and dementia classification. Interestingly, both CBM approaches matched the performance of the baseline model, achieving 82.7-82.9% on TUAB and 57.9-60.5% on CAUEEG, compared to 83.6% for the baseline.

However, the pure CBM exhibited a concern known as concept fidelity, where the predicted concepts deviated from their actual values. This deviation led to a significant loss in accuracy, with a 26-point difference on TUAB and a 22-point difference on CAUEEG.

To mitigate this issue, the researchers increased the weight of the concept loss, aligning it with the ground truth concepts. While this approach reduced the accuracy cost on TUAB, it resulted in a more substantial accuracy loss on CAUEEG. The experiment code and full logs are available at https://www.github.com/mhmdrdwn/cbm.