Using EEG Frequency Attributions to Explain the Classifications of a Deep Neural Network for Sleep Staging

Electroencephalography (EEG) signals contain rich frequency information, which is difficult to assess using many state-of-the-art post hoc explainable AI (XAI) methods that typically provide attributions in the time domain. To address this, we introduce a novel post hoc XAI framework, entitled FLEX (Frequency Layer Explanation), which expands the Integrated Gradients (IG) framework with the Discrete Cosine Transform (DCT) to generate attributions in the frequency domain instead of the time domain. To demonstrate the effectiveness of this framework, we performed two experiments: (i) Using a simple deep neural network (DNN) trained on synthetic EEG signals, i.e., sine waves. (ii) Using a state-of-the-art DNN trained on real EEG signals stemming from the Sleep Heart Health Study (SHHS) to analyze the association between frequency attributions and predicted sleep stages. Our results demonstrate that (i) the resulting attributions of simulated EEG signals match the used input frequencies for generating the synthetic signals, confirming its effectiveness in controlled settings. (ii) The real-world validation demonstrates that the DNN's attributions for sleep staging align with established medical knowledge, e.g the high relevance of delta waves as a marker of deep sleep. This highlights the potential of FLEX to complement existing post hoc XAI workflows.Clinical relevance- The FLEX framework bridges the gap between black-box DNNs and medical textbook knowledge, potentially enhancing clinicians trust in AI applications.