TL;DR
This paper introduces Layer-wise Relevance Propagation (LRP), a gradient-based method, to explain how deep learning models use multimodal data (EEG, EOG, EMG) for sleep stage classification. Unlike traditional ablation methods, LRP provides more nuanced insights into feature importance without altering the input data, demonstrating its potential for understanding complex biomedical models.
What Problem Does This Paper Address?
As biomedical studies increasingly leverage multimodal data to enhance model performance, interpreting these complex models becomes a significant challenge. Many existing explainability methods, like ablation, require modifying input data, which can lead to out-of-distribution samples and potentially inaccurate explanations. This is particularly problematic in fields like electrophysiology, where abnormal data points might not reflect real-world scenarios. The paper addresses the need for explainability techniques that can effectively reveal the contributions of different data modalities in multimodal deep learning models without introducing such artifacts.
Methods
The authors propose using Layer-wise Relevance Propagation (LRP), a gradient-based feature attribution method, as an alternative to ablation for explaining multimodal deep learning models. LRP works by assigning a "relevance" score to each input feature based on its contribution to the model's output. This relevance is then propagated backward through the network layers. The paper utilizes two LRP rules: the epsilon (ɛ) rule and the alpha-beta (αβ) rule, with variations in their parameters to explore different aspects of relevance. They trained a 1D Convolutional Neural Network (CNN) on multimodal electrophysiology data for automated sleep stage classification. The LRP method was then applied globally to understand the overall importance of each modality (EEG, EOG, EMG) for classification across all samples and locally for specific classification outcomes (correct vs. incorrect).
Datasets
The study used the Sleep Telemetry dataset from the Sleep-EDF Database [10] on Physionet [11]. This dataset comprises approximately 9-hour recordings from 22 subjects, including EEG, EOG, and EMG data sampled at 100 Hz, along with polysomnogram labels for sleep stages. The dataset was preprocessed by segmenting recordings into 30-second epochs, consolidating NREM3 and NREM4 stages, and z-scoring each modality.
Key Insights and Ideas
- LRP offers a more robust alternative to ablation: By avoiding data modification, LRP circumvents the issue of creating out-of-distribution samples, leading to potentially more reliable explanations.
- Multimodal contributions vary by sleep stage: The study reveals that different modalities play distinct roles in classifying various sleep stages, aligning with established clinical knowledge.
- Consistency in correct classifications, variability in errors: LRP showed consistent modality importance for correctly classified samples across different cross-validation folds, suggesting the model learns generalizable features. Conversely, inconsistent importance for misclassified samples hints at the model making different types of errors or learning different patterns in each fold.
- Gradient-based methods provide deeper insights: LRP can differentiate between features supporting a correct classification and those that might be misleading, offering a richer understanding than methods that simply indicate overall importance.
Key Results
- Overall Modality Importance: Across all sleep stages and samples, EEG was consistently found to be the most important modality for classification, followed by EOG, and then EMG.
- Stage-Specific Importance:
- EEG and EOG showed higher relevance for classifying awake and NREM1 stages.
- EOG was most important for classifying REM sleep.
- EEG was most relevant for classifying NREM2 and NREM3 stages.
- Consistency of Explanations: LRP demonstrated consistent modality importance across cross-validation folds for correctly classified samples, indicating stable feature learning. For incorrectly classified samples, the importance varied more significantly across folds.
Limitations and Future Work
- Exploration of other GBFA methods: The authors acknowledge that other gradient-based feature attribution (GBFA) methods might offer superior explanations.
- Parameter tuning for LRP: While they explored different LRP rules and parameters, further investigation into optimal parameter settings for specific tasks could be beneficial.
- Model Architecture: The CNN architecture was adapted from a previous EEG-focused model. Examining architectures specifically designed to extract features optimally from EOG and EMG could yield more refined explainability results.
- Classifier Performance: The study's primary focus was on explainability, and the classifier's performance was not state-of-the-art. Applying LRP with a higher-performing classifier could lead to more impactful biomarker discovery.
- Local vs. Global Explanations: While global explanations were provided, local LRP explanations could offer higher-resolution insights for identifying novel multimodal biomarkers.
- Quantitative Evaluation of Explanations: The authors suggest applying metrics developed for quantifying explanation quality to further assess the effectiveness of different methods.
Summary updated