Technical Summary: A Conflict-aware Evidential Framework for Reliable Sleep Stage Classification

Problem Statement

Automatic sleep stage classification is a critical task for diagnosing sleep disorders, traditionally performed manually by experts using polysomnography (PSG) data, a process that is time-consuming and labor-intensive. While machine learning models have accelerated this process, existing multi-view methods (utilizing multiple modalities like EEG and EOG) face a significant reliability challenge. Most current approaches assume that different views are well-aligned and assign fixed or equal weights to them. However, in real-world scenarios, different modalities often provide conflicting information due to signal noise, physiological variability, or inherent differences in signal characteristics. Existing methods frequently fail to detect or appropriately handle these inter-view conflicts, leading to compromised staging reliability. Furthermore, standard deep learning models often produce overconfident predictions even when faced with low-quality or conflicting data.

Methodology: ConfSleepNet

The authors propose ConfSleepNet, a conflict-aware evidential framework designed to dynamically resolve inter-view conflicts and produce reliable joint decisions. The framework operates in two primary phases: multi-view evidence extraction and conflict-aware aggregation.

1. Multi-View Evidence Extraction with Hybrid Structures

Unlike traditional methods that treat all modalities with a uniform five-class classification structure, ConfSleepNet leverages the physiological complementarity of EEG and EOG signals through a hybrid category structure:

• EEG and Multi-view Inputs: Processed via evidential Deep Neural Networks (DNNs) ($f_1$ and $f_2$) to learn fine-grained evidence for all five sleep stages (Wake, N1, N2, N3, REM).
• EOG Inputs: Processed via evidential DNNs ($f_3$ and $f_4$) to learn coarse-grained evidence for three categories (Wake, REM, and a combined NREM stage). This design acknowledges that EOG is particularly effective for identifying REM but less discriminative for specific NREM sub-stages compared to EEG.

The architecture employs parallel convolutional branches with different kernel sizes, enhanced by GCNet-1D attention modules and cross-attention mechanisms to capture time-frequency features and inter-modal interactions. A Bidirectional LSTM layer models stage-transition patterns across consecutive epochs.

2. Evidence Mapping

Before fusion, the coarse-grained evidence vectors (3-class) from EOG-based networks are mapped to fine-grained vectors (5-class) using a fixed evidence mapping layer. This transformation distributes the evidence for the combined NREM category equally among its sub-stages (N1, N2, N3) while preserving the total evidence mass for Wake and REM.

3. Conflict-Aware Multi-View Aggregation (CMAM)

The core innovation lies in the aggregation phase, which utilizes Evidential Deep Learning (EDL) and Subjective Logic:

• Opinion Construction: Each view's evidence is converted into a Dirichlet distribution, generating a "subjective opinion" consisting of belief masses (prediction probabilities) and an uncertainty mass.
• Conflict Metric: A novel metric quantifies the degree of conflict between any two opinions based on the divergence of their belief masses.
• Aggregation Rule: The framework introduces a Conflict-Aware Multi-View Aggregation (CMAM) method. Unlike previous methods that may blindly reduce uncertainty when adding consistent views or fail to account for conflict, CMAM explicitly adjusts the final joint opinion based on the conflict degree:
  • Consistent Views: Incorporating consistent opinions reduces overall uncertainty, increasing confidence.
  • Conflicting Views: Incorporating a conflicting opinion with high uncertainty increases the overall uncertainty of the joint decision, effectively down-weighting unreliable views rather than forcing a consensus.

4. Training Strategy

The model is trained jointly using a loss function that combines:

• Accuracy Loss: Encourages evidence for the ground-truth category to approach the total evidence.
• KL Divergence: Penalizes the model for assigning high evidence to incorrect classes, encouraging the total evidence to shrink (high uncertainty) for misclassified samples. An annealing coefficient is used to gradually increase the weight of the KL term to prevent premature convergence.

Key Contributions

1. ConfSleepNet Framework: The proposal of an evidential framework that introduces a hybrid category structure (coarse-grained for EOG, fine-grained for EEG) to facilitate differentiated and physiologically aligned evidence learning.
2. Conflict-Aware Aggregation: The development of a novel aggregation method (CMAM) that explicitly quantifies and manages inter-view conflicts. Theoretical analysis demonstrates that this method correctly reduces uncertainty when views are consistent and increases uncertainty when views conflict, addressing a limitation in existing EDL-based fusion methods.
3. Empirical Validation: Extensive evaluation on four public datasets (SleepEDF-20, SleepEDF-78, MASS-SS3, and SHHS) showing superior performance over state-of-the-art baselines, including single-view, multi-view, and uncertainty-aware methods.

Experimental Results

The paper reports that ConfSleepNet achieves state-of-the-art performance across all tested datasets:

• Accuracy and F1-Score: ConfSleepNet outperforms baselines like DeepSleepNet, XSleepNet, and the current uncertainty-aware state-of-the-art TMCEK. For instance, on the SleepEDF-20 dataset, it achieved an accuracy of 87.2% and a macro-F1 of 81.8%, surpassing the next best method by a margin of 0.4% to 2.0% in accuracy.
• Specific Stage Improvement: The model shows particular improvement in classifying REM and Wake stages, which are often prone to confusion due to feature similarities.
• Ablation Studies: Experiments confirm that both the hybrid category structure and the conflict-aware aggregation strategy are essential. Removing the conflict-aware mechanism (replacing it with average fusion) or using a uniform category structure resulted in performance drops.
• Conflict Analysis: Visualization of conflict degrees reveals that inter-view conflicts are significantly higher during sleep stage transitions (e.g., N2 to N3) compared to stable stages, validating the necessity of a conflict-aware mechanism.
• Generalizability: The CMAM method was also tested on general multi-view benchmarks (HandWritten, Scene15, CUB, PIE), where it consistently outperformed other evidential and fusion methods, suggesting its applicability beyond sleep staging.

Significance and Claims

The authors claim that ConfSleepNet provides a more robust and reliable approach to automatic sleep staging by explicitly modeling the uncertainty and potential conflicts inherent in multi-modal physiological data.

• Clinical Relevance: By improving the reliability of staging, particularly for clinically critical stages like REM, the method holds positive implications for the practical application of automated sleep disorder diagnosis.
• Theoretical Soundness: The paper provides theoretical proofs demonstrating that the proposed aggregation method satisfies the principles of reducing uncertainty with consistent views and increasing it with conflicting views, a property often lacking in existing fusion techniques.
• General Utility: The framework's success on general multi-view datasets suggests that the conflict-aware evidential approach is a generalizable solution for reliable decision-making in scenarios involving heterogeneous and potentially conflicting data sources.

The work concludes that by dynamically adjusting the importance of each view based on conflict assessment, ConfSleepNet effectively mitigates the risks of misleading predictions, offering a significant step forward in the automation of sleep staging.