[Paper Review] Addressing Class Imbalance in Classification Problems of Noisy Signals by using Fourier Transform Surrogates

Justus T. C. Schwabedal, John Snyder+3 more|arXiv (Cornell University)|Jun 20, 2018

EEG and Brain-Computer InterfacesNeuroscience19 citations

TL;DR

This paper proposes using Fourier Transform surrogates to address class imbalance in biomedical signal classification, particularly for rare sleep stages in noisy EEG/EOG/EMG data. By generating statistically valid synthetic examples via phase randomization and applying them to data augmentation, the method improved the mean F1-score by 7% in a convolutional neural network for sleep staging, while also enabling novel saliency maps to interpret model decisions.

ABSTRACT

Randomizing the Fourier-transform (FT) phases of temporal-spatial data generates surrogates that approximate examples from the data-generating distribution. We propose such FT surrogates as a novel tool to augment and analyze training of neural networks and explore the approach in the example of sleep-stage classification. By computing FT surrogates of raw EEG, EOG, and EMG signals of under-represented sleep stages, we balanced the CAPSLPDB sleep database. We then trained and tested a convolutional neural network for sleep stage classification, and found that our surrogate-based augmentation improved the mean F1-score by 7%. As another application of FT surrogates, we formulated an approach to compute saliency maps for individual sleep epochs. The visualization is based on the response of inferred class probabilities under replacement of short data segments by partial surrogates. To quantify how well the distributions of the surrogates and the original data match, we evaluated a trained classifier on surrogates of correctly classified examples, and summarized these conditional predictions in a confusion matrix. We show how such conditional confusion matrices can qualitatively explain the performance of surrogates in class balancing. The FT-surrogate augmentation approach may improve classification on noisy signals if carefully adapted to the data distribution under analysis.

Motivation & Objective

• To address class imbalance in biomedical signal classification, especially for under-represented sleep stages like S1 and S3.
• To develop a data-augmentation method that preserves the statistical properties of rare signal classes without overfitting.
• To create interpretable saliency maps using partial surrogate replacement to reveal which signal features drive model predictions.
• To evaluate surrogate quality by analyzing classifier performance on surrogates, using conditional confusion matrices.
• To explore whether IAAFT surrogates improve generalization over simple FT surrogates in low-data regimes.

Proposed method

• Generate Fourier transform surrogates by preserving the amplitude spectrum of original signals while randomizing the phase components.
• Use these surrogates to augment under-represented sleep stages in the CAPSLPDB database, creating synthetic training examples.
• Train a 1D convolutional neural network on the augmented dataset for sleep stage classification.
• Construct saliency maps by replacing short signal segments with partial surrogates and measuring changes in predicted class probabilities.
• Evaluate surrogate fidelity by testing a trained classifier on surrogates of correctly classified examples and summarizing results in conditional confusion matrices.
• Compare simple FT surrogates with IAAFT (Iterated Amplitude Adjusted Fourier) surrogates to assess their impact on model performance and generalization.

Experimental results

Research questions

• RQ1 Can Fourier transform surrogates effectively augment imbalanced biomedical signal datasets to improve deep learning model performance?
• RQ2 How do surrogate-based saliency maps reveal which temporal features are most influential in classification decisions?
• RQ3 Do IAAFT surrogates lead to better generalization than standard FT surrogates in low-data regimes?
• RQ4 To what extent do surrogate-based conditional confusion matrices reflect the true data distribution and model robustness?
• RQ5 Can partial surrogate replacement detect transient features like K-complexes that are critical for distinguishing sleep stages?

Key findings

• The proposed surrogate-based data augmentation improved the mean F1-score by 7% in sleep stage classification, demonstrating significant performance gains on imbalanced data.
• Saliency maps generated via partial surrogate replacement successfully highlighted the relevance of K-complexes in distinguishing REM from S2 stages, even when overall accuracy was low.
• Conditional confusion matrices computed on surrogates revealed that IAAFT surrogates produced higher accuracy and lower label mixing than simple surrogates, suggesting better distributional fidelity.
• The method detected that simple surrogates failed to preserve localized features like transient alpha waves, leading to misleading visual representations and potential misinterpretation.
• The results suggest that IAAFT surrogates may reduce overfitting risks compared to standard surrogates, though the underlying distributional topology remains poorly understood.
• The approach enables systematic identification of ambiguous or mislabeled epochs due to transient events, offering a path to improve data quality in sleep staging.