Examining Effects of Schizophrenia on EEG with Explainable Deep Learning Models

Charles A. Ellis, Abhinav Sattiraju, Robyn L. Miller, Vince D. Calhoun

TL;DR

This paper introduces explainability methods to deep learning models for diagnosing schizophrenia from raw EEG data, a crucial step for clinical adoption. By analyzing which brain regions and frequency bands the models focus on, they found that both CNN and CNN-LSTM architectures prioritize specific electrodes (T8, C3) and frequency bands (delta, gamma), aligning with existing neurological findings.

What Problem Does This Paper Address?

Diagnosing schizophrenia (SZ) is challenging due to symptom variability and the lack of definitive biomarkers. While electroencephalography (EEG) offers a promising avenue for objective diagnosis, traditional machine learning approaches often require complex manual feature engineering. More recent deep learning models can automatically extract features from raw EEG, simplifying the process. However, these "black box" models lack transparency, making it difficult to understand why they make a particular diagnosis. This lack of explainability is a significant barrier to their use in healthcare, where understanding the reasoning behind a diagnosis is critical. This paper aims to bridge this gap by applying explainability techniques to deep learning models trained on raw EEG for SZ diagnosis.

Methods

The researchers developed two distinct deep learning architectures: a Convolutional Neural Network (CNN) and a CNN combined with a Long Short-Term Memory network (CNN-LSTM). Both models were trained on raw EEG data using a 10-fold cross-validation strategy. To enhance robustness, they employed data augmentation techniques, including adding Gaussian noise and a novel sinusoidal noise.

The core contribution lies in their application of perturbation-based explainability approaches:

  • Spatial Explainability: They systematically "ablated" (perturbed with noise) individual EEG channels one by one. By measuring the resulting drop in diagnostic accuracy (Balanced Accuracy - BACC), they could infer the importance of each channel to the model's decision-making.
  • Spectral Explainability: They perturbed specific frequency bands (delta, theta, alpha, beta, gamma) within the EEG data. Similar to spatial explainability, they measured the change in BACC to understand which frequency bands were most crucial for the models' SZ diagnosis.

Datasets

The study utilized a scalp EEG dataset originally presented in [6]. This dataset comprises EEG recordings from 47 individuals diagnosed with schizophrenia (SZ) and 54 healthy controls (HCs). The data was collected during a 5-minute resting state at 1000 Hz and preprocessed by downsampling to 250 Hz, segmenting into 25-second epochs, and applying techniques like random upsampling of SZ samples for class balance, outlier removal, and z-score normalization. The final dataset consisted of 5765 HC and 5764 SZ samples.

Key Insights and Ideas

  • EEG as a Biomarker: The study reinforces the potential of EEG, particularly raw EEG, as a source of biomarkers for schizophrenia, even when processed by deep learning models.
  • Shared Neural Signatures: Both the CNN and CNN-LSTM models, despite their architectural differences, converged on prioritizing similar brain regions (electrodes T8 and C3) and frequency bands (delta and gamma). This convergence suggests these are robust indicators of SZ within the EEG data.
  • Clinical Relevance of Explainability: The paper strongly advocates for explainability in clinical AI. By understanding what the models are learning, researchers can validate their findings against existing neurological knowledge and build trust for clinical deployment.
  • Novel Data Augmentation: The introduction of sinusoidal noise as a data augmentation technique is a novel contribution, potentially helping models generalize better to specific types of noise found in real-world EEG.

Key Results

  • Model Performance: Both the CNN and CNN-LSTM models achieved classification performance metrics above chance level (all above 70%). The CNN generally outperformed the CNN-LSTM in terms of overall accuracy.
  • Spatial Importance: The T8 and C3 electrodes were identified as the most important for both models. This finding aligns with previous research linking activity in the temporal lobe (near T8) and central regions (near C3) to schizophrenia.
  • Spectral Importance: The delta (δ) and gamma (γ) frequency bands were found to be most important for diagnosis. The CNN prioritized gamma bands, while the CNN-LSTM prioritized delta bands, with both showing significant importance. This also aligns with existing literature that has identified differences in these frequency bands in individuals with SZ.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Channel-Specific Frequency Analysis: While they analyzed channel importance and frequency band importance across all channels, future work could investigate the importance of specific frequency bands within each individual channel.
  • Temporal Dynamics: The current explainability methods don't fully capture the temporal dynamics of SZ effects on EEG. Future research could explore how to gain insight into the temporal distribution of importance.
  • Interpretable Classifiers: Developing more inherently interpretable classifiers, rather than relying solely on post-hoc explainability, is another promising avenue.
  • Gradient-Based Methods: The authors suggest that gradient-based explainability methods could offer further insights into learned features.
  • Attention Mechanisms: Integrating attention mechanisms into the CNN-LSTM could potentially provide a more interpretable way to understand temporal dependencies.

Summary updated

Abstract

Abstract Schizophrenia (SZ) is a mental disorder that affects millions of people globally. At this time, diagnosis of SZ is based upon symptoms, which can vary from patient to patient and create difficulty with diagnosis. To address this issue, researchers have begun to look for neurological biomarkers of SZ and develop methods for automated diagnosis. In recent years, several studies have applied deep learning to raw EEG for automated SZ diagnosis. However, the use of raw time-series data makes explainability more difficult than it is for traditional machine learning algorithms trained on manually engineered features. As such, none of these studies have sought to explain their models, which is problematic within a healthcare context where explainability is a critical component. In this study, we apply perturbation-based explainability approaches to gain insight into the spectral and spatial features learned by two distinct deep learning models trained on raw EEG for SZ diagnosis for the first time. We develop convolutional neural network (CNN) and CNN long short-term memory network (CNN-LSTM) architectures. Results show that both models prioritize the T8 and C3 electrodes and the δ- and γ-bands, which agrees with previous literature and supports the overall utility of our models. This study represents a step forward in the implementation of deep learning models for clinical SZ diagnosis, and it is our hope that it will inspire the more widespread application of explainability methods for insight into deep learning models trained for SZ diagnosis in the future.
BibTeX
@inproceedings{ellis2022examining,
  author = {Charles A. Ellis and Abhinav Sattiraju and Robyn L. Miller and Vince D. Calhoun},
  title = {{Examining Effects of Schizophrenia on EEG with Explainable Deep Learning Models}},
  booktitle = {BIBE},
  year = {2022},
  doi = {10.1109/bibe55377.2022.00068},
}

Loading related content...

Figure