Source-linked AI summary
Deep learning with convolutional neural networks for decoding and visualization of EEG pathology
Robin Tibor Schirrmeister, Lukas Gemein, Katharina Eggensperger, Frank Hutter, Tonio Ball
TL;DR
EEG diagnosis requires substantial expert time and specialized knowledge, motivating machine-learning assistance. This study applies shallow and deep ConvNets to distinguish pathological from normal EEG, achieving the best reported accuracies on the TUH EEG Abnormal Corpus and supporting progress toward automated diagnosis.
Problem
EEG diagnosis requires expert time and specialized knowledge, limiting its accessibility and efficiency.
Method
The study applies shallow and deep ConvNet architectures to raw time-domain EEG for distinguishing pathological from normal recordings.
Results
Both ConvNets exceeded the published baseline by more than 5%, while high accuracies were also achieved from individual 6-second EEG crops.
Takeaways & Limitations
The ConvNets achieved the best published accuracies on the TUH EEG Abnormal Corpus and represent a step toward clinically useful automated EEG diagnosis.
Takeaways & Limitations
The amount of EEG required and the design of models integrating information across longer timescales remain to be determined.
Abstract
from arXiv · showhide
We apply convolutional neural networks (ConvNets) to the task of distinguishing pathological from normal EEG recordings in the Temple University Hospital EEG Abnormal Corpus. We use two basic, shallow and deep ConvNet architectures recently shown to decode task-related information from EEG at least as well as established algorithms designed for this purpose. In decoding EEG pathology, both ConvNets reached substantially better accuracies (about 6% better, ~85% vs. ~79%) than the only published result for this dataset, and were still better when using only 1 minute of each recording for training and only six seconds of each recording for testing. We used automated methods to optimize architectural hyperparameters and found intriguingly different ConvNet architectures, e.g., with max pooling as the only nonlinearity. Visualizations of the ConvNet decoding behavior showed that they used spectral power changes in the delta (0-4 Hz) and theta (4-8 Hz) frequency range, possibly alongside other features, consistent with expectations derived from spectral analysis of the EEG data and from the textual medical reports. Analysis of the textual medical reports also highlighted the potential for accuracy increases by integrating contextual information, such as the age of subjects. In summary, the ConvNets and visualization techniques used in this study constitute a next step towards clinically useful automated EEG diagnosis and establish a new baseline for future work on this topic.
I. Introduction
EEG diagnosis is clinically valuable but constrained by the time and expertise required for reliable interpretation. This study applies deep ConvNets to distinguish normal from pathological EEG using the TUH Abnormal EEG Corpus.
- EEG is widely used because it is low-cost and noninvasive, supporting screening and hypothesis-based diagnosis.
- Reliable EEG diagnosis requires substantial time and specialized expert knowledge.
- Computer-aided EEG diagnosis has used diverse machine-learning methods, but the best approach across diagnoses remains unsettled.
- The TUH Abnormal EEG Corpus contains about 3000 recordings created to support distinguishing pathological from normal EEG.
- The study applies deep ConvNets to pathological-versus-normal EEG and reports better accuracies than the only published baseline identified by the authors.
- The work establishes an improved baseline for future EEG pathology-decoding research.
II. Methods
The study evaluates deep and shallow ConvNets adapted for six-second raw EEG inputs and tests how recording duration affects training and prediction.
- EEG ConvNet architectures and training: The study uses generic deep and band-power-focused shallow ConvNet architectures previously shown to decode raw time-domain EEG competitively with hand-engineered methods.
- EEG ConvNet architectures and training: The ConvNets were adapted to approximately 600 input samples, corresponding to 6 seconds at 100 Hz.
- EEG ConvNet architectures and training: The deep ConvNet’s pooling strides were moved to convolutional layers immediately before pooling to increase input length without strongly increasing computation time.
- Reduced-data analyses used 1, 2, 4, 8, or 16 minutes after the artifact-prone first minute, varying training data, test data, or both.
C. Automatic architecture optimization
The study combines automated architecture optimization with spectral analysis and input-perturbation visualizations to investigate ConvNet design and EEG decoding behavior.
- C. Automatic architecture optimization: SMAC optimized deep and shallow ConvNet hyperparameters, including filter lengths, strides, and nonlinearities, using cross-validation performance.
- Band-power analysis examined delta, theta, alpha, low-beta, high-beta, and low-gamma frequency ranges.
- Spectral analysis computed pathological-to-normal median band-power log ratios and displayed electrode-wise results as topographical maps.
- Input-perturbation network-prediction correlation maps measured how amplitude perturbations at different frequencies affected ConvNet predictions and provided scalp maps.
F. Analysis of word frequencies in the medical reports
The study analyzes word-frequency differences between correctly and incorrectly predicted clinical reports to identify contextual aspects associated with ConvNet failures.
- F. Analysis of word frequencies in the medical reports: Clinical reports were analyzed separately for normal and pathological recordings to compare words in incorrectly versus correctly predicted cases.
- For each word, the analysis computed relative frequencies in incorrect and correct predictions and their ratio r_i = f_i−/f_i+.
- Words with very large or very small frequency ratios were examined in report context to identify clinical or contextual aspects correlated with ConvNet failures.
- The deep and shallow architectures differ in their schematic network designs, with inputs and feature maps shown as black cuboids and convolution/pooling kernels as brown cuboids.
G. Dataset
The TUH EEG Abnormal Corpus contains manually labeled normal and pathological clinical EEG recordings, with recordings drawn from a much larger longitudinal corpus. Recordings use standard electrode positions and typically provide about 20 minutes of EEG data.
- The Abnormal Corpus contains 3017 recordings: 1529 normal and 1488 pathological.
- The corpus derives from over 16000 clinical recordings involving more than 10000 subjects across over 12 years.
- Recordings were acquired from at least 21 standard electrode positions and usually contain around 20 minutes of EEG data.
- Preprocessing selected 21 shared electrodes, removed the first minute, capped amplitudes at ±800 µV, and resampled data to 100 Hz.
III. Results
Both deep and shallow ConvNets outperformed the published TUH Abnormal EEG baseline. Their performance remained strong on short EEG inputs, with high specificity and comparable confusion patterns.
- Both ConvNets outperformed the only previously published result on the TUH Abnormal EEG Corpus.
- The ConvNets reduced the error rate from about 21% to about 15% relative to the baseline.
- The deep and shallow ConvNets produced very similar confusion matrices across five independent runs.
- Both ConvNets exceeded 90% specificity and achieved approximately 75–78% sensitivity, with more errors on pathological recordings.
- Single predictions from 6-second crops already achieved higher accuracy than the baseline, while averaging crops improved per-recording accuracy.
B. Deep ConvNet reached best accuracies using only 1 minute per test-recording
The deep ConvNet achieved its best trialwise accuracy using only one minute of test-recording data. Accuracy declined only slightly over recording time, and the reduced-duration analyses require cautious interpretation because they used one run each.
- Best trialwise accuracies were reached with only 1 minute of data used for prediction.
- Reducing training duration decreased cropwise and trialwise test accuracy, whereas reducing test duration did not have the same effect.
- When only test recordings were shortened, maximal accuracies occurred with 1 minute per recording.
- The reduced-duration results were based on one run per condition, so slightly better performance than Table II may reflect noise.
- Accuracy decreased slightly with increasing recording time, but the decrease was below 0.5% and should be interpreted cautiously.
C. Architecture optimization yielded unexpected new models
Automated architecture optimization produced ConvNets that differed markedly from the hand-designed models, including a shallow network using max pooling as its only nonlinearity. Training improvements did not transfer to the evaluation set.
- The optimized deep and shallow architectures used 1.8 and 3.7 seconds of EEG, respectively, versus about 6 seconds in the original versions.
- Training performance improved by 2.1% for the deep ConvNet and 1.4% for the shallow ConvNet, but evaluation performance deteriorated by 0.9% and 1.5%, respectively.
- The optimized shallow ConvNet used max pooling as its only nonlinearity, omitting the original nonlinearities.
- SMAC selected the final shallow architecture through automated hyperparameter optimization evaluated with 10-fold cross-validation on the training data.
D. Power spectra and ConvNet visualizations
Pathological EEG showed increased low-frequency power and decreased beta and low-gamma power, while ConvNet perturbation maps indicated reliance on delta and theta activity. Clinical-report words also exposed error patterns and missing context.
- Pathological recordings showed broadly increased delta and theta power, decreased beta and low-gamma power, and electrode-dependent alpha changes.
- Both ConvNets strongly relied on delta and theta frequencies, with perturbation effects consistent with the observed spectral power differences.
- The words “small” and “amount” occurred 15.5 times more frequently in incorrectly than correctly predicted pathological reports, often describing subtle abnormalities.
- “Age” appeared 9.7 times more frequently, while “sleep” occurred three times in incorrect pathological reports and was absent from correct reports.
- EEG-only ConvNets lacked contextual cues such as subject age and vigilance, highlighting their potential integration into training or decoding.
- Misclassified normal reports offered little insight because they were typically very short, such as “Normal EEG.”
IV. Discussion
The ConvNets achieved the best published accuracies on the TUH EEG Abnormal Corpus and their behavior aligned with spectral and textual evidence. Further gains may require longer-timescale modeling, richer clinical context, and more interpretable visualizations.
- The ConvNet architectures achieved the best published accuracies on the TUH EEG Abnormal Corpus and were only slightly modified from earlier task-related EEG architectures.
- Good accuracies from very limited EEG segments suggest that models using longer timescales could provide further accuracy improvements, although the required amount of EEG remains undetermined.
- Automated architecture optimization found markedly different configurations with similar accuracies, while training improvements failed to improve evaluation accuracy.
- Human-level performance provides a benchmark because clinicians use EEG patterns together with available contextual information.
- Perturbation visualizations linked decoding to delta and theta power changes, particularly in temporal channels, consistent with spectral analysis and reports mentioning “temporal” and “slowing.”
- Clinical usefulness would benefit from visualization methods that better explain individual diagnoses, including attention mechanisms and saliency maps.
Conclusion
The deep ConvNets achieved the best published accuracies on the largest available dataset for EEG pathology, advancing clinically useful automated EEG diagnosis.
- The deep ConvNets achieved the best published accuracies on the largest available dataset for decoding EEG pathology.