Deep learning with convolutional neural networks for decoding and visualization of EEG pathology

• EEG is widely used because it is low-cost and noninvasive, supporting screening and hypothesis-based diagnosis.Source: §I · p. 1
• Reliable EEG diagnosis requires substantial time and specialized expert knowledge.Source: §I · p. 1
• Computer-aided EEG diagnosis has used diverse machine-learning methods, but the best approach across diagnoses remains unsettled.Source: §I · p. 1
• The TUH Abnormal EEG Corpus contains about 3000 recordings created to support distinguishing pathological from normal EEG.Source: §I · p. 1
• The study applies deep ConvNets to pathological-versus-normal EEG and reports better accuracies than the only published baseline identified by the authors.Source: §I · p. 1
• The work establishes an improved baseline for future EEG pathology-decoding research.Source: §I · p. 2

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.Source: A. EEG ConvNet architectures and… · p. 2
• The ConvNets were adapted to approximately 600 input samples, corresponding to 6 seconds at 100 Hz.Source: A. EEG ConvNet architectures and… · p. 2
• The deep ConvNet’s pooling strides were moved to convolutional layers immediately before pooling to increase input length without strongly increasing computation time.Source: A. EEG ConvNet architectures and… · p. 2

• Reduced-data analyses used 1, 2, 4, 8, or 16 minutes after the artifact-prone first minute, varying training data, test data, or both.Source: B. Decoding from reduced EEG time… · p. 2

C. Automatic architecture optimization

• SMAC optimized deep and shallow ConvNet hyperparameters, including filter lengths, strides, and nonlinearities, using cross-validation performance.Source: C. Automatic architecture optimization · p. 2

• Band-power analysis examined delta, theta, alpha, low-beta, high-beta, and low-gamma frequency ranges.Source: D. Visualizations of the spectral… · p. 2
• Spectral analysis computed pathological-to-normal median band-power log ratios and displayed electrode-wise results as topographical maps.Source: D. Visualizations of the spectral… · p. 2
• Input-perturbation network-prediction correlation maps measured how amplitude perturbations at different frequencies affected ConvNet predictions and provided scalp maps.Source: E. Visualizations based on the effects… · p. 2

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.Source: F. Analysis of word frequencies in the… · p. 2

• For each word, the analysis computed relative frequencies in incorrect and correct predictions and their ratio r_i = f_i−/f_i+.Source: F. Analysis of word frequencies in the… · p. 2
• Words with very large or very small frequency ratios were examined in report context to identify clinical or contextual aspects correlated with ConvNet failures.Source: F. Analysis of word frequencies in the… · p. 2
• 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.Source: Figure 1 · p. 3, Figure 2 · p. 3

• The Abnormal Corpus contains 3017 recordings: 1529 normal and 1488 pathological.Source: G. Dataset · p. 3
• The corpus derives from over 16000 clinical recordings involving more than 10000 subjects across over 12 years.Source: G. Dataset · p. 3
• Recordings were acquired from at least 21 standard electrode positions and usually contain around 20 minutes of EEG data.Source: G. Dataset · p. 3
• Preprocessing selected 21 shared electrodes, removed the first minute, capped amplitudes at ±800 µV, and resampled data to 100 Hz.Source: H. Preprocessing · p. 3

• Both ConvNets outperformed the only previously published result on the TUH Abnormal EEG Corpus.Source: A. Deep and shallow ConvNets reached… · p. 3
• The ConvNets reduced the error rate from about 21% to about 15% relative to the baseline.Source: A. Deep and shallow ConvNets reached… · p. 4
• The deep and shallow ConvNets produced very similar confusion matrices across five independent runs.Source: A. Deep and shallow ConvNets reached… · p. 3, Figure 3 · p. 4
• Both ConvNets exceeded 90% specificity and achieved approximately 75–78% sensitivity, with more errors on pathological recordings.Source: A. Deep and shallow ConvNets reached… · p. 4
• Single predictions from 6-second crops already achieved higher accuracy than the baseline, while averaging crops improved per-recording accuracy.Source: A. Deep and shallow ConvNets reached… · p. 4

• Best trialwise accuracies were reached with only 1 minute of data used for prediction.Source: B. Deep ConvNet reached best… · p. 4
• Reducing training duration decreased cropwise and trialwise test accuracy, whereas reducing test duration did not have the same effect.Source: B. Deep ConvNet reached best… · p. 4
• When only test recordings were shortened, maximal accuracies occurred with 1 minute per recording.Source: Figure 4 · p. 4
• The reduced-duration results were based on one run per condition, so slightly better performance than Table II may reflect noise.Source: Figure 4 · p. 4
• Accuracy decreased slightly with increasing recording time, but the decrease was below 0.5% and should be interpreted cautiously.Source: B. Deep ConvNet reached best… · p. 4

• The optimized deep and shallow architectures used 1.8 and 3.7 seconds of EEG, respectively, versus about 6 seconds in the original versions.Source: C. Architecture optimization yielded… · p. 5
• 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.Source: C. Architecture optimization yielded… · p. 5
• The optimized shallow ConvNet used max pooling as its only nonlinearity, omitting the original nonlinearities.Source: Figure 6 · p. 4, C. Architecture optimization yielded… · p. 5
• SMAC selected the final shallow architecture through automated hyperparameter optimization evaluated with 10-fold cross-validation on the training data.Source: C. Architecture optimization yielded… · pp. 4–5

• Pathological recordings showed broadly increased delta and theta power, decreased beta and low-gamma power, and electrode-dependent alpha changes.Source: D. Power spectra and ConvNet… · p. 5
• Both ConvNets strongly relied on delta and theta frequencies, with perturbation effects consistent with the observed spectral power differences.Source: D. Power spectra and ConvNet… · p. 5
• The words “small” and “amount” occurred 15.5 times more frequently in incorrectly than correctly predicted pathological reports, often describing subtle abnormalities.Source: E. Insights from the textual reports… · p. 5
• “Age” appeared 9.7 times more frequently, while “sleep” occurred three times in incorrect pathological reports and was absent from correct reports.Source: E. Insights from the textual reports… · p. 5
• EEG-only ConvNets lacked contextual cues such as subject age and vigilance, highlighting their potential integration into training or decoding.Source: E. Insights from the textual reports… · p. 5
• Misclassified normal reports offered little insight because they were typically very short, such as “Normal EEG.”Source: E. Insights from the textual reports… · p. 5

• 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.Source: §IV · p. 5
• 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.Source: §IV · p. 5
• Automated architecture optimization found markedly different configurations with similar accuracies, while training improvements failed to improve evaluation accuracy.Source: §IV · p. 5
• Human-level performance provides a benchmark because clinicians use EEG patterns together with available contextual information.Source: §IV · p. 6
• Perturbation visualizations linked decoding to delta and theta power changes, particularly in temporal channels, consistent with spectral analysis and reports mentioning “temporal” and “slowing.”Source: §IV · p. 6
• Clinical usefulness would benefit from visualization methods that better explain individual diagnoses, including attention mechanisms and saliency maps.Source: §IV · p. 6

• The deep ConvNets achieved the best published accuracies on the largest available dataset for decoding EEG pathology.Source: Conclusion · p. 6