PaperlayerView exact revision on arXiv

Source-linked AI summary

Neonatal seizure detection from raw multi-channel EEG using a fully convolutional architecture

Alison O'Shea, Gordon Lightbody, Geraldine Boylan, Andriy Temko

arXiv:2105.13854v1cs.LGcs.AIstat.AP

TL;DR

Many existing neonatal seizure algorithms relied on hand-crafted features and precise labels, while deep learning work on neonatal seizure detection was limited. This paper presents an end-to-end CNN framework using raw EEG; its FCN outperforms an SVM-based system using hand-crafted features.

  • Problem

    Most existing data-driven neonatal seizure algorithms relied on hand-crafted features and extremely precise seizure labels.

    Source: §1 · p. 2

  • Method

    The paper develops a novel end-to-end convolutional neural network that operates on raw EEG signals.

    Source: §5 · p. 26

  • Results

    The FCN outperforms an SVM-based system that uses hand-crafted features.

    Source: §5 · p. 26

  • Takeaways & Limitations

    The conclusion presents raw-EEG representation learning with an FCN as an alternative to hand-crafted-feature seizure detection.

    Source: §5 · p. 26

    Most earlier data-driven neonatal seizure algorithms required extremely precise seizure labels.

    Source: §1 · p. 2

Abstract

from arXiv · show

A deep learning classifier for detecting seizures in neonates is proposed. This architecture is designed to detect seizure events from raw electroencephalogram (EEG) signals as opposed to the state-of-the-art hand engineered feature-based representation employed in traditional machine learning based solutions. The seizure detection system utilises only convolutional layers in order to process the multichannel time domain signal and is designed to exploit the large amount of weakly labelled data in the training stage. The system performance is assessed on a large database of continuous EEG recordings of 834h in duration; this is further validated on a held-out publicly available dataset and compared with two baseline SVM based systems. The developed system achieves a 56% relative improvement with respect to a feature-based state-of-the art baseline, reaching an AUC of 98.5%; this also compares favourably both in terms of performance and run-time. The effect of varying architectural parameters is thoroughly studied. The performance improvement is achieved through novel architecture design which allows more efficient usage of available training data and end-to-end optimisation from the front-end feature extraction to the back-end classification. The proposed architecture opens new avenues for the application of deep learning to neonatal EEG, where the performance becomes a function of the amount of training data with less dependency on the availability of precise clinical labels.

1. Introduction

The introduction motivates automated EEG seizure detection as clinical decision support amid limited interpretation expertise and the prevalence of subclinical seizures. It presents end-to-end CNNs operating on raw time-domain EEG and weak seizure labels as the paper’s approach.

1. Introduction

  • Approximately two thirds of neonatal seizures occur only at the sub-clinical level, while many seizures lack physical signs and NICU expertise to interpret EEG is limited.The introduction notes that diagnosis in practice is still based only on clinical signs, reliably detecting as little as 10% of seizure events.Source: §1 · p. 2
  • Most prior data-driven algorithms relied on handcrafted features and extremely precise seizure labels, which require substantial time, skill, and niche expertise.The introduction also notes that complex features can increase feature-extraction time and that feature design relies on prior knowledge of neonatal EEG.Source: §1 · pp. 2–3
  • The paper’s FCN framework differs from a prior CNN feature extractor followed by a decision-tree classifier by using an end-to-end deep-learning framework.The introduction describes the earlier CNN-plus-decision-tree approach as stacking.Source: §1 · p. 3
  • Prior experiments with raw EEG as input showed mixed results: one setup performed better, while other setups showed no reduction in performance.The introduction also cites computational efficiency of training deep-learning algorithms with raw time-series inputs.Source: §1 · p. 3
  • The proposed approach uses CNNs on time-domain EEG and a single end-to-end optimization routine, with weak seizure labels that require less clinical labor to obtain.The introduction contrasts this with baselines requiring annotations precisely located in time and space.Source: §1 · p. 4

2.1 Datasets

The Cork and Helsinki datasets provide complementary neonatal EEG cohorts, while their annotations distinguish event timing from channel-localized seizure labels. The section also explains why weak labels are useful and strong labels costly, and describes the evolving patterns used to identify neonatal seizures.

2.1.1 Dataset I – Cork

  • 834 hours of EEG from 18 full-term newborns with hypoxic ischemic encephalopathy and 1,389 seizure events comprise the Cork dataset.Eight bipolar channels were recorded at 256 Hz; two independent neonatal electro-encephalographers annotated seizure events.Source: §2.1.1 · p. 5

2.1.2 Dataset II – Helsinki

  • The Helsinki test dataset contains 18-channel EEG from 79 full-term babies in 112 hours of excerpts, including 342 seizure events.Thirty-nine babies experienced seizures, and annotations were based on consensus among three EEG experts.Source: §2.1.2 · pp. 5–6

2.1.3 Neonatal seizure annotations: weak and strong labels

  • Weak labels specify seizure start and end times but not involved channels, whereas strong labels localize activity and are costly to acquire.Weak labels suffice for temporal detection evaluation but not for extracting channel-specific seizure signatures; only a minority of Cork events had strong labels.Source: §2.1.3 · pp. 6–7

2.1.4 Neonatal seizures

  • Neonatal seizures are repetitive, evolving EEG patterns lasting over ten seconds with voltage above 2 μV.They can evolve across brain regions and in frequency; seizure burden is the percentage of time spent in seizure and is associated with more adverse outcomes.Source: §2.1.4 · pp. 7–8

2.2 Machine learning approaches

The section describes two feature-based SVM baselines, which extract features from individual EEG channels and require strong labels. It details their feature sets and processing, then contrasts their training datasets and post-processing routines.

2.2 Machine learning approaches

  • Both feature-based SVM approaches extract seizure signatures from individual EEG channels, requiring strong labels that specify seizure time and location.The section contrasts a data-driven approach using relatively simple features with one using more complex, hand-engineered features.Source: §2.2 · p. 8

2.2.1 SVM Cork

  • The Cork SVM uses 55 engineered features from 8-second EEG windows and a Gaussian-kernel SVM, then combines channel probabilities during post-processing.Its pipeline filters and downsamples raw EEG, uses 4-second shifts, and post-processes with a one-minute moving average, background adaptation, and a collar.Source: §2.2.1 · p. 9

2.2.2 SVM Helsinki

  • The Helsinki SVM uses 21 features from 32-second windows, including bespoke time-frequency correlation features, with a 4-second shift.The engineered features are reported to correlate well with neonatal seizures and require more computation than the Cork feature set.Source: §2.2.2 · p. 9

2.2.3 Comparison of SVM algorithms

  • The Cork algorithm was developed on long EEG recordings with around 10% seizure burden, whereas Helsinki used short recordings with around 20% burden and smoothing and decision making on a 48-second window.The CNN directly substitutes for the feature extraction and SVM blocks, while Cork SVM preprocessing and post-processing routines are applied.Source: §2.2.3 · p. 10, Figure 2 · p. 10

2.3 Deep learning for neonatal seizure detection

The section develops a fully convolutional approach that learns from raw EEG rather than relying on hand-engineered features. It explains the architecture and how its depth and pooling choices affect the receptive field.

2.3 Deep learning for neonatal seizure detection

  • Raw-waveform inputs outperformed image-representation architectures in the experiments, and showed an ability to learn underlying EEG characteristics.Earlier spectrogram experiments were inferior to previous baselines; reformulating spectral-envelope filters yielded some improvement but remained below the baseline.Source: §2.3 · p. 11

2.3.1 Building blocks of the fully convolutional architecture

  • The fully convolutional network uses convolutional layers for both feature extraction and classification, without fully connected layers.Global average pooling and softmax produce probabilistic outputs from the final feature maps.Source: §2.3.1 · p. 12

2.3.2 Convolutions as data-driven feature extraction

  • Convolutional layers preserve temporal ordering along the feature-map width until global pooling, allowing the network to learn representations from the signal.Each layer maps its inputs to feature maps using convolution, bias, and a ReLU nonlinearity.Source: §2.3.2 · p. 12

2.3.3 Parameters

  • Varying feature-extraction block count and pool stride controls network complexity and the final convolutional layer’s receptive field.The largest receptive field used in these experiments was 256 EEG samples, or 8 seconds at 32 Hz.Source: Figure 3 · p. 12, §2.3.3 · p. 13
  • The final convolutional layer acts as the classifier, and its receptive field determines how much input EEG each element sees before global average pooling.Deeper layers widen the receptive field; adding convolutional layers or down-sampling with pooling can increase it.Source: §2.3.3 · pp. 13–14

2.4 1D FCN architecture: deep learning baseline

The 1D FCN processes one EEG channel as a temporal waveform and combines feature extraction with classification in one end-to-end routine. Its training requires strong labels, limiting the data available for training.

2.4 1D FCN architecture: deep learning baseline

  • Each input is an 8-second EEG window sampled at 32 Hz, represented as a 256×1 vector.The input is represented as a 1D temporal waveform.Source: §2.4 · p. 14
  • The 1D FCN combines feature extraction and classification in an end-to-end optimization routine while operating on a single EEG channel.It shares the SVM-based system’s pre- and post-processing steps.Source: §2.4 · p. 14
  • Strong-label training means that only a small subset of the available data can be used to train the 1D FCN.The system requires strong labels for training.Source: §2.4 · p. 14

2.5 2D FCN architecture: levering more data by using weak labels

The 2D FCN takes multiple EEG channels and detects seizure activity across them without requiring channel-specific seizure labels. This design allows weakly labelled data to be used for training.

2.5 2D FCN architecture: levering more data by using weak labels

  • The 2D FCN takes multiple EEG channels as input and detects whether a seizure occurs in any channel.Its input can contain an arbitrary number of channels; each channel is represented by an 8-second window sampled at 32 Hz.Source: §2.5 · p. 14
  • A final maximum-pooling layer takes the maximum across feature maps from all EEG channels, incorporating a 1D FCN post-processing step into end-to-end learning.The architecture also uses the same layers as the 1D FCN, with one-dimensional filters.Source: §2.5 · pp. 14–15
  • The 2D FCN is designed to distinguish seizure from non-seizure activity across channels using weakly labelled data, without channel-specific seizure labels.The shared convolutional filters use samples from all signals, and the architecture is designed to avoid learning baby-specific seizure-location dependencies.Source: §2.5 · p. 15
  • Using weak labels allows the entire Cork dataset to be used for training instead of only the subset with channel-specific labels.The authors hypothesize that over 40 times more EEG samples may compensate for noisier labels.Source: §2.5 · p. 15

2.6 Experimental procedure

The experiments use patient-independent leave-one-out testing and a separate Helsinki dataset test. They report AUC and AUC90 to assess classification performance, including performance at high specificity.

2.6.1 Leave one out experiment

  • Leave-one-out testing trains on all babies except one and tests on the unseen baby, repeating until each baby has been held out.For the 2D FCN, the remaining 17 training babies are split into 3 validation babies and 14 training babies; three trained networks’ probabilities are averaged for testing.Source: §2.6.1 · pp. 15–16
  • Strong-label requirements restrict 1D FCN training to a small portion of the data, leaving the rest of the training babies’ EEG available for validation and early stopping.Training stops when the validation score fails to improve for more than 25 iterations.Source: §2.6.1 · p. 15

2.6.2 Public dataset and algorithm

  • Cork-trained 1D and 2D networks are tested on the unseen Helsinki dataset and compared with a Helsinki-trained SVM; the Helsinki SVM is also tested on Cork.The Helsinki evaluation is a once-off test of generalization to a publicly available dataset.Source: §2.6.2 · p. 16

2.6.3 Metrics

  • AUC is used as the primary performance measure, enabling comparison with baselines reported using that metric.AUC can be computed across concatenated recordings, including babies without annotated seizures.Source: §2.6.3 · pp. 16–17
  • AUC90 measures area under the curve for specificities over 90%, focusing on seizure epochs detected at high specificity.The paper describes it as a more clinically relevant measure because high specificity is needed for an acceptable number of false detections per hour.Source: §2.6.3 · p. 16

3. Results

On Cork leave-one-out evaluation, deep learning achieved a 56% relative improvement over the SVM baseline; the 2D FCN trained on weakly labelled Cork data also performed best on Cork and unseen Helsinki. The results include per-baby comparisons and an architectural sweep across feature-extraction depth and pool stride.

3. Results

  • In nearly every Cork baby, the 1D FCN improved AUC, with the average across all AUC scores rising by over 1% absolute.Source: Figure 6 · p. 17, §3 · pp. 17–18
  • Figure 7 compares AUC across 1–5 feature-extraction blocks and pool strides 1–3; each point averages three full leave-one-out experiments, with standard-deviation error bars.Source: Figure 7 · p. 18, §3 · pp. 18–19
  • Figure 7 test scores include post-processing, whereas validation scores omit it; the caption attributes the higher test values to this difference.Source: Figure 7 · p. 18, §3 · p. 19
  • 56% relative performance improvement over the Cork SVM baseline was achieved: (98.5-96.6)/(100-96.6).Source: Table 2 · p. 19, §3 · p. 20
  • The 2D FCN trained with weakly labelled Cork data had the best performance on both Cork leave-one-out and the unseen Helsinki test dataset.Source: Table 3 · p. 19, §3 · p. 20

4. Discussion

The discussion examines the FCNs’ performance relative to SVM baselines across datasets, and describes factors associated with their advantages, including raw-EEG processing, weak-label training at scale, and receptive-field width. It also covers inspection of temporal outputs, validation-informed architecture selection, and processing speed.

4.1 Comparison with the SVM baseline on Cork dataset

  • The 1D FCN outperformed the Cork SVM in the LOO experiment, improving average performance by over 1% absolute under a shared-data and shared-protocol comparison.Both systems used the same strongly labelled training data, single-channel inputs, pre- and post-processing, and LOO testing on the same dataset.Source: §4.1 · p. 20
  • The FCN learns task-appropriate features from raw EEG and jointly optimizes feature extraction and classification, unlike the SVM’s hand-crafted feature approach.Using the temporal signal avoids losing information through an intermediate representation; the FCN’s feature selection is part of backpropagation.Source: §4.1 · pp. 20–21
  • The FCN uses fewer parameters than a CNN with fully connected layers, can process inputs of any length, and preserves temporal relationships between input and output.The discussion notes that variable-length input could facilitate integrating post-processing into a single optimization routine, though this was not explored in the study.Source: §4.1 · p. 21
  • The FCN’s final convolutional feature maps can be visualized as a heatmap to identify EEG patterns that contribute strongly to seizure decisions.A softmax maps corresponding values in the final two feature maps to values between zero and one; higher values indicate stronger contributions.Source: §4.1 · p. 22, Figure 8 · p. 22

4.2 Receptive field

  • Most AUC improvement came from wider receptive fields; more parameters or greater depth alone did not guarantee higher performance.At very low receptive fields, the five-block setup with pool stride one had many parameters but lower performance than other depths investigated.Source: §4.2 · pp. 22–23
  • Validation trends reflected test results, supporting validation-based architecture selection, although the architecture search was not exhaustive.The authors note that other building-block designs or combinations of pool strides could yield larger receptive fields.Source: §4.2 · p. 23
  • The study varies network depth, width, and input resolution using compounding scaling and applies parameter-balancing principles to physiological signal-processing networks.Source: §4.2 · p. 24

4.3 More data with weak labels vs. less data with strong labels

  • The 2D FCN uses weak labels indicating seizure presence in at least one channel rather than channel-specific labels, whose annotation is time-consuming and often limited to small dataset subsets.Source: §4.3 · p. 24
  • The 2D FCN’s superior performance followed training on over 40 times more EEG samples: over 2.5 million 8-second examples (~5B samples) versus over half a million (~128M) for the 1D FCN.Source: §4.3 · p. 24

4.4 Comparison with a publicly available algorithm

  • The 2D FCN performed best in matched and mismatched tests, with the smallest performance drop and higher Helsinki AUC than the Helsinki SVM achieved on Cork.Source: §4.4 · p. 25
  • The Helsinki dataset’s short EEG segments are not representative of typical NICU recordings, and longer public recordings would better reflect clinical use.Source: §4.4 · pp. 25–26
  • The Helsinki SVM took over 4.5 hours to process one hour of EEG, while both FCNs completed the task in less than one minute on the same CPU-based PC.Source: §4.4 · p. 26

5. Conclusions

The conclusions report stronger performance for the proposed FCNs than feature-based alternatives, including on a public Helsinki test set. They also highlight raw-EEG end-to-end learning, weak labels, and receptive-field width as key architectural findings.

5. Conclusions

  • The approach uses a fully convolutional network on raw EEG and jointly optimises feature extraction and classification.Source: §5 · p. 26
  • Given the same initial conditions, the 1D FCN learned from raw annotated data and outperformed an SVM system using hand-crafted features.Source: §5 · p. 26
  • Receptive-field width in the final convolutional filters influenced performance more than network depth or the overall number of tuneable parameters.Source: §5 · p. 26
  • The 2D FCN exploits larger amounts of weakly labelled data, with the stated aim of reducing reliance on laborious precise per-channel labelling.Source: §5 · pp. 26–27
  • On the publicly available Helsinki dataset, the 2D FCN showed superior performance to existing published alternatives.Source: §5 · p. 27
  • The deep-learning-only algorithm outperformed algorithms relying on complex feature sets without the computational load typically associated with intensive feature calculations.Source: §5 · p. 27

Related papers

Loading 2105.13854v1…