Source-linked AI summary
ChronoNet: A Deep Recurrent Neural Network for Abnormal EEG Identification
Subhrajit Roy, Isabell Kiral-Kornek, Stefan Harrer
TL;DR
Manual EEG interpretation is slow and resource-intensive, motivating automated abnormal-versus-normal classification. The paper proposes ChronoNet, a raw-time-series RNN architecture combining multiscale convolutions and dense GRU connections, and reports improved benchmark performance with speech-command transfer.
Problem
Manual EEG interpretation requires experts and substantial time for large recordings, motivating automation of abnormal-versus-normal classification.
Method
ChronoNet processes raw EEG time series with stacked Conv1D layers using varying filter sizes followed by densely connected GRU layers.
Results
1.17% better accuracy than recently published state-of-the-art performance is reported on the TUH Abnormal EEG Corpus, setting a new benchmark.
Takeaways & Limitations
ChronoNet successfully classifies EEG abnormalities and speech commands, supporting its reported applicability to time-series analysis beyond EEG.
Takeaways & Limitations
The architecture can suffer degradation during deep-network training and may yield higher training errors on simpler problems that do not require its full complexity.
Abstract
from arXiv · showhide
Brain-related disorders such as epilepsy can be diagnosed by analyzing electroencephalograms (EEG). However, manual analysis of EEG data requires highly trained clinicians, and is a procedure that is known to have relatively low inter-rater agreement (IRA). Moreover, the volume of the data and the rate at which new data becomes available make manual interpretation a time-consuming, resource-hungry, and expensive process. In contrast, automated analysis of EEG data offers the potential to improve the quality of patient care by shortening the time to diagnosis and reducing manual error. In this paper, we focus on one of the first steps in interpreting an EEG session - identifying whether the brain activity is abnormal or normal. To solve this task, we propose a novel recurrent neural network (RNN) architecture termed ChronoNet which is inspired by recent developments from the field of image classification and designed to work efficiently with EEG data. ChronoNet is formed by stacking multiple 1D convolution layers followed by deep gated recurrent unit (GRU) layers where each 1D convolution layer uses multiple filters of exponentially varying lengths and the stacked GRU layers are densely connected in a feed-forward manner. We used the recently released TUH Abnormal EEG Corpus dataset for evaluating the performance of ChronoNet. Unlike previous studies using this dataset, ChronoNet directly takes time-series EEG as input and learns meaningful representations of brain activity patterns. ChronoNet outperforms the previously reported best results by 7.79% thereby setting a new benchmark for this dataset. Furthermore, we demonstrate the domain-independent nature of ChronoNet by successfully applying it to classify speech commands.
1 Introduction
EEG is a practical diagnostic signal, but interpreting long recordings manually is costly and slow. The paper targets automated abnormal-versus-normal classification with ChronoNet and evaluates it on TUH Abnormal EEG Corpus.
- Clinical motivation: EEG measures brain activity noninvasively and is used to diagnose neurological conditions including epilepsy.Its excellent temporal resolution and low cost support clinical use despite lower spatial resolution than MRI and CT.
- Clinical motivation: 2–weeks of treatment delay can result from manually interpreting the large EEG datasets produced by long-term monitoring.The passage states that limited expert availability and high data volume make interpretation time-consuming.
- Task: The first interpretation step is deciding whether recorded activity appears abnormal or normal, influencing medication and further investigation.Clinicians consider recording patterns and the patient’s state of consciousness when making this distinction.
- Approach: ChronoNet uses raw EEG time series and recurrent neural networks to automate this first interpretation step on the TUH Abnormal EEG Corpus.This differs from earlier work using hand-engineered features or other machine-learning approaches without RNNs for this task.
- Results: 7.77% overall improvement over earlier reported results is achieved by combining raw time series, RNNs, and the ChronoNet architecture.The reported gains are 3.51% from the raw-time-series RNN combination and an additional 4.26% from ChronoNet.
- Contributions: ChronoNet combines inception layers with exponentially varying 1D convolution filters and densely connected recurrent layers.The paper also reports state-of-the-art EEG classification and successful application to speech commands.
2 Background and Theory
The background develops recurrent and convolutional components for modeling EEG time series across temporal scales. It motivates GRUs, inception filters, and dense connections as building blocks for ChronoNet.
- EEG time series: Raw EEG time series contain patterns and periodicities at multiple time scales, motivating recurrent models for temporal signal classification.The section presents these principles as the basis for understanding ChronoNet.
- Recurrent neural networks: RNNs process variable-length sequences by updating a hidden state from the current input and previous hidden state.The recurrent state carries information across time steps.
- Recurrent neural networks: Classical RNN gradients can grow or decay exponentially over long sequences, making long-term dependencies difficult to learn.This motivates more sophisticated gated recurrent units.
- Gated recurrent units: GRUs use update and reset gates to interpolate a candidate hidden state with the previous hidden state.The supplied equations specify the GRU state update and gate computations.
- Inception modules: Inception modules apply filters of varied sizes to capture features at different abstraction scales and aggregate the resulting information.The conventional module uses 1 × 1, 3 × 3, and 5 × 5 filters plus a max-pooling path.
- Inception modules: Using more convolutional filters increases parameters and computational intensity, motivating 1 × 1 bottleneck filters for dimensionality reduction.The bottleneck mechanism is presented as a response to the cost of naive multi-filter convolution.
- Dense connections: DenseNet connects every layer to all preceding layers in a feed-forward fashion, producing L(L + 1)/2 direct connections for L layers.These shortcut paths are described as mitigating vanishing or exploding gradients in deep networks.
3 Methods
The methods develop ChronoNet by combining multiscale 1D convolutions with densely connected GRU layers for temporal EEG processing. These components address computational cost, time-scale flexibility, and degradation in deep recurrent networks.
- C-RNN: Conv1D layers reduce EEG sequence length and extract local information before GRU layers capture short- and long-term dependencies.This makes the computationally expensive recurrent stage more tractable while beginning temporal representation learning.
- C-RNN: C-RNN processes temporal inputs with Conv1D layers followed by stacked GRU layers, using layer specifications for filter lengths, filter counts, strides, and GRU filters.The paper presents this as the base convolutional recurrent architecture.
- IC-RNN: Multiple exponentially varying filters in each Conv1D layer let IC-RNN extract information across multiple time-scales without hand-picking one filter size.The authors report better performance for exponentially varying lengths on their dataset and motivate them by the wider time-scale range of time series.
- C-DRNN: Dense feed-forward connections among GRU layers form C-DRNN and are intended to mitigate degradation by allowing lower-complexity paths through the recurrent stack.Each GRU layer receives outputs from every other GRU layer in a feed-forward manner.
- ChronoNet: ChronoNet combines C-RNN’s convolutional-recurrent structure, IC-RNN’s exponentially varying filters, and C-DRNN’s densely connected GRUs.The resulting architecture stacks multiple Conv1D layers before multiple densely connected GRU layers.
4 Experiments
Experiments evaluated ChronoNet and related recurrent architectures on the TUH Abnormal EEG Corpus, including expanded training windows and comparisons with prior systems. ChronoNet achieved the strongest reported EEG accuracy and also performed well on speech-command classification.
- Data Selection: The TUH Abnormal EEG Corpus contains clinically abnormal or normal EEG sessions selected from the TUH EEG Corpus.The dataset was formed as a demographically balanced subset through manual review.
- Data Preparation: Up to 11 minutes of training data could be used without performance degradation, producing an 11-fold increase in training data.Testing continued to use the first minute for comparison with prior work.
- Results: ChronoNet surpassed the original CNN-MLP result by 7.77% and the previously reported state-of-the-art result by 1.17% in accuracy.It also achieved the best training and testing accuracy among the four recurrent architectures evaluated.
- Results: With only the first minute of training data, ChronoNet achieved 85.27% accuracy, 6.47% better than the original study.The number of EEG records used for training matched that of other works on the dataset.
- Results: Replacing exponentially varying filters with linearly varying filters reduced testing accuracy to 85.12% or 85.92%, depending on filter length.The shorter-filter variant achieved 89.15% training accuracy, while the longer-filter variant achieved 91.25%.
- Beyond EEG Identification: On the 30-class Speech Commands Dataset, ChronoNet achieved 92.84% testing accuracy using the same architecture.This experiment extended evaluation beyond EEG classification.
5 Conclusion
The paper introduces ChronoNet to automate abnormal-versus-normal EEG identification, addressing the expense and delay of manual interpretation. ChronoNet sets a new benchmark on the evaluated dataset and successfully classifies speech, supporting its use beyond EEG.
- Conclusion: ChronoNet targets the first diagnostic distinction between abnormal and normal EEG activity, which is expensive and time-consuming to assess manually.The paper presents it as a flexible and adaptable architecture for EEG time-series analysis.
- Conclusion: ChronoNet outperforms the best previously reported accuracy on the evaluated dataset by 1.17%, setting a new benchmark.The paper also reports successful speech classification to demonstrate broader time-series applicability.