SeizureTransformer: Scaling U-Net with Transformer for Simultaneous Time-Step Level Seizure Detection from Long EEG Recordings

We design model architecture based on the U-Net to do end-to-end learning from raw waveforms for time-step-level classification to achieve simultaneous seizure detection. Our model consists of three primary modules: an encoder, a scaling embedding component, and a decoder, as shown in Fig. 1. Taking the continuous long-term EEG signals from the epilepsy monitoring unit, the encoder extracts features by recognizing patterns through one-dimensional convolution layers. The feature vectors are further embedded by a ResCNN stack and a Transformer encoder stack with a global attention mechanism to generate high-level representations that capture rich temporal dependencies. The streamlined decoder then converts these representations into a sequence of probability, indicating the presence or absence of seizures at every time step. Residual connections between each encoder layer and decoder layer are used to ease the gradient flow and to avoid degradation problems in the deep neural network. More details about network architecture selection are provided in the methodology section.

We combine two datasets by concatenating segmented one-minute-long time series windows together, i.e., $60\times 256=15360$ time steps per window. A $75\%$ overlap ratio between two consecutive windows was set as a hyperparameter during the segmentation process to augment training examples. To improve the model’s ability to distinguish seizure signals from background noise, we statistically categorize training windows into three classes: no-seizure, full-seizure, and partial-seizure, and uniformly sample a certain number of windows from each class to create a balanced dataset. Specifically, our training dataset is constructed as follows:

where $\mathcal{D}_{ps}$ contains all partial-seizure windows, $\mathcal{D}_{fs}^{*}$ and $\mathcal{D}_{ns}^{*}$ is a randomly selected subset of full-seizure and no-seizure window with $|\mathcal{D}_{fs}^{*}|=0.3\times|\mathcal{D}_{ps}|$ and $|\mathcal{D}_{ns}^{*}|=2.5\times|\mathcal{D}_{ps}|$.

We used TUSZ’s predefined test set, consisting of 42.7 hours of waveforms from 43 subjects with 469 seizure activities, to evaluate the detection performance of SeizureTransformer with other traditional and deep-learning algorithms. The test set of TUSZ is a list of blind EEG signals that are completely separated from its training set and validation set, which ensures the generalization of model performance.

We quantify the model’s performance using the area under the receiver operating characteristics(AUROC). For each continuous EEG recording, the ROC curve plots the true and false positive rates across all possible decision thresholds, and the AUC represents the area under the ROC curve, which summarizes the model’s performance. We compare our model’s performance using the same evaluation metric under the TUSZ’s predefined test set with other seizure detection models, namely, Zhu-Transformer[6], EEGWaveNet[8], and DCRNN[7], to demonstrate the effectiveness of our proposed approach. Models used here for the comparison are pre-trained models based on different training sets. All of these pre-trained models are implemented by [21] and are publicly available. As shown in Figure 3, our model demonstrated the highest performance, with a mean AUROC of $0.876$ and a distribution tightly concentrated toward higher values.

The 2025 Seizure Detection Challenge ¹¹1competition website and leaderboard is available in: https://epilepsybenchmarks.com/challenge/, organized as part of the International Conference on Artificial Intelligence in Epilepsy and Other Neurological Disorders, provides a completely blind private dataset consisting of continuous EEG recordings for evaluation, which makes it an ideal place to test the performance and generalization of our model fairly. The test dataset was collected at the EMU of the Filadelfia Danish Epilepsy Center in Dianalund from January 2018 to December 2020 with the NicoleteOne^TM v44 amplifier. The dataset contains 4360 hours of EEG recordings from 65 subjects with various ages, where for each subject, at least one seizure during the hospital stay with a visually identifiable electrographic correlate to the seizures recorded on the video. The ground truth labels were annotated by three board-certified neurophysiologists with expertise in long-term video-EEG monitoring. The F1-score, sensitivity, precision, and false positive per day were used as the primary ranking criterion to align with real-world requirements. The event-based scoring evaluates annotations at the event level by assessing the degree of overlap between predicted and reference events.

As shown in Table I, our model largely outperforms the other algorithms in terms of F1-score. It is noteworthy that we set the picking threshold to be $80\%$ in the competition, which leads to a relative low sensitivity but comes with the best precision and False Positive rate. Van Gogh Detector and Zhu-Transformer are window-level classification models that also take advantage of both convolutional and transformer encoder units; however, their performance did not reach that of SeizureTransformer. This points to the beneficial effects of time-step-level end-to-end learning. Similarly, SeizUnet, like our model, is a time-step-level classification algorithm using U-Net; but different to SeizureTransformer, it chooses to add LSTM layers, instead of transformer encoders, after the U-Net decoder, instead of embedding into the U-Net, and turns out to be not as good as our results.

Window-level classification models assign predictions individually to each segmented window. Mapping window labels to the final annotation output followed by the SCORE compliant [10] that contains the start time and duration time of a seizure requires the model to segment windows with a great overlap ratio to ensure the start and stop time’s precision. This led to tremendous redundant computing and complicated mapping procedures. On the other hand, the time-step-level classification models do not require such post-processing steps as their predictions can directly indicate the onset time and activity duration. This approach inherently mitigates the redundant computations associated with overlapping windows and significantly simplifies the annotation pipeline, which makes this method align more closely with the practical clinical requirement for efficient automated seizure detection.

We further show our model’s efficiency by comparing the inference time with other models using TUSZ’s testing set in Table II. Our model demonstrate the lowest running time with the ability to handle a one-hour-long recording in 3.98 seconds.

The better performance of the proposed method for seizure detection could be due to several factors. Here, we show each model component’s necessity by testing multiple partial models after removing certain components. As shown in Figure 4, vanilla U-Net has an underwhelming performance with a low AUROC mean. Solely adding a ResCNN stack or a transformer stack will marginally improve the model performance but also lead to a bigger variance with some extreme false cases. By contrast, integrating both the ResCNN and Transformer stacks produces not only higher mean AUROC but also reduced variance, indicating that these components complement each other effectively. These results underscore the importance of each proposed element in achieving robust and accurate seizure detection.

The competition leaderboard shows a relatively low F1-score across every algorithm compared to the results shown in previously published reviews[22, 23] and the self-reported performance of computing algorithms. To comprehensively understand the model’s performance, we test our model with several published algorithms, namely, EventNet [24], Zhu-Transformer [6], DCRNN [7], EEGWaveNet [8], and the Gotman algorithm [25], in the TUSZ’s predefined testing set using the same evaluation metrics(F1-score, sensitivity, and precision) implemented by the challenge organizers [21]. The testing tools provide both sample-level and event-level evaluation. As shown in Table III, while our model keeps the state-of-the-art performance, all model achieved better F1-scores. Such result difference might be due to the distribution shift between datasets. As described by the organizer, the private evaluation dataset include recordings from various ages, and the data was collected by portable EEG amplifiers, allowing patients to move freely within the building, which will likely lead to unique attributes in the recording that depart from the training set.

U-Net[18] architecture was first proposed in the field of CV for image segmentation tasks. Considering the temporal continuity of time series data, such networks have been widely deployed in various scientific signal processing applications, such as seismic phase detection[16], sleep-staging classification [12, 26], denoising heart sound signals [27], and Seizure detection[28].

There are some works exploring combining U-Net with Transformer together for other fields. For example, in a medical image segmentation task, [29] used self and cross-attention with U-Net; [30] incorporated hierarchical Swin Transformer into U-Net to extract both coarse and fine-grained feature representations. In seismic analysis, [31] proposed a deep neural network that can be regarded as a U-Net with global and self-attention but without a residual connection. However, in the signal processing area, to the best of our knowledge, there is no existing work to scale U-Net using transformer blocks. The closest work to this paper [28], where multiple attention-gated U-Net are used and a following LSTM network is implemented to fusion results.

For a continuous EEG waveform, before segmenting it to uniform windows as training examples, we resample all data to a common, i.e., 256, sampling rate using the Fourier method [32], to fix the time resolution for the convolutions in the model to be meaningful across subjects, and implement a Gaussian normalization to each channel, calculated by

The generated dataset, after slicing, is denoted as $\mathcal{D}=(\mathcal{X},\mathcal{Y})=\{(x_{i},y_{i})\mid i=1,\dots,N\}$, where $N$ represents the number of training samples. Each input window $x_{i}\in\mathbb{R}^{T\times d}$ represents a multivariate time series with $T=256\times 60=15{,}360$ time steps and $d=19$ channels. The corresponding time-step-level label $y_{i}\in\{0,1\}^{T}$ is a binary, box-shaped ground truth signal indicating the presence of seizure activity at each time step.

We then implement a transformer encoder stack [33] to scale the model and to capture long-range dependencies across the tokenized signal. Specifically, the sine and cosine functions of different frequuencies are used to be positional encodings,

which can then be summed with the input embedding. The refined representation, denoted as $Z$, will then be projected into equally-shaped query, key, and value spaces,

and processed with the use of the global-attention mechanism,

The attention output is combined with tokens with a residual connection and layer normalization and a subsequent feed-forward network to transform the output with another residual addition.

Such hierarchical processing scales the model and integrates both local features and global context, enabling the model to learn complex temporal dependencies.