SinTransNet: an EEG-based deep learning framework for infantile epileptic spasms syndrome detection Authors: Junyuan Feng, Zhenzhen Liu, Linlin Shen, Xiaoling Luo, Yan Chen, Lin Li, Tian Zhang Published: 2026 ABSTRACT Infantile Epileptic Spasms Syndrome (IESS) represents a severe form of developmental epileptic encephalopathy in infancy, characterized by clusters of spasms and hypsarrhythmia patterns on electroencephalogram (EEG), which often lead to long-term neurodevelopmental impairments if not diagnosed promptly. The inherent non-stationarity and polymorphic complexity of EEG signals complicate interpretation, resulting in time-consuming and error-prone diagnostics that hinder timely therapeutic interventions. To address these challenges, we propose SinTransNet, an innovative EEG-based deep learning framework that combines multi-band signal decomposition, adaptive sinusoidal convolutions, and Transformer-based attention mechanism. This architecture decomposes EEG into five key frequency bands (δ, θ, α, β, γ) to isolate oscillatory features such as spike-and-wave complexes, employs sinusoidal convolutions for frequency-adaptive feature extraction, and utilizes Transformer attention to capture inter-band correlations and long-range dependencies essential for accurate IESS detection. Evaluated on a proprietary dataset comprising 129 EEG recordings with 1,941 epileptic spasm events, SinTransNet demonstrates superior performance with average accuracy of 85.69%, sensitivity of 80.55%, and specificity of 90.76%. By providing an automated, efficient tool for IESS identification, SinTransNet holds promise for enhancing clinical workflows and supporting early interventions in pediatric neurology. Introduction Infantile epileptic spasms syndrome (IESS), a severe developmental and epileptic encephalopathy, presents with epileptic spasms in infants aged 1 to 24 months, typically peaking between 3 and 12 months. The most frequently observed ictal behaviors in infantile spasms include symmetric flexion, extension, or a combination of both [1, 2]. During ictal events, electroencephalogram (EEG) recordings typically exhibit fast activity resembling spindles, high-voltage spikes, spike-and-wave complexes, slow waves, and attenuation [1, 2]. The interictal EEG is characterized by hypsarrhythmia, an abnormal pattern comprising disorganized, high-voltage slow waves and spikes [3]. The subtle, brief spasms and complex EEG patterns, including transient spike waves and cross-band interactions, pose unique challenges for IESS detection, necessitating early diagnosis to mitigate severe developmental consequences. Traditionally, IESS detection relies on labor-intensive visual EEG inspection, which is time-consuming, costly, and prone to human error. The implementation of automated IESS detection system holds the potential to significantly enhance the diagnostic process for clinicians, thereby improving diagnostic efficiency and accuracy. Early studies on EEG-based epilepsy and IESS detection primarily relied on traditional machine learning (ML) frameworks, where handcrafted features were extracted from EEG recordings and subsequently classified using conventional classifiers such as Support Vector Machines (SVM), Random Forests (RF), and K-Nearest Neighbors (KNN) [4–6]. Typical feature extraction strategies included time-domain statistical descriptors, spectral power analysis, entropy measures, and wavelet-based time-frequency representations. For example, Subasi et al. combined Discrete Wavelet Transform (DWT) with SVM optimized using Genetic Algorithms and Particle Swarm Optimization for seizure classification [7]. Aayesha et al. introduced Fuzzy Rough Nearest Neighbors to address EEG nonlinearity [8]. Omidvar et al. further employed multi-level Db4 wavelet decomposition for multi-scale EEG analysis [9], while Pattnaik et al. utilized Tunable-Q Wavelet Transform to improve time-frequency resolution [10]. In addition, Melek demonstrated that frequency-domain representations outperform purely time-domain features in KNN-based epilepsy classification, highlighting the importance of rhythmic EEG activity analysis [11]. Although these approaches achieved promising results in controlled settings, they heavily depend on manual feature engineering and expert prior knowledge. Consequently, they often struggle to capture the highly nonlinear, non-stationary, and cross-band coupling characteristics of IESS EEG signals, thereby limiting their generalization capability and clinical robustness. The advent of deep learning has catalyzed progress in epilepsy detection, with Convolutional Neural Networks (CNNs) automating feature extraction from raw EEG data to discern intricate patterns [12, 13]. Wei et al., for example, devised a multi-view CNN incorporating Feature Pyramid Networks (FPN) for interictal epileptiform discharge detection [14], and Pan et al. fused raw EEG with FFT, STFT, and DWT in a hybrid time-frequency CNN for multi-domain integration [15]. Shanmugam et al. augmented CNN with LSTM for temporal modeling [16], while Liu et al.’s CosCNN employed cosine-based kernels to mitigate overfitting in lightweight designs [17]. Nonetheless, CNNs’ constrained local receptive fields impede their grasp of expansive seizure dynamics, limiting efficacy in capturing the global context vital for IESS. Transformer architectures have recently surged as efficacious alternatives for EEG tasks, harnessing self-attention to model long-range dependencies and surmount CNN limitations [18–21]. He et al. proposed a sparse Transformer encoder to alleviate computational demands [22], Zhao et al. introduced CwA-T with channel-wise compression for efficient feature handling [23], and Yan et al.’s DARNet fused spatio-temporal convolutions and dual attention for bolstered dependency capture [24]. Yet, these models frequently overlook inter-band oscillatory linkages in brain rhythms, compromising performance in nuanced seizure scenarios for epileptic spasms. Although general epilepsy research flourishes, IESS-targeted investigations are sparse. Das et al. applied SVM to differentiate IESS from akin disorders [25], Fan et al. automated the BASED scoring via multi-layer CNN [26], and Zheng et al. constructed PLV-based networks from five EEG sub-bands, training ResNet18 for detection [27]. These tailored efforts, while promising, often neglect seamless fusion of frequency-specific oscillations with global modeling, yielding suboptimal accuracy and robustness in clinical IESS contexts. To address these challenges, we introduce SinTransNet, a novel deep learning framework designed to meet the unique demands of IESS detection in pediatric EEG analysis. Tailored to the neurophysiological complexity of IESS, SinTransNet integrates multi-band decomposition, sinusoidal convolution, Transformer-based attention, and a lightweight classifier, harmonizing local oscillatory signatures with global seizure dynamics. EEG signals are decomposed into five sub-band rhythms for capturing the diverse oscillatory patterns inherent to spasms. Sinusoidal convolutions, parameterized by amplitude, frequency, phase, and bias, adaptively extract band-specific features, addressing the non-stationary nature of EEG signals. These features are then processed as tokens by a Transformer-based attention mechanism, which models cross-band interactions and long-range dependencies critical for delineating IESS-specific patterns. Finally, a lightweight classifier generates the diagnostic output. Our key contributions are:Sinusoidal Convolution: A novel convolutional approach that dynamically captures frequency-specific EEG features, such as spikes and slow waves, enhancing robustness to IESS’s non-stationary signals.SinTransNet Framework: A unified framework that synergizes multi-band decomposition, sinusoidal convolution, and Transformer attention, bridging local and global representations to address IESS diagnostic challenges.Superior Performance: Validated on a proprietary IESS dataset characterized by high-density spasm events and high-precision expert annotations, SinTransNet achieves a robust accuracy of 85.69%, sensitivity of 80.55%, and specificity of 90.76%, outperforming state-of-the-art methods and supporting timely clinical interventions. Sinusoidal Convolution: A novel convolutional approach that dynamically captures frequency-specific EEG features, such as spikes and slow waves, enhancing robustness to IESS’s non-stationary signals. Sinusoidal Convolution: A novel convolutional approach that dynamically captures frequency-specific EEG features, such as spikes and slow waves, enhancing robustness to IESS’s non-stationary signals. SinTransNet Framework: A unified framework that synergizes multi-band decomposition, sinusoidal convolution, and Transformer attention, bridging local and global representations to address IESS diagnostic challenges. SinTransNet Framework: A unified framework that synergizes multi-band decomposition, sinusoidal convolution, and Transformer attention, bridging local and global representations to address IESS diagnostic challenges. Superior Performance: Validated on a proprietary IESS dataset characterized by high-density spasm events and high-precision expert annotations, SinTransNet achieves a robust accuracy of 85.69%, sensitivity of 80.55%, and specificity of 90.76%, outperforming state-of-the-art methods and supporting timely clinical interventions. Superior Performance: Validated on a proprietary IESS dataset characterized by high-density spasm events and high-precision expert annotations, SinTransNet achieves a robust accuracy of 85.69%, sensitivity of 80.55%, and specificity of 90.76%, outperforming state-of-the-art methods and supporting timely clinical interventions. The remainder of this paper is organized as follows. Section 2 describes the SinTransNet architecture. Section 3 provides a detailed description of the proprietary IESS dataset. Sections 4 and 5 present the experimental results and discussion, respectively. Section 6 concludes the paper. Methodology Overall architecture As illustrated in Fig. 1, the proposed SinTransNet is a hierarchical deep learning architecture. The framework encompasses five interconnected modules: signal preprocessing, multi-band decomposition, sinusoidal convolution, Transformer encoding, and classification, forming a seamless pipeline for robust feature extraction and decision-making.Fig. 1The framework of SinTransNet, including a signal preprocessing module, a multi-band decomposition module, a sinusoidal convolution module, a Transformer encoding module, and a classification module The framework of SinTransNet, including a signal preprocessing module, a multi-band decomposition module, a sinusoidal convolution module, a Transformer encoding module, and a classification module The preprocessing module filters, re-references, and standardizes raw EEG inputs to ensure high-quality data. Subsequently, the multi-band decomposition module partitions the signals into five brain wave sub-bands, aligning with the frequency-specific manifestations of IESS. The sinusoidal convolution module then processes each sub-band adaptively, extracting nuanced oscillatory characteristics that reflect the dynamic nature of epileptic spasms. These localized features are tokenized and fed into the Transformer encoding module, which employs multi-head self-attention to model cross-band fusions and long-range dependencies, yielding comprehensive global representations. Finally, the classification module delivers a binary output, enabling reliable IESS identification suitable for clinical integration. Signal preprocessing module In this section, some steps are used to preprocess the raw EEG data to mitigate noise, enhance spatial patterns, and standardize amplitudes. First, a band-pass filtering in the range of 0.3 to 70 Hz is applied to remove low-frequency artifacts and high-frequency noise, while preserving the neurophysiological components relevant to IESS [28]. Next, the average reference montage reduces common-mode signals across channels, thereby enhancing the signal-to-noise ratio of localized brain activity. For a given channel’s filtered signal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_c(t)$$\end{document}, where c denotes the channel index and t represents time, the average-referenced signal is: 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{c,\mathrm{ref}}(t) = x_c(t) - \frac{1}{C} \sum_{c\prime=1}^C x_{c\prime}(t),$$\end{document} where C is the number of channels and the mean is calculated across all channels at each time point. Subsequently, the Z-score standardization normalizes the signal referenced by the average montage of each channel to eliminate the variability of the amplitude, which facilitates the convergence of the model. The normalized signal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{c,\mathrm{norm}}(t)$$\end{document} is computed as: 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{c,\mathrm{norm}}(t) = \frac{x_{c,\mathrm{ref}}(t) - \mu_c}{\sigma_c},$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu_c$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma_c$$\end{document} are the mean and standard deviation of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{c,\mathrm{ref}}(t)$$\end{document} on the temporal dimension. The output is a preprocessed EEG tensor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X \in \mathbb{R}^{B \times C \times L}$$\end{document}, where B is the batch size and L is the temporal length of the decision window. These steps ensure that the input signals are clean and standardized, providing a consistent foundation for subsequent extraction and classification of features. Multi-band decomposition module The human brain consists of multiple specialized regions that interact to produce distinct brain wave rhythms [29], reflecting activity in functional areas. We decompose EEG signals into five frequency sub-bands: δ (0.3–4 Hz), θ (4–8 Hz), α (8–13 Hz), β (13–30 Hz), and γ (30–70 Hz). This is achieved using finite impulse response (FIR) band-pass filters, which provide linear phase response and precise frequency selectivity without phase distortion, preserving temporal patterns in IESS-related signals. For each sub-band \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k \in \{1, \ldots, 5\}$$\end{document}, we design an FIR filter with M=101 coefficients using the windowed method with a Hamming window. This balances sharpness in the transition band and computational efficiency while minimizing ripples in passband and stopband. The coefficients \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h_k(m)$$\end{document} are generated via Scipy’s firwin function for the passband \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[f_{\mathrm{low},k}, f_{\mathrm{high},k}]$$\end{document} and sampling frequency \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_s$$\end{document}. The frequency response is: 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_k(e^{j\omega}) = \sum_{m=0}^{M-1} h_k(m) e^{-j\omega m},$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega = 2\pi f / f_s$$\end{document} is the normalized angular frequency. Based on this response, the filtering operation is performed as a convolution between the EEG signal X and filter coefficients. For each channel and sub-band k, the filtered signal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_k$$\end{document} is computed as: 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_k(n) = \sum_{m=0}^{M-1} h_k(m) X(n-m),$$\end{document} where n is the time index, and X(n-m) is the input signal shifted by m time samples. To maintain the output length, zero-padding is applied with (M-1)/2 time samples on each side. This is efficiently implemented using a 1D convolutional layer with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h_k(m)$$\end{document} as a fixed kernel, applied independently per channel. The module outputs a tensor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y \in \mathbb{R}^{B \times C \times L \times K}$$\end{document}, where each \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_k \in \mathbb{R}^{B \times C \times L}$$\end{document} isolates frequency-specific components while retaining temporal and spatial structure. Sinusoidal convolution module Traditional convolution kernels lack explicit frequency-dependent training parameters, making it difficult to adapt to the non-stationary oscillations of EEG. The sinusoidal convolution module extracts localized oscillatory features from each sub-band of the EEG signal, tailored to the neurophysiological characteristics of IESS. Our design incorporates four learnable parameters, amplitude, frequency, phase offset, and bias, for superior adaptability to dynamic EEG patterns in pediatric epilepsy. The module receives the multi-band input tensor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y \in \mathbb{R}^{B \times C \times L \times K}$$\end{document} from the decomposition stage. For each sub-band k, a specialized processing unit handles \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y_k \in \mathbb{R}^{B \times C \times L}$$\end{document} to yield a condensed feature set. At the heart of each convolutional layer lies a sinusoidal kernel, formulated for the i-th output channel and j-th input channel as: 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k_{i,j}(m) = A_{i,j} \sin \left( 2\pi \frac{\omega_{i,j}}{f_s} \bar{m} + \phi_{i,j} \right) + b_{i,j},$$\end{document} where m = 0, 1, …, M-1 denotes the kernel index with M=25, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{m} = m - \frac{M-1}{2}$$\end{document} ensures symmetry, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_s$$\end{document} represents the sampling frequency. The frequency \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega_{i,j}$$\end{document} is normalized by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_s$$\end{document} to align with the discrete signal domain. The trainable parameters include:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_{i,j}$$\end{document}: amplitude, modulating oscillation intensity to adapt to EEG amplitude variations;\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega_{i,j}$$\end{document}: frequency, tuning the period to target sub-band oscillatory components;\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\phi_{i,j}$$\end{document}: phase shift, fine-tuning alignment for capturing ephemeral events like spikes;\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b_{i,j}$$\end{document}: bias, offsetting baseline shifts in the signal. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_{i,j}$$\end{document}: amplitude, modulating oscillation intensity to adapt to EEG amplitude variations; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_{i,j}$$\end{document}: amplitude, modulating oscillation intensity to adapt to EEG amplitude variations; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega_{i,j}$$\end{document}: frequency, tuning the period to target sub-band oscillatory components; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega_{i,j}$$\end{document}: frequency, tuning the period to target sub-band oscillatory components; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\phi_{i,j}$$\end{document}: phase shift, fine-tuning alignment for capturing ephemeral events like spikes; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\phi_{i,j}$$\end{document}: phase shift, fine-tuning alignment for capturing ephemeral events like spikes; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b_{i,j}$$\end{document}: bias, offsetting baseline shifts in the signal. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b_{i,j}$$\end{document}: bias, offsetting baseline shifts in the signal. To confine \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega_{i,j}$$\end{document} within the sub-band bounds \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[f_{\mathrm{min},k}, f_{\mathrm{max},k}]$$\end{document} Hz, we employ a logarithmic parameterization: 6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega_{i,j} = \exp(\omega\prime_{i,j}), \quad \omega\prime_{i,j} \in [\ln(f_{\mathrm{min},k}), \ln(f_{\mathrm{max},k})],$$\end{document} with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega\prime_{i,j}$$\end{document} initialized uniformly and refined through training, promoting stable and bounded frequency adjustments. Featuring \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C_{\mathrm{out}}=128$$\end{document} output channels per layer, the convolution for the i-th channel is computed as: 7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_{k,i}(n) = \sum_{j=1}^{C} \sum_{m=0}^{M-1} k_{i,j}(m) Y_{k,j}(n-m),$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y_{k,j}(n)$$\end{document} is the j-th channel of the k-th sub-band at time n, generating 128 feature maps. Each sub-band module consists of two cascaded layers, each incorporating convolution, batch normalization (BN), LeakyReLU activation with a 0.3 negative slope, and max pooling (kernel size 2, stride 2). The first layer’s output is: 8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y_k^{(1)} = \mathrm{MaxPool}_2 \left( \mathrm{LeakyReLU}_{0.3} \left( \mathrm{BN} \left( y_k \right) \right) \right),$$\end{document} with the second layer following suit on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y_k^{(1)}$$\end{document}. Global average pooling then produces \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$z_k \in \mathbb{R}^{B \times C_{\mathrm{out}}}$$\end{document} for each sub-band. The aggregated output \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Z \in \mathbb{R}^{B \times K \times C_{\mathrm{out}}}$$\end{document} encapsulates the oscillatory and baseline traits across sub-bands, enabling the module to adeptly model IESS-specific waveforms, from δ slow waves to γ spikes. This parameterized approach fosters resilient feature extraction tailored to the intricacies of epileptic spasms. Transformer encoding module The global patterns in EEG with IESS, such as hypsarrhythmia, a hallmark of abnormal brain activity [30], demand models that transcend local receptive fields to capture long-range dependencies and inter-band interactions [31]. The Transformer encoding module fulfills this by leveraging multi-head self-attention (MHSA) to fuse sub-band features into cohesive global representations, drawing from neural network dynamics to mirror epileptic propagation across frequencies. Features from the sinusoidal convolution (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Z \in \mathbb{R}^{B \times K \times C_{\mathrm{out}}}$$\end{document}) are restructured into a token sequence for efficient processing. For each sub-band k, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$z_k \in \mathbb{R}^{C_{\mathrm{out}}}$$\end{document} is divided into S=4 non-overlapping segments, each linearly projected to a token \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{k,s} \in \mathbb{R}^{E}$$\end{document} (E=32), reducing computational load while preserving oscillatory essence. A learnable classification (CLS) token \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\mathrm{cls}} \in \mathbb{R}^{E}$$\end{document} is prepended to synthesize global insights, forming the sequence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T \in \mathbb{R}^{(1 + K \cdot S) \times E}$$\end{document}: 9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T = [t_{\mathrm{cls}}; t_{1,1}, \ldots, t_{1,S}, t_{2,1}, \ldots, t_{K,S}],$$\end{document} augmented with sinusoidal positional encoding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$PE \in \mathbb{R}^{(1 + K \cdot S) \times E}$$\end{document}: 10T\prime = T + PE. The encoder comprises two stacked layers, each integrating MHSA and a feed-forward network (FFN) with layer normalization (LN) and residual connections for stable gradient flow. MHSA employs H=8 heads; for head h, scaled dot-product attention computes: 11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{Attention}_h(Q_h, K_h, V_h) = \mathrm{softmax} \left( \frac{Q_h K_h^\top}{\sqrt{d_h}} \right) V_h,$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q_h, K_h, V_h \in \mathbb{R}^{(1 + K \cdot S) \times d_h}$$\end{document} (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_h = E/H = 4$$\end{document}) are projections. The concatenated MHSA output is: 12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{MHSA}(T\prime) = \mathrm{Concat}(\mathrm{Attention}_1, \ldots, \mathrm{Attention}_H) W^O,$$\end{document} with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W^O \in \mathbb{R}^{E \times E}$$\end{document}. The sub-layer result is: 13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T^{(1)} = \mathrm{LN} \left( T\prime + \mathrm{MHSA}(T\prime) \right),$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{LN}(x) = \frac{x - \mu}{\sigma} \cdot \gamma + \beta$$\end{document} normalizes embeddings (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu, \sigma$$\end{document}: mean/std; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma, \beta$$\end{document}: learnable). The FFN applies two linear layers with LeakyReLU: 14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{FFN}(x) = \mathrm{LeakyReLU} \left( x W_1 + b_1 \right) W_2 + b_2,$$\end{document} (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_1 \in \mathbb{R}^{E \times D}$$\end{document}, D=128; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_2 \in \mathbb{R}^{D \times E}$$\end{document}; biases \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b_1, b_2$$\end{document}), yielding: 15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T^{(2)} = \mathrm{LN} \left( T^{(1)} + \mathrm{FFN}(T^{(1)}) \right).$$\end{document} Post two layers, the refined CLS token \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\mathrm{cls}}\prime\prime \in \mathbb{R}^{E}$$\end{document} is extracted for downstream classification. Classification module The classification module decodes the CLS token, a comprehensive feature representation, to produce the final classification result for IESS detection. The module receives the CLS token \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_{\mathrm{cls}}\prime\prime \in \mathbb{R}^{E}$$\end{document}, where E=32 is the embedding dimension. A linear transformation maps it to classification logits: 16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$z = t_{\mathrm{cls}}\prime\prime W + b,$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W \in \mathbb{R}^{E \times N}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b \in \mathbb{R}^{N}$$\end{document}, N=2 is the number of classes (IESS or non-IESS), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$z \in \mathbb{R}^{B \times N}$$\end{document} contains the logits for each sample in the batch B. The softmax function converts logits to probabilities: 17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_{b,i} = \mathrm{softmax}(z)_{b,i} = \frac{\exp(z_{b,i})}{\sum_{j=1}^N \exp(z_{b,j})},$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$z_{b,i}$$\end{document} is the logit for sample b and class i, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_{b,i} \in \mathbb{R}^{B \times N}$$\end{document} is the predicted probability for sample b and class i, and the predicted label is the class with the highest probability. The cross-entropy loss is used to train the SinTransNet framework: 18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{L} = -\frac{1}{B} \sum_{b=1}^B \sum_{i=1}^N \hat{y}_{b,i} \log(y_{b,i}),$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y}_{b,i} \in \{0, 1\}$$\end{document} is the ground-truth label for sample b and class i. Evaluation metrics For evaluation, four metrics are computed: accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). Accuracy measures the proportion of correctly classified samples: 19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{Accuracy} = \frac{\mathrm{TP} + \mathrm{TN}}{\mathrm{TP} + \mathrm{TN} + \mathrm{FP} + \mathrm{FN}},$$\end{document} where TP, TN, FP, and FN are the true positives, true negatives, false positives, and false negatives, respectively. Sensitivity (or recall) quantifies the true positive rate for the IESS class: 20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{Sensitivity} = \frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FN}}.$$\end{document} Specificity measures the true negative rate for the non-IESS class: 21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{Specificity} = \frac{\mathrm{TN}}{\mathrm{TN} + \mathrm{FP}}.$$\end{document} The AUC represents the area under the ROC curve, which plots the true positive rate against the false positive rate across varying classification thresholds. It quantifies the model’s ability to distinguish IESS from non-IESS samples by ranking their predicted probabilities, with a value of 1 indicating perfect discrimination and 0.5 indicating no discriminative ability. Dataset Data source The proprietary dataset was collected by Shenzhen Children’s Hospital, specifically designed for IESS detection in pediatric patients. Data were recorded in a controlled hospital EEG monitoring ward using a Compumedics Greal EEG system with dry disk electrodes. The study received ethical approval from the Shenzhen Children’s Hospital Ethics Committee (approval number 202305802), with informed consent obtained from legal guardians. The dataset comprises 129 multi-session EEG recordings from 10 pediatric subjects diagnosed with IESS based on EEG and clinical features. While the subject size is limited, this reflects the rare nature of IESS, with an incidence of approximately 2 per 10,000 infants [32]. The specialized clinical requirements and strict diagnostic criteria further constrain data availability. Nevertheless, the comprehensive multi-session recordings per subject provide sufficient data for developing clinically meaningful detection algorithms. EEG analysis identified background activity, interictal epileptiform discharges, and ictal patterns. Clinically, subjects exhibited epileptic spasms and psychomotor developmental delay (PDD), with additional comorbidities such as CDKL5 deficiency disorder (CDD), glucose-6-phosphate dehydrogenase deficiency (G6PD), or neonatal hypoxic-ischemic encephalopathy (HIE), confirmed by pediatric neurologists. The clinical backgrounds of the subjects, including age, gender, and comorbidities, are summarized in Table 1.Table 1Clinical backgrounds of subjects in the proprietary IESS datasetPatient IDAgeGenderComorbiditiesA18 monthsFemaleCDD, PDDC10 monthsMalePDDD11 monthsFemalePDDF10 monthsMalePDDG28 monthsMalePDDH11 monthsFemalePDDI17 monthsFemalePDDL5 monthsFemalePDDM36 monthsMaleG6PDN48 monthsMaleHIE Table 1Clinical backgrounds of subjects in the proprietary IESS datasetPatient IDAgeGenderComorbiditiesA18 monthsFemaleCDD, PDDC10 monthsMalePDDD11 monthsFemalePDDF10 monthsMalePDDG28 monthsMalePDDH11 monthsFemalePDDI17 monthsFemalePDDL5 monthsFemalePDDM36 monthsMaleG6PDN48 monthsMaleHIE Clinical backgrounds of subjects in the proprietary IESS dataset The dataset uses 25 channels, including 19 standard electrodes from the International 10–20 system (Fp1, Fp2, F7, F3, Fz, F4, F8, T3, C3, Cz, C4, T4, T5, P3, Pz, P4, T6, O1, O2) and 6 additional electrodes (AFz, Oz, C5, C6, S1, S2) to enhance detection of IESS-related frontal and occipital epileptiform activity. Recordings span 2–5 minutes, totaling approximately 630 minutes of EEG signals. The 1,941 epileptic spasm events of 129 EEG recordings from 10 pediatric subjects were initially independently annotated by three experienced neurologists, and the final diagnosis was confirmed by their consensus. Seizure durations range from 0.4 to 7.5 seconds (predominantly 0.5–1.5 seconds), recorded at a 1024 Hz sampling rate to capture transient epileptic discharges. Characteristics of the proprietary IESS dataset are summarized in Table 2.Table 2Characteristics of the proprietary IESS datasetAttributeDescriptionSourceShenzhen Children’s HospitalPatients10 pediatric patientsRecordings129 EEG recordingsTotal Duration≈ 630 minutesSampling Rate1024 HzChannels25 (19 standard + 6 additional)Electrode TypeDry disk electrodesEquipmentCompumedics GrealRecording EnvironmentHospital EEG monitoring wardSeizure Events1,941 annotated eventsSeizure Duration0.4–7.5 secondsAnnotation3 experienced neurologists Table 2Characteristics of the proprietary IESS datasetAttributeDescriptionSourceShenzhen Children’s HospitalPatients10 pediatric patientsRecordings129 EEG recordingsTotal Duration≈ 630 minutesSampling Rate1024 HzChannels25 (19 standard + 6 additional)Electrode TypeDry disk electrodesEquipmentCompumedics GrealRecording EnvironmentHospital EEG monitoring wardSeizure Events1,941 annotated eventsSeizure Duration0.4–7.5 secondsAnnotation3 experienced neurologists Characteristics of the proprietary IESS dataset Data segmentation and splitting To ensure robust model training and evaluation, the dataset was preprocessed by the signal preprocessing module, subjected to segmentation into decision windows, and subsequently split into training, validation, and test sets. Considering the predominantly brief duration of most seizure episodes, to augment seizure samples, 0.5-second decision windows with 0.1-second overlapping sliding were applied specifically to seizure segments within each EEG recording. Concurrently, an equivalent number of non-overlapping 0.5-second non-seizure decision windows were selected to balance the dataset. To prevent data leakage, the dataset was partitioned by subject into training, validation, and test sets in a ratio of 7:1:2, yielding 24,128 training windows, 3,484 validation windows, and 8,244 test windows. Results Implementation details The model was implemented in Python 3.8 with PyTorch 1.9.0, with training accelerated through an NVIDIA V100 GPU (32GB VRAM). The Adam optimizer was adopted with learning rates of 1e-4. Meanwhile, a batch size of 128 and a maximum of 50 training epochs were configured, and the early stopping strategy is adopted. The training will stop if the loss function value on the validation set does not decrease for 10 consecutive epochs, and the model parameters corresponding to the lowest validation loss will be restored. To further demonstrate the training stability and robustness of SinTransNet, we present the training and validation accuracy curves in Fig. 2. The model converges smoothly with a gap between training and validation curves, and early stopping effectively prevents overfitting, achieving the best validation accuracy of 84.37% at epoch 23. This indicates good generalization capability under the current dataset constraints.Fig. 2Training and validation accuracy curves of SinTransNet Training and validation accuracy curves of SinTransNet Comparison with existing methods To validate the clinical utility of SinTransNet in addressing IESS diagnostic challenges, we conducted a comprehensive benchmark against state-of-the-art methods on our proprietary dataset, encompassing general-purpose time-series models (TimesNet [33], TimeMixer [34], iTransformer [35]), EEG-specific architectures (EEG Conformer [19], LMDA-NET [21], DARNet [24], CosCNN [17]), and the IESS-tailored PLV-ResNet18 [27]. The experiments were repeated over 5 independent runs, and results are reported as mean ± standard deviation. For fair comparison, all baseline models were trained under the same data splits, preprocessing pipeline, and optimization budget. Hyperparameters (e.g., number of the layers and hidden dimension) were initialized according to the original publications and mildly tuned on the validation set when necessary. As detailed in Table 3, SinTransNet achieved average accuracy of 85.69%, sensitivity of 80.55%, and specificity of 90.76%, surpassing all baselines. The ROC curve in Fig. 3 yields an average AUC of 0.9214 (95%CI: 0.9155–0.9273), demonstrating superior and statistically robust performance over EEG-specific models (average AUC: 0.8606) and general-purpose models (average AUC: 0.7696). In particular against PLV-ResNet18, SinTransNet exhibited an approximately 13% higher sensitivity, underscoring its enhanced ability to detect subtle epileptic spasms, which is critical for timely interventions in pediatric neurology.Table 3Comparison of state-of-the-art methods on the proprietary IESS datasetModelTypeAcc. (%)Spec. (%)Sens. (%)AUCTimesNet [33]General-purpose70.18±0.5374.73±1.8465.57±1.960.7649±0.0050TimeMixer [34]General-purpose69.77±0.7685.89±1.3153.47±2.820.7633±0.0051iTransformer [35]General-purpose71.12±0.2972.99±2.6469.22±3.020.7807±0.0041EEG Conformer [19]EEG-specific74.99±0.7979.58±0.3570.35±1.720.8198±0.0094LMDA-NET [21]EEG-specific81.58±2.1992.33±0.9570.71±4.510.8992±0.0131DARNet [24]EEG-specific79.89±0.1787.12±1.6672.58±1.680.8654±0.0026CosCNN [17]EEG-specific78.35±0.4685.85±0.4170.76±0.670.8580±0.0028PLV-ResNet18 [27]IESS-specific70.23±1.3473.32±11.6267.10±9.450.7812±0.0089SinTransNetIESS-specific85.69±0.1390.76±1.7380.55±1.930.9214±0.0059Fig. 3ROC curve with 95% confidence interval for SinTransNet on the proprietary IESS dataset Table 3Comparison of state-of-the-art methods on the proprietary IESS datasetModelTypeAcc. (%)Spec. (%)Sens. (%)AUCTimesNet [33]General-purpose70.18±0.5374.73±1.8465.57±1.960.7649±0.0050TimeMixer [34]General-purpose69.77±0.7685.89±1.3153.47±2.820.7633±0.0051iTransformer [35]General-purpose71.12±0.2972.99±2.6469.22±3.020.7807±0.0041EEG Conformer [19]EEG-specific74.99±0.7979.58±0.3570.35±1.720.8198±0.0094LMDA-NET [21]EEG-specific81.58±2.1992.33±0.9570.71±4.510.8992±0.0131DARNet [24]EEG-specific79.89±0.1787.12±1.6672.58±1.680.8654±0.0026CosCNN [17]EEG-specific78.35±0.4685.85±0.4170.76±0.670.8580±0.0028PLV-ResNet18 [27]IESS-specific70.23±1.3473.32±11.6267.10±9.450.7812±0.0089SinTransNetIESS-specific85.69±0.1390.76±1.7380.55±1.930.9214±0.0059 Comparison of state-of-the-art methods on the proprietary IESS dataset ROC curve with 95% confidence interval for SinTransNet on the proprietary IESS dataset Ablation study An ablation study was performed to dissect the contributions of the sinusoidal convolution module (SCM) and Transformer encoder module (TEM), elucidating their roles in capturing IESS-specific EEG dynamics. By systematically omitting each component, we evaluated their impact on performance over 5 independent runs. The results are reported as mean ± standard deviation in Table 4.Table 4Ablation study on the proprietary IESS datasetConfigurationAcc. (%)Spec. (%)Sens. (%)AUCw/o SCM80.57 ± 1.0888.64 ± 2.8872.40 ± 5.030.8742 ± 0.0085w/o TEM82.21 ± 0.4490.57 ± 4.0273.74 ± 4.920.8975 ± 0.0010SCM + TEM85.69 ± 0.1390.76 ± 1.7380.55 ± 1.930.9214 ± 0.0059 Table 4Ablation study on the proprietary IESS datasetConfigurationAcc. (%)Spec. (%)Sens. (%)AUCw/o SCM80.57 ± 1.0888.64 ± 2.8872.40 ± 5.030.8742 ± 0.0085w/o TEM82.21 ± 0.4490.57 ± 4.0273.74 ± 4.920.8975 ± 0.0010SCM + TEM85.69 ± 0.1390.76 ± 1.7380.55 ± 1.930.9214 ± 0.0059 Ablation study on the proprietary IESS dataset As shown in Table 4, replacing the SCM with standard convolution resulted in a mean accuracy drop of 5.12% (from 85.69% to 80.57%) and a sensitivity drop of 8.15% (from 80.55% to 72.40%). These differences were statistically significant (paired t-test, p < 0.01), supporting the effectiveness of sinusoidal convolution for adaptive extraction of non-stationary oscillatory features. Spectral visualizations in Fig. 4 further reveal enhanced energy in clinically relevant bands (8–12 Hz and 12–30 Hz) after sinusoidal modulation, aligning with the slow-wave dominance in IESS hypsarrhythmia. In addition, deactivating TEM led to a mean accuracy reduction of 3.48% and a sensitivity reduction of 6.81%, highlighting its contribution to modeling cross-frequency temporal dependencies.Fig. 4Spectral comparison of three sub-bands before and after sinusoidal kernel modulation Spectral comparison of three sub-bands before and after sinusoidal kernel modulation EEG reference montage analysis Exploring the influence of EEG reference montages on detection robustness is crucial for clinical applicability, given the susceptibility of pediatric recordings to artifacts. We evaluated five configurations across 25 channels (19 from the 10–20 system plus 6 extended), with outcomes in Table 5. Bipolar montages, however, resulted in a reduced number of channels due to differential computations (24 for longitudinal and 19 for transverse montages).Table 5Comparative analysis of model performance with various EEG reference montagesReference MontageUsed channelsAcc. (%)Spec. (%)Sens. (%)AUCDefault reference2584.1891.8976.370.9152Linked ear reference2584.4788.6480.260.9105Longitudinal montage2483.6490.6476.570.9081Transverse montage1984.0992.1875.920.9142Average reference2585.6990.7680.550.9214 Table 5Comparative analysis of model performance with various EEG reference montagesReference MontageUsed channelsAcc. (%)Spec. (%)Sens. (%)AUCDefault reference2584.1891.8976.370.9152Linked ear reference2584.4788.6480.260.9105Longitudinal montage2483.6490.6476.570.9081Transverse montage1984.0992.1875.920.9142Average reference2585.6990.7680.550.9214 Comparative analysis of model performance with various EEG reference montages The average reference montage excelled, boosting sensitivity by 4.63% over the transverse variant through global artifact mitigation and improved seizure pattern discernment. The linked ear reference performed comparably but risked diminished sensitivity to hemispheric asymmetries, while the default montage was hampered by artifact vulnerability. These findings advocate for average referencing in IESS workflows to enhance diagnostic reliability in noisy clinical environments. Parameter sensitivity analysis Parameter tuning is vital for optimizing SinTransNet’s balance between expressiveness and efficiency in EEG processing. We first assessed sinusoidal convolution kernel sizes from 5 to 25, contrasting with cosine convolution, as depicted in Fig. 5. A size of 25 yielded peak accuracy, with stability from 21 to 29, demonstrating effective capture of intricate EEG patterns without parameter inflation. Sinusoidal kernels consistently outperformed cosine variants, better suiting the oscillatory nuances of IESS signals.Fig. 5The effect of different convolutional operations and kernel sizes on the model performance The effect of different convolutional operations and kernel sizes on the model performance Additionally, Fig. 6 evaluates Transformer encoder layers, revealing that two layers optimize performance, striking an equilibrium between depth for dependency modeling and computational tractability, essential for real-time clinical integration.Fig. 6The effect of different transformer encoder layers on the model performance The effect of different transformer encoder layers on the model performance Interpretability analysis To improve the interpretability of SinTransNet and alleviate the black-box concern, a multi-band Gradient-weighted Class Activation Mapping (Grad-CAM) analysis was conducted on the sinusoidal convolution branches. Specifically, intermediate feature activations and corresponding gradients were captured through backward hooks, and channel-wise importance weights were computed to generate temporal activation maps for each EEG sub-band. The activation maps were normalized to [0,1] and overlaid with the corresponding filtered EEG signals for visualization, where red regions indicate high model attention and blue regions indicate low attention. Fig. 7 presents Grad-CAM visualizations for the F8 channel (right anterior temporal region) under both IESS and non-IESS conditions. As shown in Fig. 7(a), the IESS sample exhibits strong activations primarily in the δ and θ bands during the central temporal region, indicating that the model focuses on slow-wave oscillatory patterns commonly associated with hypsarrhythmia in infantile epileptic spasms syndrome. In addition, sparse high-frequency activations can be observed in the β and γ bands, corresponding to transient epileptiform discharges and spike-related activities. In contrast, the non-IESS sample in Fig. 7(b) demonstrates substantially weaker and less structured activations across frequency bands. Most regions remain low-attention, suggesting that the model does not rely on nonspecific fluctuations for classification. These observations indicate that SinTransNet captures physiologically meaningful oscillatory patterns and provides clinically relevant interpretability for EEG-based IESS detection.Fig. 7Grad-CAM visualizations on channel F8 (right anterior temporal region). IESS positive sample with strong activations in delta and theta bands (left). IESS negative sample showing weak and diffuse activations (right). Color intensity reflects the importance for model prediction Grad-CAM visualizations on channel F8 (right anterior temporal region). IESS positive sample with strong activations in delta and theta bands (left). IESS negative sample showing weak and diffuse activations (right). Color intensity reflects the importance for model prediction Model complexity and efficiency analysis As shown in Table 6, SinTransNet contains 0.616 M parameters, which is significantly lower than EEG Conformer (1.706 M) and moderately higher than DARNet (0.463 M), while achieving superior accuracy (85.69% vs. 74.99% and 79.89%). The model adopts a multi-branch sinusoidal convolutional design combined with a lightweight two-layer Transformer encoder, enabling effective feature extraction and cross-band modeling without introducing excessive parameter overhead.Table 6Comparison of model complexity and inference efficiencyModelParamsInference Time(M)(ms)EEG Conformer [19]1.7066.25DARNet [24]0.4631.57SinTransNet0.6164.57 Table 6Comparison of model complexity and inference efficiencyModelParamsInference Time(M)(ms)EEG Conformer [19]1.7066.25DARNet [24]0.4631.57SinTransNet0.6164.57 Comparison of model complexity and inference efficiency In terms of deployment feasibility, SinTransNet achieves an average inference time of 4.57 ms per sample, outperforming EEG Conformer (6.25 ms) and remaining within a practical range for real-time applications. Given the sliding-window setting (0.5 s window), the model is capable of supporting near real-time EEG analysis. These results indicate that SinTransNet achieves a balanced trade-off between computational efficiency and detection performance, making it suitable for clinical decision-support scenarios. Discussion The development and evaluation of SinTransNet demonstrate the potential of IESS-related spasm detection using EEG signals. By integrating multi-band EEG decomposition, the sinusoidal convolution module, and the Transformer encoder module, SinTransNet is able to capture both local oscillatory features and global seizure patterns in non-stationary and polymorphically complex EEG signals, providing an assistive approach for supporting neurologists in reviewing complex pediatric EEG recordings. Compared to previous approaches, SinTransNet addresses several key limitations in the literature. While conventional CNN-based methods (e.g., CosCNN [17], LMDA-NET [21]) are effective at local feature extraction, they lack the capability to model long-range dependencies and cross-frequency interactions essential for hypsarrhythmia patterns. Existing Transformer-based models (e.g., EEG Conformer [19], DARNet [24]) capture global context but often overlook the frequency-specific non-stationarity inherent in IESS signals. In contrast, our method combines multi-band decomposition, a sinusoidal convolution module with jointly learnable amplitude, frequency, phase, and bias parameters for frequency-adaptive local modeling, and Transformer attention for global integration. This design enables more flexible characterization of non-stationary oscillatory patterns while preserving physiologically meaningful sub-band structures. From a clinical perspective, balancing sensitivity and specificity is particularly important in automated IESS detection. SinTransNet achieved a sensitivity of 80.55% and a specificity of 90.76%, indicating its potential as an assistive screening and decision-support tool rather than a standalone diagnostic system. Higher sensitivity may help reduce missed spasms, which is critical in pediatric neurology, while high specificity can reduce unnecessary alarms and manual EEG review burden. Compared with PLV-ResNet18 [27], which incorporates phase-locking information but exhibits lower sensitivity, SinTransNet showed improved detection performance for subtle epileptic spasms while maintaining a relatively low false-positive rate. The interpretability analysis further supports the clinical plausibility of the proposed framework. Grad-CAM results indicate that SinTransNet exhibited stronger activations in the low-frequency delta and theta bands for IESS samples, with relatively sparse responses in the beta and gamma bands. This pattern is consistent with reported EEG characteristics of IESS, including slow-wave abnormalities and transient epileptiform discharges. Such interpretability may facilitate clinical adoption, as decision-support systems should not only provide predictions but also offer evidence that can be reviewed by clinicians. The proprietary IESS dataset provides a robust foundation for training and evaluating automated detection algorithms. Unlike public datasets such as CHB-MIT [36] and Bonn [37], which are limited by low sampling rates and lack annotations for epileptic spasms. The proprietary dataset includes 1,941 epileptic spasm events with high density and high precision annotations, capturing the transient and subtle nature of IESS seizures. The high resolution of this dataset and the focus on IESS-specific characteristics enable robust training and evaluation, rendering the dataset suitable for developing IESS-related spasm detection algorithms and supporting clinical decision making. Nevertheless, several limitations remain. Although SinTransNet achieves a favorable efficiency profile with only 0.616 M parameters and an average inference time of 4.57 ms per 0.5-s window, the Transformer-based architecture still incurs relatively higher computational cost compared to purely convolutional models. This may limit its deployment in extremely resource-constrained bedside or portable EEG devices for prolonged continuous monitoring. In addition, the relatively small patient cohort may restrict generalization to broader populations. Future work will explore multi-center external validation and subject-level evaluation strategies such as leave-one-subject-out (LOSO) to further assess robustness. To improve generalization under limited data conditions, knowledge-guided representation learning and multimodal fusion with complementary physiological signals (e.g., EMG or video) will also be investigated. Finally, model compression techniques, including pruning and quantization, along with data augmentation strategies, offer promising directions for even more efficient and clinically deployable IESS detection systems. Conclusion This study presents SinTransNet, a deep learning framework for automated detection of IESS. By combining sinusoidal convolution for adaptive extraction of local oscillatory features with Transformer-based attention for modeling global seizure dynamics, SinTransNet effectively addresses the non-stationarity and polymorphic complexity of pediatric EEG. Validated on a clinically annotated proprietary dataset, it outperforms state-of-the-art baselines, demonstrating strong potential to reduce diagnostic delays and enable timely therapeutic interventions. By bridging local and global EEG representations, SinTransNet enhances diagnostic precision and offers a practical foundation for integrating automated EEG analysis into pediatric clinical workflows.