NeuroPACT: Skuteczne wykrywanie padaczkowych ataków upadków z rejestru EEG i EMG przy użyciu hierarchicznie kompresowanego transformatora
NeuroPACT: Efficient Detection of Epileptic Drop Attacks from EEG and EMG Using a Hierarchically Compressed Transformer
W skrócie
[Preprint - wstępne wyniki] Naukowcy opracowali sztuczną inteligencję NeuroPACT do wykrywania padaczkowych ataków upadków – niebezpiecznych, krótkotrwałych napadów, które mogą powodować urazy u dzieci z opornią padaczką. System analizuje zapisy aktywności elektrycznej mózgu (EEG) i mięśni (EMG), zmniejszając złożoność obliczeń o ponad 40% przy utrzymaniu wysokiej czułości (92%). Model został przetestowany na danych od 625 pacjentów w szpitalu rozwojowym i niezależnie zweryfikowany na 88 pacjentach w innych ośrodkach, wykazując potencjał do wspomagania lekarzy w szybkim rozpoznawaniu tych niebezpiecznych napadów.
Oryginalny abstract (angielski)
Abstract Objective Epileptic drop attacks (DAS) are catastrophic, fleeting ictal events that pose a substantial risk of physical trauma in patients with refractory pediatric epilepsy. In clinical practice, identifying these events within days of continuous video-EEG (VEEG) monitoring is an immense challenge, often complicated by the extreme brevity of the events and the computational overhead of processing high-density, multi-channel electrophysiological signals. Methods To address this, we developed NeuroPACT (Neurophysiological Progressive Attentive Compression Transformer), a deep learning architecture designed for high-efficiency segment-level detection of DAS. By integrating hierarchical spatiotemporal compression with a novel Summary-Augment Module (SAM), NeuroPACT distills high-dimensional EEG-EMG inputs into compact representations without sacrificing discriminative precision. We evaluated the model using a multi-center framework: a developmental cohort of 625 patients (8,068 segments) from Peking University People's Hospital and an independent external validation cohort of 88 patients (1,628 segments) from Nanjing and Wuhan Children's Hospitals. Results NeuroPACT achieved robust discriminative performance with high sensitivity (0.9183) and a low false-negative rate (0.0817) in external validation. Compared to state-of-the-art baselines such as EEGPT, NeuroPACT realized a substantial reduction in computational complexity—requiring 38.07% fewer parameters and 42.22% fewer GFLOPs—while maintaining, and in some metrics exceeding, existing detection standards. Systematic ablation revealed that hierarchical compression serves as the primary driver of computational efficiency, while cortical EEG provides the core discriminative substrate, with EMG offering limited marginal utility for binary segment detection. Conclusion Our findings demonstrate that strategic spatiotemporal compression enables clinical-grade ictal detection with a significantly reduced footprint, addressing the "efficiency-discrimination" trade-off inherent in long-term neuro-monitoring. NeuroPACT offers a scalable foundation for future real-time, low-latency diagnostic aids. Future clinical translation will prioritize prospective validation, robustness against motion-induced artifacts, and multiclass subtyping to provide comprehensive, automated support for epilepsy care.