Adaptive Prototype-Based Subtle Transient Pattern Transformers for Enhanced Neonatal Seizure Classification and...

Neonatal seizures are important neurological episodes that need to be identified and managed early to prevent adverse effects. Cohort comparison and rule-based models do not account for the nuances of electroencephalography (EEG) signals and require much time to analyse and interpret raw neonatal EEG signals. The latest advancements in AI-based methods demonstrate the possibility of such tasks, but they still possess some drawbacks, such as a requirement for big labelled datasets, inefficiency in computation processes and noise sensitivity, which hinder clinical use. In this regard, to overcome these limitations, the new multi-component framework named Adaptive Prototype-Based Subtle Transient Pattern Aware Transformer (APSTPT) is introduced for neonate seizure detection, classification and its severity quantification. Pre-processing is the first stage, where noise and artefacts are removed, and only relevant brain signals are amplified. This is succeeded by feature extraction, where power spectral density and phase locking value components are used to identify important spectral and phase-synchronisation characteristics. These aspects are fine-tuned using cross-channel covariance attention to handle inter-channel dependencies. Real-time adaptation of the prototype with the use of prototype learning makes the classification of seizure types better and more dynamic because the finer details of the signal are captured. Moreover, multiscale entropy analysis measures the signal complexity across different temporal scales and properly differentiates the severity of the seizure into mild, moderate and severe cases, respectively. This structured approach allows for accurate separation of seizure events on the time-series and also flexibility according to the characteristics of different datasets. Experiments conducted using the TUH EEG Corpus and Zenodo dataset prove the effectiveness of the proposed framework, with a classification accuracy of 99.74% and a severity assessment accuracy of 98.87%, which is higher than previous approaches. Therefore, the APSTPT framework presents stable performance irrespective of window lengths and the condition of different datasets, showing its flexibility for real-time implementation in the clinical setting.