Sarah Almaghrabi

Transformer-based analysis of physiological signals critically depends on how input tokens are constructed, yet conventional tokens derived from raw temporal windows or spectrogram patches may overlook the intrinsic oscillatory organization of electroencephalographic (EEG) activity. To address this limitation, we propose AMST-Former, an adaptive mode-wise synchrosqueezing Transformer for patient-independent epileptic seizure detection. Each EEG channel is first decomposed into oscillatory modes using Adaptive Local Iterative Filtering (ALIF), after which the Synchrosqueezing Transform (SST) is applied independently to each mode. This decomposition-before-synchrosqueezing strategy reduces cross-component interference in instantaneous-frequency estimation while preserving mode-specific time–frequency dynamics. The resulting time–frequency patches are embedded together with channel, mode-frequency, and positional information. A hierarchical factorized-attention mechanism then models token relevance across time–frequency regions, oscillatory modes, and EEG channels before a Transformer encoder captures their interactions. Evaluation using a repeated patient-independent five-fold cross-validation protocol on the CHB-MIT database (22 subjects) yields an F1-score of 70.76% and an area under the precision–recall curve (AUPRC) of 0.870. At the event level, AMST-Former achieves a sensitivity of 81.4% at an average false detection rate of 1.39 false detections per 24 h. Ablation experiments show that mode-wise SST provides more effective representations than direct SST, per-mode Hilbert spectra, and fixed filter-bank decomposition. Synthetic experiments confirm the cross-component bias predicted by the analysis and shows that the SST of a mode remains accurate when separation is imperfect, whereas its Hilbert instantaneous frequency does not. Furthermore, the factorized-attention mechanism provides mode, time–frequency, and channel-level explanations, whose faithfulness is assessed through a deletion analysis. These results support adaptive mode-wise time–frequency tokenization as a promising strategy for interpretable, patient-independent EEG seizure detection.

Transformer-based analysis of physiological signals depends critically on how input tokens are constructed, yet conventional tokens derived from raw temporal windows or spectrogram patches may overlook the intrinsic oscillatory organization of electroencephalographic (EEG) activity. To address this limitation, we propose AMST-Former, an adaptive mode-wise synchrosqueezing Transformer for patient-independent epileptic seizure detection. Each EEG channel is first decomposed into oscillatory modes using Adaptive Local Iterative Filtering (ALIF), after which the Synchrosqueezing Transform (SST) is applied independently to each mode. This decomposition-before-synchrosqueezing strategy reduces cross-component interference in instantaneous-frequency estimation while preserv- ing mode-specific time–frequency dynamics. The resulting time–frequency patches are embedded together with channel, mode-frequency, and positional information. A hierarchical factorized-attention mechanism then models token relevance across time–frequency regions, oscillatory modes, and EEG channels before a Transformer encoder captures their interactions. Evaluation using a repeated patient-independent five-fold cross-validation protocol on the CHB-MIT database (22 subjects) yielded an F1-score of 70.76% and an area under the precision–recall curve (AUPRC) of 0.870. At the event level, AMST-Former achieved a sensitivity of 81.4% at one false detection per 24 h. Ablation experi- ments showed that mode-wise SST provided more effective representations than direct SST, per-mode Hilbert spectra, and fixed filter-bank decomposition. Furthermore, the factorized-attention mechanism provides mode, time–frequency, and channel-level explanations, whose faithfulness was assessed through deletion analysis. These results support adaptive mode-wise time–frequency tokenization as a promising strategy for interpretable, patient-independent EEG seizure detection.

Background/Objectives: Typhoid fever remains a major public health challenge in many low-income countries, where overlapping clinical symptoms and the limited reliability of conventional diagnostic procedures hinder accurate diagnosis. This study aims to develop a reliable and efficient diagnostic framework that automates typhoid fever detection from clinical data while preserving patient privacy. Methods: To achieve this objective, we propose a hybrid framework combining genetic algorithm (GA)–based feature selection, a Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) deep learning classifier, and federated learning. The GA identifies the most informative clinical features, reducing redundancy and computational complexity. The selected features are then used to train a CNN–LSTM model in a federated learning setup using the Federated Averaging (FedAvg) algorithm, enabling collaborative model training across multiple clients without sharing raw patient data. Results: Experimental results show that the proposed framework achieves 92% accuracy, with a strong F1-score and satisfactory sensitivity. Compared to models trained on the full feature set, the proposed approach requires less memory and shorter training time, while maintaining balanced performance under class imbalance. Conclusions: These results demonstrate that integrating evolutionary feature selection, deep sequential learning, and federated training provides an effective and privacy-aware solution for multi-class typhoid fever diagnosis. The proposed framework is particularly suitable for clinical environments with limited data access and constrained resources.

Context: In numerous low- and middle-income nations, malaria remains a significant issue due to the challenges associated with diagnosing it through thin blood smears. The appearance of images can vary significantly depending on the microscope type, magnification, lighting conditions, slide preparation methods, and staining techniques. Due to the delicate morphology of parasites, false negatives might adversely affect patient care. Objective: To achieve optimal outcomes from validation, it is essential to construct a robust and easily replicable process. This pipeline should integrate the optimal elements of classical machine learning and end-to-end deep learning, enhance reliability by pairwise ensembling, and select ensemble weights in a logical, data-driven manner. Method: To achieve our objective, we propose two tracks. The initial track encompasses real-time augmentation, convolution-based feature extraction, and the training of calibrated classical classifiers. The second module focuses on training many convolutional networks from inception to completion. Subsequently, we construct paired ensembles and employ a hybrid methodology to select convex weights for combining the findings. This method initially evaluates a set of candidate weights and then refines them to maximise validation accuracy. Results: The precision of the two-track architecture consistently improves, transitioning from conventional baselines to end-to-end models. Optimal and consistent enhancements are achieved through weighted ensembling. Utilising optimised fusion reduces the incidence of false negatives for subtle parasites and false positives caused by staining artefacts. This yields an accuracy of 96.35% on the reserved data and reduced variance across folds. Conclusions: The integration of augmentation, multiple modelling tracks, and optimal pairwise ensembling yields the highest accuracy in categorising malaria smears. It facilitates further enhancements by incorporating supplementary models, multi-class extensions, and operating-point calibration.

Forecasting the power generation from intermittent renewable energy sources, such as Photovoltaic (PV) systems, is crucial for the reliable operations of power systems. In this paper, we consider the task of spatially aggregated PV power generation from large-scale, grid-connected and geographically dispersed PV sites. PV power generation data is highly uncertain, non-linear and non-stationary, making accurate forecasting very challenging. We present a new approach, Wavelet Convolutional Neural Networks (WCNNs), by combining Wavelet Transformation (WT) with Convolutional Neural Networks (CNNs). The WCNNs approach first applies time-invariant WT to decompose the highly fluctuating PV power time series into multiple components. It then predicts the approximation (i.e., low frequency smoothed time series) and details (i.e., high frequency random noise) using CNNs and linear regression, respectively. Extensive evaluation using a real dataset from the Australian Energy Market Operator (AEMO) shows that WCNNs is an effective approach and outperforms the state-of-the-art machine learning models both with and without WT.