Timely detection of seizures and seizure-like EEG abnormalities—such as periodic discharges and rhythmic delta activity—is essential in neurocritical care to reduce mortality and improve outcomes. Manual EEG interpretation is labor-intensive, variable, and infeasible for continuous monitoring, especially in resource-limited settings. Existing automated methods predominantly target binary seizure detection, demand extensive labeled data, and operate as “black boxes,” limiting clinical trust and deployment. This paper introduces ProtoEEG, a few-shot prototypical network that classifies EEG spectrograms via a pretrained EfficientNetV2-S encoder and episodic metric learning. To handle intra-class variability and support-set heterogeneity, ProtoEEG-QA is proposed, refining class prototypes with a query-aware attention module that weights support embeddings by their relevance to each query. On the HMS Harmful Brain Activity dataset under 5-, 10-, and 15-shot regimes, ProtoEEG-QA achieves up to 85.37%±1.02% accuracy (95% CI) and a macro-AUROC of 0.97—surpassing static prototype baselines. Attention weights offer per-instance explanations by highlighting influential support examples, enhancing model transparency and clinical trust. By uniting high accuracy, calibration, and inherent interpretability in a few-shot framework, ProtoEEG-QA provides a scalable solution for explainable EEG classification in critical care and remote health-analytics applications.