Intracranial hemorrhage (ICH), a life-threatening emergency, requires rapid diagnosis through analysis of computed tomography (CT) scans. Automating intracranial hemorrhage detection using artificial intelligence is crucial, but is hampered by the scarcity of labeled training data. To overcome this limitation, we present a novel weakly supervised deep learning framework. This framework integrates Attention-based Multiple Instance Learning (Att-MIL) with Sparse Variational Gaussian Processes (SVGP) to improve diagnostic accuracy and interpretability. Att-MIL mimics clinical radiology practice by identifying abnormal CT slices and modeling critical dependencies between slices. Meanwhile, SVGP provide probabilistic diagnostic results, explicitly highlighting the influence of training data size on decision confidence. Additionally, we introduce a data redundancy reduction technique that enhances learning efficiency by removing semantically redundant information across imaging modalities. Our proposed method demonstrates superior performance in binary ICH classification compared to existing Gaussian Processes-based MIL approaches. Furthermore, it exhibits robust performance in sub-label categorization, highlighting its effectiveness and generalizability across diverse classification tasks. The source code is available and maintained in the GitHub repository ( https://github.com/nhatminhtrieu/Brain-Stroke-Diagnosis/ ).