A deep learning system called MedSegNet-AXU segmented multiple sclerosis lesions from 3D MRI with a reported 98.58% Dice score and used radiomics plus graph neural networks to predict sensory, motor, and visual impairment. The approach focuses on converting imaging and clinical data into outcome-linked predictions. Separately, a blinded evaluation found that AI chatbots were mostly safe on sudden cardiac death advice, but recurring safety gaps appeared across multiple models—particularly around emergency escalation and CPR guidance. That result reinforces that clinical performance is not the only hurdle; response behavior under real-world prompts matters. In combination, the advances show rapid gains in algorithmic capability while also stressing that model evaluation frameworks must cover clinical actionability and safety, not only accuracy.