Knowledge graphs (KGs) provide a powerful framework for representing complex, interrelated data - an essential capability for navigating complex domains such as medicine. This is particularly true in the study of stroke, one of the leading causes of death and long-term disability worldwide. Currently, there is no comprehensive medical KG for the stroke domain, largely because the specialized knowledge embedded in the vast medical literature makes manual construction prohibitively time-consuming. In this work, we explore the steps towards automated construction of a Stroke KG by comparing four relation-extraction methods and proposing novel LLM-as-a-judge-based evaluation. Specifically, we benchmark one unsupervised system (OpenIE), two supervised frameworks (REBEL and ReLiK), and one large-language-model approach (Gemma 2 9B). Our contributions are twofold: first, we provide a systematic evaluation of multiple extraction strategies tailored to a domain-specific KG; second, we propose a hybrid evaluation protocol – combining traditional statistical metrics with an LLM-as-a-judge paradigm – to assess graph quality more comprehensively. The results demonstrate that LLM-based methods hold particular promise for generating a robust, high-coverage Stroke KG, a key resource for accelerating both research and clinical decision-making in stroke care.