This paper presents an evaluation of four large language models (LLMs)—GPT-4o, Claude-3.5-Sonnet, Gemini-1.5-Pro, and Llama-3.1-405B-T based on word problems that were formulated to check LLMs level of “understanding” of medical knowledge and their “inferential capabilities”. We constructed a small synthetic dataset by providing pattern to LLM and then asking LLM to generate few word problems related to heartbeat, and evaluated responses of different LLMs for each word problem. Models were evaluated based on their robustness to changing scenarios. Moreover, their responses were compared in order to understand their level of learning about medical knowledge. Consistency of models was checked by providing them answer choices in random order. Then models’ answers were compared with their previous solutions. It was found that all the LLMs involved in the study showed inconsistent behavior. The study also reveals that Gemini and LLama did not show required resilience to changing input. GPT-4o and Claude gave similar answers for most of the queries and showed robustness to changing inputs. The study reveals very interesting insights about strengths and weaknesses of each of LLM and can be very helpful for healthcare professionals in deciding level of adoption of modern artificial intelligence (AI) technologies for identification of heartbeat related disorders like Arrhythmia.