A25-07 Trust Me, I’m a Robot: The Harms of Medical Self Diagnosis Using Large Language Models

Abstract Introduction Artificial Intelligence (AI) Large Language Model (LLM) applications, such as ChatGPT, have surged in popularity among physicians and patients. Traditional web searches typically provide physician-reviewed articles with adequate context and clear guidance on when to seek further care. On the other hand, LLM applications often generate a list of diseases and treatments that patients can readily act upon. In this case report, we present a patient that used ChatGPT to self-diagnose themself with Deep Vein Thrombosis (DVT) and suffered significant neurologic injury and paralysis due to complications with anticoagulation. Case Description A 50-year-old male with no significant medical conditions developed right calf pain. He consulted ChatGPT, and through his search learned about DVT. Deferring further medical evaluation or imaging, the patient purchased warfarin from an unregulated online pharmacy. He self-administered 5mg of oral warfarin daily without INR checks. After 8 months of daily warfarin use, the patient developed progressive bilateral lower extremity weakness, sensory loss, and urinary retention. Upon presentation to the emergency department his labs revealed a supratherapeutic INR of 9.1. Neurologic examination showed minimal sensation or strength caudal to the mid-thorax. MR imaging of the spine revealed a loculated compressive epidural hematoma spanning from the C6 to L3 spinal levels. After warfarin reversal, the patient underwent emergent multilevel spinal decompression from T2 to L1. Postoperatively, he was diagnosed with bilateral DVTs and treated sequentially with IVC filter placement and eventual therapeutic anticoagulation. The patient remained in the intensive care unit for 18 total days of treatment. At the time of discharge, he had significant neurologic recovery and was able to ambulate independently. Discussion The use of AI by patients and physicians is the highest it has ever been. Therefore, informed medical practice must involve appropriate counseling towards patients about proper usage of AI tools. While AI/LLMs are revolutionizing medicine, our case highlights the potential harm to patients when recommendations from LLMs are followed without adequate physician oversight. Studies assessing diagnostic accuracy of ChatGPT show the potential for a high degree of error in its responses. Further awareness of this growing practice will help drive changes in both medical education and organization-level guidelines and regulations about the proper use of AI in medical care. This abstract is funded by: None