In a study published on MedRxiv, a prestigious online health sciences journal, researchers evaluated the ability of the most well-known artificial intelligence system to perform clinical reasoning by testing its performance on the United States Medical Licensing Examination (USMLE), which covers a wide variety of topics (clinical reasoning, medical management, bioethics, etc.).
Out of the 376 test questions (available on the official USMLE website), the researchers presented 305 to the system, excluding those containing visual elements (clinical images, medical photographs) and ensuring that the content was not indexed on Google before January 1, 2022 (the last accessible date for the AI training dataset).
The system achieved an accuracy of 60% or higher, meaning it would have passed the medical licensing exam in most analyses (a surprising result considering this was the first experiment to reach this benchmark).
Regarding the ability to implement the human learning process, the AI bot’s responses were highly consistent, allowing a human learner to easily follow the internal language, logic, and directionality of the relationships contained in the explanation text. This results in high internal consistency and low self-contradiction, indicators of solid clinical reasoning and important parameters of explanation quality.
Can we therefore say that the new AI bots have the partial ability to teach medicine, bringing up non-obvious concepts that might be outside the students’ awareness?
Could Artificial Intelligence become a tool that facilitates the practice of medicine and provides creative insights in the diagnosis and treatment of diseases?
In fact, it already is: it is widely used for analyzing radiological images to determine the presence of pathologies. However, the qualitative result of these tests is an encouraging foundation for future studies on the effectiveness of generative AI as a tool to enhance the human medical training process and the performance of Artificial Intelligence in diagnostics.
Certainly beyond what can be anticipated for the near future, AI with vision systems already allows us to impact the quality of monitoring critical or disoriented patients today, for example, through widespread and active monitoring of patients at risk with simple cameras installed in patient rooms.
TapMyLife’s AI algorithms can capture and evaluate remote monitoring indices, anticipating the alert to the nearest operator thanks to predictive signals of possible bed exits, slips, or even extubations. Evaluating the effectiveness of these systems is straightforward: physical restraint of patients is replaced by active monitoring, and the risk of incidents is drastically reduced.
Artificial intelligence is thus a discipline already generating added value for patients (and resource savings for healthcare facilities).
We are working towards a future where AI is comparable to laparoscopy or minimally invasive robotic surgery, a working mode that helps healthcare professionals perform their tasks more effectively, quickly, and safely, providing improvement insights that, combined with irreplaceable human intelligence, deeply impact the efficiency of patient analysis, monitoring, and care processes.
