ChatGPT proves unreliable in heart risk assessment
The free-to-use AI system ChatGPT is hoped to be a useful tool in the future of healthcare but failed to deliver consistent analysis of patient data in the assessment of heart risk.
A utopic look at artificial intelligence (AI) would suggest it can help tackle at least very basic problems in healthcare. However, implementing AI software such as ChatGPT in routine practice is still very much in its infancy.
Chest pain is a frequent complaint in emergency departments, with patients often unnecessarily hospitalized to determine potential acute coronary syndromes, even when they do not present with significant heart disease. High-risk cases can be easier to spot but some low-risk patients are harder to detect. A team at the University of Washington, WA, USA recently evaluated ChatGPT-4 as a risk assessment tool for patients with acute nontraumatic chest pain.
TIMI and HEART are two current protocols to predict major adverse cardiac events, though they have faced previous scrutiny over their sensitivities. The sophisticated AI of ChatGPT-4 represents an opportunity to upgrade current approaches to stratify patients with acute nontraumatic chest pain.
A spoonful of sugar makes the AI count down
Researchers have designed an AI device using the world’s first implementation of physical reservoir computing, paving the way for a new generation of AI software.
The study involved thousands of randomly simulated patient datasets. Though ChatGPT’s assessment correlated well with TIMI and HEART in assessing patients with major adverse cardiac events, it offered inconsistencies and even gave different diagnostics and recommendations when presented with the same repeated data. Additionally, ChatGPT failed to recognize variables e.g. gender and race as key risk factors that would increase the likelihood of a patient being at a higher risk of a cardiac event.
“ChatGPT was not acting in a consistent manner,” commented lead author Thomas Heston. “Given the exact same data, ChatGPT would give a score of low risk, then next time an intermediate risk, and occasionally, it would go as far as giving a high risk.”
Even though current protocols for predicting major adverse cardiac events have faced scrutiny, to consider ChatGPT as an upgrade could result in unpredictable patient care The researchers suggested the inconsistencies in ChatGPT’s reporting and analysis are likely a result of its attempt to best replicate the variance in human language – potential for a Turing test but not when delivering medical diagnoses.
However, there were some positives from the study. ChatGPT showcased its ability to handle and integrate a lot of clinical data in a short space of time, indicating that ChatGPT could offer multiple diagnoses to support clinicians speed up their decision-making.
ChatGPT-4 is a product of its vast knowledge base: the equivalent of hundreds of millions of resources, both medical and non-medical. The authors suggested that training a specialized ChatGPT model on medical literature alone would improve its ability in the future to make more informed decisions.
