The historical context of this study is essential to understanding its implications. Edward Hoffer, MD, a co-author of the study and a member of the LCS at MGH, noted that the current hype surrounding generative AI often obscures the long-standing successes of earlier models. "Amid all the interest in large language models, it’s easy to forget that the first AI systems used successfully in medicine were expert systems like DXplain," Dr. Hoffer said. These expert systems were built with a specific purpose: to serve as a repository of medical knowledge that could be accessed systematically. Unlike modern LLMs, which function as "black boxes" that predict the next likely word in a sentence, systems like DXplain are built on explicit rules and clinical relationships.
The utility of these systems lies in their ability to augment the human mind, which, despite its brilliance, is subject to fatigue and cognitive limitations. Mitchell Feldman, MD, the corresponding author of the study and also a member of MGH’s LCS, emphasized that these tools are designed to serve as a backup to the physician’s own reasoning. "These systems can enhance and expand clinicians’ diagnoses, recalling information that physicians may forget in the heat of the moment and isn’t biased by common flaws in human reasoning," Dr. Feldman explained. He pointed out that human clinicians can sometimes fall victim to "anchoring bias"—the tendency to rely too heavily on the first piece of information encountered—or simply forget a rare condition that fits a patient’s profile. An AI system, whether it is an LLM or a DDSS, does not suffer from these specific human fallibilities.
To put these systems to the test, the MGH team developed a rigorous methodology involving 36 diverse patient cases. These cases were carefully selected to span a wide range of racial, ethnic, age, and gender categories, ensuring that the AI systems were tested against a representative sample of the human population. This diversity is crucial in medical AI research, as there are ongoing concerns about how different models handle demographic-specific medical nuances. For each of the 36 cases, the investigators presented the clinical findings to DXplain, ChatGPT, and Gemini.
The testing was conducted in two distinct phases: one in which the systems were provided only with the patient’s clinical symptoms and history, and another in which lab data was included. The inclusion of lab data is a critical variable, as it often provides the definitive evidence needed to confirm or rule out a diagnosis. The results showed that all three systems benefited significantly from the addition of laboratory results. With lab data included, the systems were able to identify the correct diagnosis a majority of the time. DXplain led the pack with a 72% success rate, followed by ChatGPT at 64% and Gemini at 58%.
The gap between the specialized system and the general-purpose models became even more pronounced when lab data was withheld. In these scenarios, which mimic the early stages of a patient consultation where a doctor must rely on physical exams and patient history alone, DXplain maintained a correct diagnosis rate of 56%. In contrast, ChatGPT’s accuracy dropped to 42%, and Gemini fell to 39%. While DXplain’s lead in this category was clear, the researchers noted that the results were not statistically significant, likely due to the relatively small sample size of 36 cases. Nevertheless, the trend suggested that DXplain’s deep, structured database allowed it to make more accurate inferences from limited information than the broader, less structured knowledge bases of the LLMs.
