A machine with no medical training just matched the diagnostic accuracy of a specialist who spent a decade in university.

I looked at the data from the University of Chicago and the University of Illinois Chicago and the results are clear.

Software now mimics the skill of a veteran neurotologist. Large Language Models identified inner-ear disorders using only the text of a patient history. These machines had no prior medical training and no fine-tuning for the clinic. It is a tough pill to swallow for those who believe a decade of university is the only path to a diagnosis.

The AI used logic to solve cases that usually require years of human experience.

The researchers gave the models ten specific patient scenarios. The cases included Meniere’s disease and vestibular neuritis and vestibular migraine and Benign Paroxysmal Positional Vertigo. I saw how the algorithms parsed the descriptions of spinning and dizziness without the need for a physical exam.

The machine did not look into the eyes of a patient to check for nystagmus or perform the Dix-Hallpike maneuver. It simply processed the words. The accuracy of GPT-4 matched the performance of five board-certified doctors who spent their lives studying the mechanics of the ear.

The speed of the software creates a new reality for patients in rural clinics.

A general practitioner could use the model to identify a disorder during a first visit instead of waiting for a specialist referral. This is not a perfect metaphor for human care but the efficiency is undeniable. The software pulls the signal of the disease out of the background noise of the patient’s life. I noticed that the models relied on the frequency of specific nouns to reach their verdicts.

The diagnosis arrives in seconds rather than the weeks or months required for a specialist appointment.

Behind the Scenes

The research team utilized EurekAlert! to share how these findings might change the triage process in emergency rooms. During the testing phase, the team ensured the models operated on zero-shot learning to test pure reasoning.

The experiment proves that the collective medical knowledge of the internet is already embedded in the code of these systems. The software did not hesitate when presented with conflicting symptoms.

It weighed the probability of each disorder based on the linguistic markers found in the patient’s own voice.

Tell us what you think

We are asking for your feedback because this study suggests a shift where the patient’s voice becomes the primary data set for a computer. This could change how you interact with your doctor and how long you wait for answers.

• Does the idea of a machine diagnosing your vertigo without a physical exam make you feel more or less confident in your healthcare?
• If a general practitioner used an AI to confirm a specialist-level diagnosis on the same day, would you still feel the need to see a human expert?
• Should medical schools change how they teach diagnostic skills now that algorithms can match board-certified performance?