← Back to Home

As AI Auscultation Becomes Less Accurate, Can Physician Oversight Catch the Errors?

A randomized experiment involving 48 physicians found that interpreting results independently before viewing AI advice still failed to prevent diagnostic performance from declining as the system became less reliable, highlighting the need for a more comprehensive safety net for clinical monitoring.

By SURL BioNews

When diagnostic AI starts making mistakes, can physicians with the final say detect them in time? A study published on November 13 in *npj Digital Medicine* tested this safeguard using heart sound interpretation: as AI advice became less reliable, the performance of physicians working with AI also declined.

The study randomly assigned 48 physicians to two groups to classify recorded heart sounds as normal or abnormal in a simulated remote auscultation setting. One group saw the AI advice first; the other had to classify the sounds independently before viewing the advice and making a final judgment. The latter approach was intended to encourage physicians to form their own opinions first and reduce their tendency to follow the system directly.

The experiment consisted of 40 questions divided into four phases, with the accuracy of AI advice declining successively to 90%, 60%, 50%, and 0%. These were failure scenarios deliberately arranged by the researchers. The final phase also introduced a mismatch between the audio and visual images to test whether physicians could detect a system malfunction.

Physicians were able to correct some errors, but when reliability was low, human–AI performance remained below the baseline established by physicians interpreting the sounds independently. Requiring physicians to answer on their own first did not improve overall human–AI performance either. In other words, retaining final decision-making authority may not be enough to offset the influence of incorrect advice.

These findings have clear limits: they come from a small, controlled heart sound classification experiment, did not directly measure patient treatment outcomes, and cannot represent all diagnostic AI. Real clinical settings also involve medical histories, other examinations, and time pressure, which may change how physicians use advice.

The study therefore raises questions for hospitals and developers: how should system performance be tracked continuously, and who should identify failures and initiate a response? The authors advocate combining system monitoring with organizational arrangements for accountability. These measures still need to be validated, but the experiment already serves as a reminder that clinical AI safety cannot rely solely on human oversight at each diagnosis.

References

  1. npj Digital Medicine