New Study Finds Biases Limit Benefits of Human-AI Collaboration in Radiology

A new experimental study published in a top economics journal has found that biases in how radiologists use AI assistance limit the potential benefits of human-AI collaboration in medical imaging. The paper comes at a time when AI tools are starting to be deployed in radiology workflows but there is uncertainty around the optimal way to combine human expertise with algorithmic predictions.

The study was conducted by researchers at MIT, who recruited 180 radiologists through teleradiology companies to diagnose chest x-rays under different information conditions. In some cases, the radiologists were provided with AI-generated probabilities that a pathology was present. The researchers found that while the AI algorithm performed better than two-thirds of radiologists, providing its predictions did not improve diagnostic accuracy on average.

This puzzling result occurred because radiologists did not respond uniformly to the AI predictions. When the algorithm was highly confident, radiologist accuracy increased with its assistance. However, accuracy declined when the AI predictions were more uncertain. The authors conclude that radiologists exhibit biases that lead them to underweight or improperly incorporate the AI's information relative to their own assessments.

The findings have important implications for how to optimally deploy AI tools alongside human experts. One approach is more extensive training for radiologists on how to combine their own judgments with algorithmic predictions. The study also suggests that selective automation, where cases are assigned to either the radiologist or the AI alone, may perform better than always involving both.

As AI capabilities in medical imaging continue to progress, insights from this research can help guide the design of human-AI collaboration. Understanding how humans interact with algorithmic information is critical to realizing the potential benefits in accuracy and efficiency from combining human expertise with AI. More studies on biases and integration challenges will be essential as AI becomes further enmeshed in clinical workflows.