Back

A universal translator for AI scores: Providing context using error

Chung, M.; Bernstein, M. H.; Yala, A.; Baird, G. L.

2025-03-04 radiology and imaging
10.1101/2025.02.28.25323066 medRxiv
Show abstract

Artificial intelligence (AI) programs in radiology typically provide a numeric score for each case that correlates with the underlying pathology. However, these scores are not readily interpretable by themselves. To address this, we propose improving score interpretability by providing the False Discovery Rate (FDR) and False Omission Rate (FOR) corresponding with each score threshold. Using an open-source AI program for breast cancer, we estimated FDR and FOR across a range of AI scores using data from 130,712 digital screening mammograms, of which 907 were positive and 129,805 were negative. FDR and FOR ranged from 99.27% and 0.03%, respectively, at the low end of the score distribution to 60.98% and 0.65%, respectively, at the high end of the distribution. Providing these error rates alongside AI scores allows clinicians to consider the balance of trade-offs between false positive and false negative interpretations.

Matching journals

The top 8 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.