Back

Comparing human and AI performance in medical machine learning: An open-source Python library for the statistical analysis of reader study data

McKinney, S. M.

2022-05-07 health informatics
10.1101/2022.05.06.22274773 medRxiv
Show abstract

In seeking to understand the potential effects of artificial intelligence (AI) on the practice of diagnostic medicine, many investigations involve collecting interpretations from several human experts on a common set of cases. In an effort to standardize the process of analyzing the data emerging from such studies, we have released an open-source Python library to perform applicable statistical procedures. The software implements the industry-standard Obuchowski-Rockette-Hillis (ORH) method for multi-reader multi-case (MRMC) studies. The tools can be used to compare a standalone algorithm against a panel of readers, or compare readers operating in two modalities (for example, with and without algorithmic assistance). The software supports both nonequivalence and noninferiority tests. Functions are also provided to simulate reader and model scores, useful for Monte Carlo power analysis. The code is publicly available in our Gitub repository at https://github.com/Google-Health/google-health/tree/master/analysis.

Matching journals

The top 7 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.