The authors referenced literature review findings that data and code about AI decisions in healthcare couldn’t be shared and analyzed due to the privately held “corporate black boxes” of data. This inability to share the data complicated replication, and the results that were available often demonstrated “a lack of methodological rigor” because of it. Another idea the authors introduced to address challenges with AI in healthcare is “data leakage,” where an algorithm bases decisions on its training information that wouldn’t be available in real scenarios. They think this could also be managed with shared access to the data and would improve the overall reproducibility of decisions.
Another benefit of an open system, according to the authors, is that biases that affect vulnerable populations could be avoided. For example, researchers in the US might miss inclusion of a disease variant that’s common in South America, but data analysis by South American researchers would expand the algorithm’s capabilities.
Support authors and subscribe to content
This is premium stuff. Subscribe to read the entire article.
