Michael N. Liebman

my opinion (biased of course!) is that 1-4 above reflect how every problem is more complex than we realize (or admit) and humans prefer not to deal with the "known unknowns" let alone even acknowledge or attempt to deal with the "unknown unknowns"...because they are not trained or do not know how to apply critical thinking and systems thinking. Hindsight bias simply involves evaluating the subset of that complexity that would have supported 1-4. I have concerns that many of the advanced technologies only further support these biases as their validation comes mainly from more efficiently reproducing things that we already think we know....especially in critical domains like medicine.

Most people function "operationally" and ask if you can do "x" to solve their problem, but actually have not looked more strategically to identify what the root cause of the problem actually is...I have dealt with many pharma companies and failed clinical trials and found that a very simple question was never asked at the start..of the drug discovery and development process..."how confident are you in your diagnosis and the potential heterogeneity of patients with this diagnosis, i.e. need to stratify the disease"

even more than the 5 Why's, use the questions taught in journalism (back in high school)...Ask "Who", "What", "Where", "When", "How" and then "Why"

AI analysis requires the aggregation of data, commonly from multiple sources but needs to confront the reality that data, from different sources, may have the same label but not be equivalent. Many/most tests do not have universal standards for "+"vs "-", e.g.is 1% or 10% the threshold used? Triple negative breast cancer patients may not be classified "TNBC" at different institutions because of different tests being used or. different thresholds, and yet combined into a single cohort for AI/ML analysis. It is critical for data scientists to understand specific data contexts from clinicians, etc to improve the accuracy and dependability of their analysis. Failure to do so limits the quality and interpretability of analytic results.

The most critical part of complex problem solving is the appropriate identification and specification of the problem that needs to be solved...then taking it apart to see what approaches and technologies are best suited for the component questions...and bringing those to the AI and data science experts who specialize in those technologies...this is much more effective and efficient than bringing them in too early to ask them to solve a problem that is not well defined

While I agree with Jan Berger that AI has the potential to reduce medical errors, I add this caution: if we are training AI using existing standard of care as the gold standard, we run the risk of perpetuating the system that we acknowledge is currently producing these errors. We need to venture into the "unknown unknowns" to remove the potential for conformational bias that may limit true progress in improving healthcare