Under HIPAA, one method for de-identifying PHI is determining that "a person with appropriate knowledge of and experience with generally accepted statistical and scientific principles and methods" couldn't re-identify the patients. With ready access to AI, we must assume capabilities have changed.
It’s important to get agreement on success metrics prior to starting implementation, otherwise you run the risk of a mismatch on expectations. Look carefully at whether the selected metric(s) will be impacted at all by the changes, and whether the metric is aligned with the central objective.
Your organization's AI governance maturity is an extension of your existing governance framework. If governance has historically been complex or incomplete, you'll need to fix that foundation quickly so that you can adapt to the new pace and risks. Without a firm foundation, governance fails.
Startups commonly measure how well their models perform based on retroactive data. But we can’t afford to see that data as infallible. Differing processes or perceived meanings, bad judgement, and biases all live in the data before we start. How do you normalize your data for model training?
One reason physicians struggle with alert fatigue is that hospitals default to more inclusive alerting - just in case - rather than paring them to a level where they are all meaningful. I understand the fear of missing an alert, but more-is-better results in frustration and unread critical messages.
Healthcare is complex. To be successful in healthcare tech, you need to lean into the complexity.