Very interesting read, thank you! From a data scientist's standpoint - it can fail before any of those three, in hidden assumptions baked into the modeling pipeline, ones that are often invisible long before deployment.
I'd argue there's another stage belonging to data scientists: interrogate what the model actually learned, and whether the ground truth and features it was trained on reflect the decision you're trying to support - or just the system that produced the data.
Where do you see most clinical innovations fail today: action contract, clinical placement, or scale economics? Comment here!
Very interesting read, thank you! From a data scientist's standpoint - it can fail before any of those three, in hidden assumptions baked into the modeling pipeline, ones that are often invisible long before deployment.
I'd argue there's another stage belonging to data scientists: interrogate what the model actually learned, and whether the ground truth and features it was trained on reflect the decision you're trying to support - or just the system that produced the data.
I wrote about this from the modeling pipeline side, if it's useful: https://helenapelinsaric.substack.com/p/your-ground-truth-is-not-in-the-data?utm_campaign=post-expanded-share&utm_medium=web