1.WHAT PROBLEMS WILL AI POSE FOR HEALTHCARE?
Health-related machine learning models perform particularly poorly on reproducibility measures when compared to other machine learning disciplines. A prominent computer-medicine scientist at his MIT in Cambridge, who led the review, pointed out that a key problem is the relative paucity of publicly available medical records. This can encourage prejudice and dishonesty .
2.WHY SCIENTISTS STILL HAVE HOPE IN AI?
In a machine learning study that predicts protein structure, scientists say they have been successful in excluding proteins from the test set that are very similar to those used in the training set. They really need to be good at addressing and balancing the educational aspects of the data they display in their algorithms and making their datasets representative.
3.HOW AI IS BELIEVABLE?
Checklists are widely used in healthcare and provide an easy way to reduce technical problems and improve reproducibility. In machine learning, checklists help researchers to recognize in the correct order the many small steps that need to be performed correctly in order for results to be valid and reproducible.
There is still work to be done, but the growing dialogue around reproducibility in machine learning is encouraging and helping to combat research silos. And at scientific conferences, reproducibility is a common focus. “It used to be a small group of esoterics that focused on reproducibility. Now people ask questions and it feels like the conversation is going. They want it to go faster, but At least I don’t feel like I’m screaming into space.”