Key challenges for delivering clinical impact with artificial intelligence

Source: Deep Learning on Medium

Key challenges for delivering clinical impact with artificial intelligence

What are some problems when it comes to building the healthcare systems → there are going to be some problem? (a good test set as well as data set shift and biased all of those variables will be a problem).

The methods have to be robust → and more.

AI-enabled a lot of automation → but there are limited clinical adoptions.

Google has made some progress in chest and eye diseases. (and even brain tumors → cancer skin lesions and more). (and genomics and more). (Huge EHR data was very beneficial for predicting hospital mortality). (also recognize speech and more).

AI can extract patterns and data → this can lead to good prediction.

There is a limited number of prospective study → so we do not know how the method will do in the real world. (this is a bad thing → we need to have much more data).

Many papers are only available in preprinted servers → which is low rank and not really beneficial research. (and not all of them report with an honesty → most of them are just reporting bias values).

The metric used in the real world is different from the metric used in machine learning optimization. (good test sets → same ones on what clinical trials are done → we need to test models in these data).

And we need methods → that are explainable and interpretable.

Since models are able to pick up random correlations. (generalizations is a very hard thing to do).

Also tackling new populations is hard to do as well → making sure the model does well on the new dataset is hard.

A lot of algorithms are biased → when it comes to race and more. (and being robust to adversarial attacks → the model should know how to deal with these outliers.).

And since the data format is different → from the hospital to the hospital → we are not able to create fast the pace of effect.

Human-computer interactions → , as well as other players, should come in for this research. (explainability is critical).

There seems to be a trade-off between → explainability and performance. (the deeper the model → harder to understand). (black box models are hard to use).

We also need to know how a computer algorithm affects human diagnosis.