JAMA viewpoint on deep learning in medicine

https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2718342 (very short)

We're focusing in biology, not medicine, but some of this is relevant:

Data quantity:
> Applying deep learning techniques to more complex and heterogenous disease states, such as chronic heart or kidney failure, may require tens of millions of samples to create a reliable diagnostic model, as well as data from multiple sources (eg, text inputs, imaging, laboratory values, vital signs). For many complex clinical conditions, the quantity of reliable data that is required may not be readily available.

Data quality:
> It is more difficult for deep learning models, as it is for the human brain, to recognize reliable patterns from scattered and noisy information than from structured information.

Also mentions the black box issue and others. They have some solutions too for making medical studies more amenable to DL (collect diverse, high-quality data, have appropriate model security etc)

Read the linked JAMA viewpoint and compare its points on data quantity, data quality, black-box models, and security with the project's biology focus. The issue names no manuscript file or requested edit, so clarify what content should be added and where before starting.

Written by the indexing model from the issue text.