Referee 1.1

[cgreene]

The authors summarized over 400 literature references purely by narratives. The authors provided synopsis for each important reference, but lacks synthesis of related work. It would be better to synthesize related work into a table and analyze their characteristics.

[agitter]

Before addressing this, I'd like to confirm what the reviewer is asking for. Does "related work" mean comparing/contrasting groups of related papers that are discussed in the review? Or comparing/contrasting the neural network-based methods with other types of algorithms?

[michaelmhoffman]

I understood it to mean the former. The latter does not fit all uses of "related work" in the reviewer's comments.

[agitter]

@cgreene we haven't done much to address this yet. I don't think our new table is exactly what the reviewer had in mind.

[cgreene]

I agree. I am hoping that the text that goes along with the figure in #684 will also help with this; however, I have thought about it but not actually written it (been on my to-do list for 2 days).

[agitter]

Would it help if we give examples of where different network architectures are commonly used? We can't be exhaustive, but we could say things like 1D CNNs are used for text, DNA/RNA/amino acid sequence, etc. Graph convolutions are used in PPI networks and molecular graphs.

[SiminaB]

I think that would be a good idea @agitter! In #684 I had also mentioned discussing which ones are more used for supervised vs unsupervised applications.

[SiminaB]

If the mapping works out, we could add another column to the table. Another option would be to make another table (or bulleted list), with something like:

Patient categorization:
Imaging: transfer learning
Text applications: CNNs
EHRs: CNNs, RNNs, LSTM
etc

or to build on this figure https://github.com/greenelab/deep-review/blob/master/content/images/biotm.pdf

[agitter]

Closed by #799