Is it okay to restructure a literature review to fit a broad topic?

The answer to your title question, “is it okay to restructure a literature review...” is yes. There is no one uniform way that literature reviews must be written. What is important is to make sure that your literature review does its job, which is to do a comprehensive review of the publications on a particular topic.

The only worry with the structure that you propose is the ability to do a completely comprehensive review (which is a difficulty in writing any given literature review, but is especially troublesome in your topic). By simply discussing papers and not cross referencing them within topical discussions, your literature review turns into a data-dump instead of a review. It is not necessarily bad to have a broad topic, nor is it bad to discuss and “rinse and repeat” through several papers, but it is essential that your review is well-structured and does not arbitrarily choose the pieces it is reviewing.

Restructuring is fine, but do not compromise the quality of the review because the topic is broad.

In principle, I don't think what you suggest is a good approach, assuming that your goal is to produce a helpful review of the literature (for yourself and for your readers). The reason for the "similar structure" of "most literature reviews" as you describe it is that it is a more useful format. It is far more useful to readers to understand what the literature says topic by topic than it is for them to read your paper-by-paper summary report.

An excellent explanation of the difference is presented in Webster and Watson (2002). Although what they cover is not exactly the same thing, it is pretty close: they call the recommended topic-by-topic approach "concept-centric"; as for the paper-by-paper approach that you propose, they condemn it as "author-centric". Very simply, it is not very useful for the readers.

If I understand your description correctly, you seem to think that "medical applications for machine learning" is one big topic, and so you have little choice but to then go paper-by-paper. On the contrary, every topic can be sliced into subtopics; that is very much so for a broad topic like yours. You could divide topics by different fields of healthcare; by different machine learning approaches (e.g. classification; regression; clustering; etc.); or by many other subtopics that present themselves in your reading of the literature.

However, I specifically recommend along the lines of Webster and Watson (2002) that you focus on theoretical concepts as your organizing principle. In machine learning, this would mainly mean that you would group studies based on similarity of the target variables (at least, that would apply for supervised learning studies), for example, grouping by targets that study predictors of the same kind of disease; targets of predictors of the effectiveness of certain procedures; etc. Unsupervised learning techniques might not be as readily amenable to this kind of grouping (since they do not have the causal structure of theory), so for those kinds of studies, other organizing schemes (such as those I listed in the previous paragraph) might be more suitable.

In my research on literature reviews, I have found that this kind of organizing principle (that is, based on theoretical concepts) is more useful to readers. (My most recent work is not yet published, but I have an earlier working paper on this topic that you might find helpful.)