It's probably worth noting that there's a lot of discussion about challenges reproducing the workflow of this paper, and that as-described it seems to suffer from data leakage, so much so that you can replace sections of their algorithm with random initialisation and get at least as good results. See, e.g.:

https://pubpeer.com/publications/C8CFF9DB8F11A586CBF9BD53402...

Having been on both sides of the reviewing process, it seems incredibly difficult to get good peer review of data-intensive studies in medicine, as few people have the time to really dig into the detail of these models.