I found this one really interesting. The study authors used clinical biomarkers and generated straightforward rules that doctors can interpret, but when this was tried elsewhere the reliability was much lower. We put together a summary discussing the paper and limitations here