This is an outstanding paper on a project to build an EWS to predict circulatory failure. Two things that stood out were the clear combination of clinical and medical expertise – too often one of these is missing in a project – and the creative use of feature engineering to create high-resolution lookback windows.
A clear oversight, however, is the failure to discuss the implications of throwing out over 1/3 of the dataset due to missing lactate and MAP measurements. Patients without these measurements would almost certainly not be in circulatory failure (according to a consultant cardiologist we spoke to!), making it likely that 1/3 of the healthy cohort has been thrown out. Doing this boosts the PPV of your model significantly, meaning there are serious implications for the claimed effectiveness of the model.
