physionet_challenge_2012: per-stay means mix descriptors with time series and misstate urine output #149

Problem

Source preparation (ported from the notebook that made physionet.csv) takes each variable's mean over a stay's 48-hour record. Some variables do not suit a mean:

• Weight is both a general descriptor recorded on admission and an hourly time series; the mean mixes the two.
• Urine is urine output per measurement in mL; its mean over a stay is not the stay's output, which is the clinically used quantity (its total).
• Only means are kept: the last value, extremes and trends, which ICU mortality models commonly use, are lost, and so is the number of measurements.

Scope

• Source preparation and dataset generation for physionet_challenge_2012; both budget settings.
• Changes paper results: it is in the kdd26 dataset list.

Options to decide

• Use the admission value for general descriptors and the total for Urine.
• Add other summaries (last value, minimum, maximum), at the cost of more features.
• Or keep the means and document the choice (the dataset page describes it today).

Acceptance criteria

• The aggregation of each variable is decided, implemented in PhysionetChallenge2012Dataset.make_prepared_source, and stated on the dataset page.
• just qa passes.