Preprocessing clinical variables to be used in SLSTM model
- Calibration: since different lab sites tend to provide clinical
measurements with site-specific biases, it is important to correct variables
relative to the data used to train and validate the model. This is generally
true of any model leveraging combinatorial relations among the variables.
- Scaling nominal's: conversion of variables with discrete levels (ie, nominal's) to variables
with ordered scaled levels to improve statistical efficiency and leverage
- Zing: standardize all variables to z-scores, (x-mean)/stdDev
z-scores improves model robustness.