Quantifying medical histories with the Kaplan-Meier (KM) estimator for feature extraction of electronic health records in machine learning

Abstract This protocol aimed to describe data transformation procedure of medical histories from electronic health records (EHRs) to historical rates by Kaplan-Meier (KM) estimation. The applicability is to extract features from real-world, time-varying data of EHRs, for developing but not limited to a machine learning prediction model. By this extraction technique, machine can learn medical history of a condition in each healthcare provider, as a differential quantity through time in term of affecting a future health state, without a need to access EHRs of other healthcare providers. However, this protocol needs a sufficient amount of longitudinal data from the most subjects in EHRs. The key stages consisted of time interval computation, historical rate derivation, and data transformation into historical rates.