Using the landmark method for creating prediction models in large datasets derived from electronic health records

Area of Science:

• Biomedical Informatics
• Health Data Science
• Clinical Research Methods

Background:

• Electronic health records (EHRs) provide abundant health data for research cohort creation.
• Challenges in EHR data analysis include time-varying predictors, variable patient follow-up lengths, and large dataset sizes.
• These issues hinder efficient and accurate clinical outcome prediction.

Purpose of the Study:

• To present landmark analysis as a method to overcome EHR data challenges.
• To demonstrate the use of patient birthdays as landmark times for dynamic dataset creation.
• To illustrate the application of this method in predicting clinical outcomes.

Main Methods:

• Utilized landmark analyses with patient birthdays as landmark times.
• Developed dynamic datasets for predicting clinical outcomes.
• Applied techniques to two large-scale cohort studies from Cleveland Clinic.

Main Results:

• Landmark times effectively incorporate time-varying information.
• Landmark times provide unbiased reference points independent of patient exposure.
• Dataset size is reduced compared to true time-varying analyses, enhancing computational efficiency.
• Successfully applied to datasets of 4.5 million and 17,787 unique patients.

Conclusions:

• Landmark analysis, using patient birthdays, is a valuable tool for EHR research.
• This approach effectively addresses key challenges in analyzing longitudinal health data.
• The method facilitates robust and computationally feasible prediction of clinical outcomes from EHRs.