Combining structured and unstructured data in EMRs to create clinically-defined EMR-derived cohorts.

Area of Science:

• Health Informatics
• Clinical Data Mining
• Electronic Medical Records (EMR) Research

Background:

• Limited studies exist on systematically querying production EMR systems for clinically-defined populations, especially using free-text data.
• Developing methods to construct patient cohorts from EMR data is essential for advancing clinical research and healthcare systems.

Purpose of the Study:

• To provide a generalizable methodology for creating clinically-defined patient cohorts from EMRs.
• To demonstrate the utility of combining structured and unstructured EMR data for cohort identification.

Main Methods:

• An exemplar cohort of patients with possible acute coronary syndrome (ACS) was defined using clinical criteria.
• Inclusion criteria were mapped to structured data fields (orders, investigations, procedures, ECGs, diagnostic codes) and unstructured clinical notes within the EMR.
• Data were extracted from two local health districts (LHDs) in Sydney, Australia, over a 5-year period.

Main Results:

• ECG images (54.8%) and clinical documentation keywords (41.4-64.0%) were the most frequent criteria for identifying possible ACS encounters.
• Structured data fields like orders/investigations (27.3%) and procedures (1.4%) were less common; diagnostic codes were present in only 3.7% of encounters.
• Consistent trends were observed across different LHDs and over time, indicating reliability.

Conclusions:

• Combining structured and unstructured data from EMRs enables the creation of clinically-defined patient cohorts.
• This methodology is a prerequisite for validating real-time clinical decision support and learning health systems.