Creating a cohort of women with incident diabetes and a background of gestational diabetes from UK electronic health records for retrospective cohort studies

Technical Summary: Creating a Cohort of Women with Incident Diabetes and a Background of Gestational Diabetes from UK Electronic Health Records

Problem Statement
Electronic Health Records (EHRs) offer a rich source of data for studying the long-term complications of diabetes mellitus (DM) and gestational diabetes (GDM). However, utilizing these records for retrospective cohort studies presents significant methodological challenges. Specifically, accurately deriving the date of DM diagnosis is non-trivial, as diagnostic codes may be entered retrospectively or lack temporal precision. Furthermore, distinguishing between chronic DM (e.g., Type 1 or Type 2) and GDM is difficult due to ambiguous coding practices, particularly when GDM is used to describe pre-gestational diabetes incidentally diagnosed during pregnancy. Existing methodologies often exclude younger populations or fail to differentiate between GDM and chronic DM, limiting the ability to study long-term complications (such as diabetic retinopathy) in women with a history of GDM without the confounding factor of subsequent chronic DM development.

Methodology
The study utilized the Clinical Practice Research Datalink (CPRD)-Aurum database, linking primary care records with secondary care data (Hospital Episode Statistics Admitted Patient Care – HES APC) and Office for National Statistics (ONS) death data. The cohort formation process involved extracting records of women aged 18–65 between 2007 and 2013 who had evidence of DM or antidiabetic medication prescriptions.

The core of the methodology involves a series of algorithms designed to:

1. Define the Incident Diagnosis Date: The study established a hierarchy of evidence to pinpoint the date closest to the actual diagnosis of chronic DM. This hierarchy prioritizes:

   • Unambiguous diabetes coding (e.g., "Type 1" or "Type 2" diabetes).
   • Glycaemic test results meeting diagnostic thresholds (HbA1c > 6.5% or 48 mmol/mol; fasting glucose > 6.9 mmol/L; random glucose > 11.0 mmol/L).
   • Prescriptions for antidiabetic medications exclusive to DM (e.g., sulfonylureas, GLP1RAs, DPP4 inhibitors).
   • HES ICD-10 coding for DM.
   • Exclusions: Metformin prescriptions alone were excluded due to high association with Polycystic Ovarian Syndrome (PCOS) and Non-Diabetic Hyperglycaemia (NDH). Short courses of gliclazide were excluded as they may indicate steroid-induced hyperglycaemia.

2. Distinguish Chronic DM from GDM and Other Conditions:

   • Algorithm B (Medication-based): If medication was the only evidence, prescriptions were scrutinized against records of PCOS, NDH, and pregnancy. Metformin was ignored if PCOS/NDH was present; insulin was only considered valid if prescribed outside of pregnancy and at least 3 months post-partum.
   • Algorithm C (Context-dependent Coding): Codes like "Diabetes in pregnancy" were treated as equivocal. If such a code appeared during pregnancy without subsequent unequivocal evidence, it was ignored. If an equivocal code was followed by unequivocal DM coding within a year, the earlier code was assumed to be incidentally diagnosed chronic DM. If followed by a gap, the earlier code was treated as a potential GDM episode.
   • Algorithm D (Temporal Separation): Equivocal evidence overlapping with pregnancy or occurring at least one year prior to unequivocal evidence was excluded to prevent misclassifying GDM as the onset of chronic DM.

3. Data Cleaning and Unification:

   • Identifier Unification: Patients with multiple patient identifiers (patid) due to practice changes or duplicates were unified using the CPRD Master Person Service ID (CPRD MPSid), matching registration periods, and event dates to ensure complete data capture.
   • Date Validation: Erroneous dates (e.g., pre-birth) were corrected or discarded. Subjects with death dates preceding registration or diagnosis were excluded.
   • GDM Cohort Formation: A distinct GDM cohort was formed for women with a GDM episode at least one year prior to their incident DM diagnosis. This gap ensures the GDM episode is distinct from the chronic DM diagnosis.

Key Contributions

• Robust Methodology for Younger Populations: Unlike previous studies that excluded women under 35 or 40, this protocol includes women aged 18–65, improving the capture of early-onset diabetes and longer follow-up periods for complications.
• Differentiation of GDM and Chronic DM: The study provides a detailed, stepwise approach to distinguishing between true GDM (which resolves) and pre-gestational DM incidentally diagnosed in pregnancy, a distinction often missed in retrospective studies.
• Handling Ambiguity: The algorithms specifically address the heterogeneity of EHR coding, such as context-dependent codes and the dual use of metformin, to minimize misclassification.
• Data Unification: The protocol offers a solution for unifying fragmented patient records across multiple GP practices within the CPRD system.

Results
From an initial extract of 730,353 women, the application of the developed algorithms resulted in a final cohort of 83,283 women with incident diabetes diagnosed between 2007 and 2013.

• GDM Exposed Cohort: 3,686 women were identified as having a background of GDM prior to their DM diagnosis.
• Non-GDM Exposed Cohort: 79,597 women were entered into the non-GDM cohort.
• Validation: An internal validation of 50 records confirmed the accuracy of the diagnosis dates and cohort assignment, leading to the refinement of the coding list (removing non-specific codes like "Seen in diabetes clinic").
• Power: The resulting sample size provides a study power of 98% (1 – β = 0.98) to detect a clinically meaningful hazard ratio of 1.12 for diabetic retinopathy, assuming a 30% event rate in the unexposed group.

Significance and Claims
The authors claim that this methodology provides a "potentially widely-applicable, detailed methodology" for forming cohorts of women with incident diabetes, with and without a history of GDM. The significance of this work lies in its ability to facilitate robust retrospective cohort studies exploring long-term complications of GDM, such as diabetic retinopathy, in a controlled fashion where complication rates are not confounded by the subsequent development of chronic DM. By accurately dating the onset of chronic diabetes and separating it from GDM episodes, the study enables the investigation of whether GDM serves as an independent marker for long-term risk, potentially informing earlier screening protocols for at-risk populations. The authors note that this approach addresses a gap in previous literature, which often excluded GDM histories or failed to distinguish them from incidentally diagnosed chronic diabetes.