The impact of major depressive disorder on glycaemic control in type 2 diabetes: a longitudinal cohort study using UK Biobank primary care records – BMC Medicine

Study population

The UKB is a health study of ~ 500,000 individuals recruited between 2006 and 2010 in the United Kingdom, aged 40–70 [12]. Linked primary care records are available for ~ 230,000 individuals (46%), encompassing clinical events, blood test results and prescriptions, providing longitudinal patient information [13].

T2D Classification and Validation

UKB participants with primary care records were classified into T2D cases and controls. T2D cases met specific T2D diagnostic criteria (detailed below), while controls did not. The T2D diagnostic criteria were validated using T2D polygenic scores (PGSs) in European ancestry participants meeting genetic quality control criteria [14, 15] (Additional file 1: Methods S1).

Type 2 diabetes (T2D) diagnostic criteria

T2D cases were identified based on the presence of any two of the following: 1) a primary care diagnosis code for T2D (Additional file 2: Table S1), 2) an ICD9/ICD10 diagnosis code for T2D (Additional file 2: Table S2), 3) any HbA1c measurement > 48 mmol/mol (6.5%), 4) any prescription for glucose lowering medication (Additional file 1: Methods S2 [16], Additional file 2: Table S3), and 5) a self-reported diagnosis for T2D with reported age at onset > 35 years. T2D diagnosis date was then the earliest occurrence among the identified criteria. T2D cases were excluded if they had a primary care code specific to type 1 diabetes, an insulin prescription within a year of T2D diagnosis, or a prescription for multiple diabetic medications at T2D diagnosis (Additional file 1: Methods S2-S3). Individuals prescribed one diabetic medication (monotherapy) at T2D diagnosis were not excluded.

Exclusion criteria for longitudinal analysis

For the longitudinal analysis, we excluded participants meeting any of the following criteria: 1) were a T2D control, 2) had fewer than two HbA1c measurements after T2D diagnosis, 3) had age at T2D diagnosis  38 mmol/mol (3.5%) [17] within a six-month window of T2D diagnosis, 5) had diagnostic codes for bipolar, psychotic or substance-use disorders [18], 6) had a MDD diagnosis without a diagnosis date, or 7) were missing self-reported ethnicity (Fig. 1).

Outcome measures

For the T2D diagnostic criteria validation, the outcome was T2D case–control status for UKB participants of European ancestry with primary care records available. For the longitudinal analysis, the outcome was repeated measures of HbA1c (mmol/mol) after T2D diagnosis. HbA1c data were taken from: 1) primary care records up to 2017 (Additional file 2: Table S4), where older observations recorded in %-units were converted to mmol/mol [19], and 2) all UKB biomarker assessments (2006–2016), where a validated correction was applied to account for lower average HbA1c values from the UKB biomarker panel compared to primary care [20] (Additional file 1: Methods S4). The indexing date was T2D diagnosis, with a maximum follow-up period of 10 years.

Exposures

In the longitudinal analysis, four MDD exposure variables were considered simultaneously. Two were related to individuals diagnosed with MDD prior to T2D (pre-T2D MDD), and two were related to individuals diagnosed with MDD after their T2D diagnosis (post-T2D MDD).

Pre-T2D MDD exposures: (1) History of MDD at T2D diagnosis (MDD_index). This binary variable indicates whether an individual had ever received a MDD diagnosis at the index date. (2) Pre-T2D MDD duration at index (years). This semi-continuous variable quantifies the time between MDD diagnosis and T2D diagnosis when MDD_index equals 1, and is 0 otherwise. These time-invariant exposures were used to examine the impact of pre-T2D MDD on HbA1c, including interactions with T2D disease duration.

Post-T2D MDD exposures: (1) Change in MDD diagnosis (MDD_change). A time-varying binary variable indicating, at each observation time (t), whether an individual has been diagnosed with MDD between t and the index date. It captures information about individuals diagnosed with MDD during follow-up. For individuals with no MDD diagnosis occurring before t, and those diagnosed prior to T2D diagnosis, this variable is set to 0. (2) Post-T2D MDD duration (years). A time-varying, semi-continuous variable equalling the time between MDD diagnosis and t if MDD_change equals 1, and 0 otherwise. This variable allows post-T2D MDD participants to have different HbA1c time-slopes after their MDD diagnosis. Note, UKB participants classified as having a MDD diagnosis required at least two diagnostic codes for a depressive disorder or episode diagnosis in the linked primary care records [18] (Additional file 1: Methods S5).

Covariates

Covariates were extracted from UKB assessments and/or primary care data. Covariates extracted from UKB initial assessments were: sex, assessment centre, self-reported ethnicity, Townend Deprivation Index (TDI), qualifications, ever smoked, and never consumed alcohol. Covariates extracted from both UKB assessments and primary care data were: age at T2D diagnosis, HbA1c at T2D diagnosis, body mass index (BMI) at T2D diagnosis (Additional file 1: Methods S6 [21, 22]), systolic (SBP) and diastolic (DBP) blood pressure at T2D diagnosis (Additional file 1: Methods S6), number of HbA1c, BMI and blood pressure measurements taken prior to T2D diagnosis, and T2D disease duration (time since diagnosis). Covariates extracted from primary care only were diabetic medications. Glucose lowering medication at each HbA1c observation was identified using prescription records up to three months prior to HbA1c measurement (Additional file 1: Methods S2). This information was grouped into four medication categories: 1) M0 (‘no medication’), 2) M1 (‘metformin or a single medication’/ monotherapy), 3) M2 (‘two medications’/ dual-therapy), and 4) M3 (either ‘3 or more medications’ or ‘insulin’). Two medication variables were then created. Firstly, a binary variable indicating monotherapy at T2D diagnosis versus no prescribed T2D medication (M0 vs M1 at diagnosis; recall exclusion criteria removed individuals prescribed multiple medications at diagnosis). Secondly, a time-varying medication variable using the four categories M0–M3.

Apart from T2D disease duration and time-varying medication, covariates were treated as baseline measurements, but it is important to note that the index date and UKB assessment dates are different. For example, measurements from UKB initial assessments (TDI, qualifications, etc.) were collected between 2006 and 2010, and 56% of individuals were diagnosed outside of this timeframe. However, results from models with and without UKB initial assessment covariates yielded similar conclusions (Additional file 2: Table S5). Further details on covariates are available in Additional file 2: Table S6 [18,19,20,21,22,23].

Statistical analysis

All analyses were performed using R version 4.2.2 and visualised using ggplot2.

Validation of T2D diagnostic criteria

To validate the T2D definition, we tested whether PGSs for T2D [24] predicted T2D case–control status. PGSs, calculated using PRSice v2 [25, 26] at eleven P-value thresholds, were tested for association with T2D case–control status, adjusting for six genetic ancestry principal components, assessment centre and genetic batch effect (Additional file 1: Methods S7).

Longitudinal analysis

We employed linear mixed effects models (MEMs) to investigate longitudinal associations between HbA1c and MDD using the nlme package [27, 28]. All models incorporated random intercepts and time slopes, with a continuous-time autoregressive 1 (CAR1) residual correlation structure to account for autocorrelation. A series of three analyses were performed.

(1) Unadjusted model with selection

This analysis focused on the primary fixed effects of time (T2D disease duration) and the MDD exposures. Time was included using a restricted cubic spline (RCS) with four knots to allow for a non-linear temporal trend in mean HbA1c [29, 30]. Interactions between the pre-T2D MDD exposures and time-splines were considered, with the impact of post-T2D MDD on HbA1c time-slopes captured by post-T2D MDD duration. Selection of the final model was determined by Akaike Information Criterion (AIC) and likelihood ratio tests (LRTs). Models with semi-continuous exposures required their corresponding binary indicator to be included.

(2) Adjusted model

The selected unadjusted model was extended to include covariates, plus interactions between the time-splines and HbA1c at T2D diagnosis, BMI at T2D diagnosis and the medication variables.

(3) Residual within-subject variation in HbA1c

To explore the association between within-subject variation and MDD, the adjusted model was updated to allow residual variation to differ by MDD diagnosis variables. Firstly, we assessed if pre-T2D MDD (MDD_index) individuals had different residual variation compared to other participants. Secondly, we used MDD_change to investigate if post-T2D MDD individuals had different residual variation after their MDD diagnosis. Thirdly, a retrospective, time-invariant binary indicator for post-T2D MDD individuals was introduced to assess any differences in within-subject variation over all of follow-up. Please see Additional file 1: Methods S8 [27,28,29,30] for a mathematical explanation of these models. Likelihood ratio tests (LRTs) were used to test for MDD diagnosis-related heterogeneity in residual variation.

HbA1c at T2D diagnosis had a high level of missingness (40%), but all included participants had a HbA1c measurement within a six-month window of this date. Multiple imputation (MI) with the Amelia package [31] generated 50 MI datasets (Additional file 1: Methods S9) that were used in analyses (2) and (3). MEM estimates were pooled using Rubin’s method, with pooled Wald-like P-values presented for the fixed effects [32, 33]. Parameter estimates for RCSs are difficult to interpret. Therefore, in addition to a summary of the MDD exposure fixed effects (including P-values from t-tests), results are presented using plots of predicted HbA1c [34, 35]. Details of covariates included in MI and/ or MEMs are given in Additional file 2: Tables S6-S7.