An observational cohort study was performed using comprehensive electronic medical records from all individuals with type 2 diabetes resident in Tayside and Fife, Scotland. Data were collected and integrated by the Health Informatics Centre, University of Dundee. Record linkage was through the community health index number provided to all individuals in Scotland when they register with a primary healthcare provider, which has been used continuously for all National Health Service clinical activity over the past 30 years.

The Scottish Care Information – Diabetes Collaboration database provided information on type of diabetes, diagnosis date and, from calendar year 2000 onwards, clinically measured BMI. The demography database provided information on age, sex, health board (Tayside or Fife) and social deprivation (residential areas are categorised and ranked into health board quintiles based on an individual’s postcode). Ethnicity data were not collected. Scotland is a predominantly white population (87% according to Scotland’s Census 2022 [15]), and ethnicity was therefore not considered to be relevant to this study. The biochemistry database was used to obtain measurements of HbA1c from 1995 onwards. Finally, the community-dispensed prescribing database, which contains detailed information on all fulfilled prescriptions prescribing in Tayside from 1995 onwards and Fife from 2009 onwards, was used to identify diabetes medications.

We defined the study period as from 3 years before diabetes diagnosis to 5 years after diabetes diagnosis. All BMI measurements during this time were considered; however, to allow separate pre- and post-diagnosis trajectories to be derived, there was a minimum data requirement of four measurements per individual within specific time windows: BMI measurements between 3 and 2 years and 1 and 0 years before diagnosis to derive the pre-diagnosis trajectory, and BMI measurements between 0 and 1 years and 4 and 5 years after diagnosis to derive a post-diagnosis trajectory. Figure 1 shows the pattern of all BMI measurements from the eligible individuals during the 8-year study period. There was a pattern of significant weight loss between 6 months before and 1 year after diagnosis, therefore we did not include this ‘peri-diagnosis’ period in the pre- or post-diagnosis trajectories. Our revised study design required four BMI measurements defined as: BMI−3 (that closest to 3 years before diagnosis in a −3 to −2 year window), BMI−0.5 (that closest to 6 months before diagnosis in a −18 to −6 month window), BMI1 (that closest to 1 year after diagnosis in a 1 to 2 year window) and BMI5 (that closest to 5 years after diagnosis in a 4 to 5 year window).

Pattern of all BMI measurements during the 8-year study period for all eligible individuals, with a smooth line fitted through the population mean. The solid vertical lines show the split of the study period into pre-, peri- and post-diagnosis periods, and the dotted vertical lines indicate the timing of the required BMI measurements: BMI−3 (closest to 3 years before diagnosis in the −3 to −2 year window), BMI−0.5 (closest to 6 months before diagnosis in the −18 to −6 month window), BMI1 (closest to 1 year after diagnosis in the 1 to 2 year window) and BMI5 (closest to 5 years in the 4 to 5 year window)

To ensure sufficient electronic medical record coverage and avoid the impact of COVID-19 on our data, the study period was defined as 1 January 2000 until 31 December 2019. Individuals with a clinical diagnosis of type 2 diabetes between 1 January 2003 and 31 December 2014, diagnosed at ≥ 35 years of age, with a BMI1 ≥ 25 kg/m2 and an HbA1c measurement between 1 and 2 years after diagnosis were eligible. A flow chart of the study population derivation is provided in electronic supplementary material (ESM) Fig. 1.

A model was derived for each individual’s weight change based on observed BMI measurements between 3 years and 6 months before diabetes diagnosis (BMI−3 and BMI−0.5). Covariates in the model were age at BMI−3 (split into bands: <50, 50 to <60, 60 to <70 and ≥70 years), sex, social deprivation coded as 1 (most deprived) to 5 (least deprived), calendar year at BMI−3 (split into approximate quartiles: <2006, 2006 and 2007, 2008 and 2009, 2010 and 2011) and BMI−3 (split into categories: <30, 30 to <35, 35 to <40 and ≥40 kg/m2).

A model was derived for each individual’s weight change based on observed BMI measurements between 1 year and 5 years after diabetes diagnosis (BMI1 and BMI5). BMI measurements were time-dependently adjusted for diabetes treatment and HbA1c change. BMI measurements obtained before medications were started were used as the reference group. Treatments were grouped by drug class: metformin, sulfonylureas, thiazolidinediones, dipeptidyl peptidase 4 inhibitors (DPP4i), glucagon-like peptide 1 receptor agonists (GLP-1RA), sodium–glucose co-transporter 2 inhibitors (SGLT2i) and insulin; and further grouped by insulin (including any other diabetes treatment), monotherapy and various dual and triple therapy combinations. HbA1c change was expressed as the percentage change from the first HbA1c measurement after 1 year from diagnosis, and categorised into five groups: no change (reference group), >0 to <10% decrease, ≥10% decrease, >0 to <10% increase and ≥10% increase.

Covariates in the model were age at diagnosis (split into bands: <50, 50 to <60, 60 to <70 and ≥70 years), sex, social deprivation coded as 1 (most deprived) to 5 (least deprived), calendar year of diagnosis (split into approximate quartiles: 2003 to 2008, 2009 and 2010, 2011 and 2012, 2013 and 2014), BMI1 (split into categories: 25 to <30, 30 to <35, 35 to <40 and ≥40 kg/m2) and HbA1c (measured closest to 1 year after diagnosis but between 1 and 2 years) split into approximate quartiles (<43, 43 to <48, 48 to <54 and ≥54 mmol/mol; <6.1, 6.1 to <6.5, 6.5 to <7.1 and ≥7.1%).

To model the effect of prior weight change on the post-diagnosis BMI trajectory, each individual’s pre-diagnosis trajectory was included as a covariate. In addition, as BMI measurements in the peri-diagnosis period were not included in the pre- or post-diagnosis trajectories, we defined a peri-diagnosis weight change, calculated as (BMI1 − BMI−0.5)/time (in years) between measurements.

The peri-diagnosis weight change and the pre-diagnosis trajectory were split into seven categories of BMI change: rapid loss (>0.5 kg/m2/year), moderate loss (>0.3 to 0.5 kg/m2/year), slow loss (>0.1 to 0.3 kg/m2/year), stable (±0.1 kg/m2/year), slow gain (>0.1 to 0.3 kg/m2/year), moderate gain (>0.3 to 0.5 kg/m2/year), rapid gain (>0.5 kg/m2/year).

To calculate pre- and post-diagnosis trajectories, we applied a linear mixed-effects model with both a fixed and random intercept and slope. To describe the covariance structure among the errors, we used the ‘continuous time/continuous space’ spatial data covariance structure provided within the PROC MIXED procedure in SAS version 9.4 (SAS Institute, USA). After the initial model was fitted, the Studentised residuals were examined, and any BMI measurements more than three standard deviations from the mean were removed as outliers (approximately 1.2% of all measurements) and the model was refitted.

Pre- and post-diagnosis BMI trajectories were calculated by adding together each individual’s random slope with the population average (fixed) slope. These trajectories may be interpreted as BMI change per year, indicating the direction and magnitude of the change in an individual’s BMI over the specified time period.

The models were expanded to include covariates. To model the effect of each covariate on weight change, an interaction term between the covariate and time was included. In addition, the post-diagnosis BMI trajectory models were adjusted for diabetes treatment and changes in HbA1c over time, fitted as fixed effects. We fitted univariable models, and, after a full exploratory analysis, selected a multivariable model that included all significant covariates and any important confounders.

We used the PROC SGPLOT procedure in SAS to produce scatterplots of BMI and HbA1c measurements over time (Figs 1 and 2a,b). To explore overall patterns, we used the LOESS statement to fit a smoothed line through the population mean. These plots are intended as a simple visual representation of the data, with all measurements treated as independent, and are not the output from the mixed models.

Patterns of HbA1c and BMI measurements. (a) Pattern of HbA1c measurements during the peri- and post-diagnosis periods for individuals with rapid weight loss (red line) or rapid weight gain (blue line) during the peri-diagnosis period. (b) Pattern of BMI measurements during the peri- and post-diagnosis periods for individuals with rapid weight loss (red line) or rapid weight gain (blue line) during the peri-diagnosis period. The curves are fitted through the population mean. The vertical lines highlight the time of diagnosis (solid line) and the peri-diagnosis window (dotted lines)

A p value <0.05 was considered statistically significant in all analyses.