Evaluation of the obesity paradox in diabetes: a longitudinal case control study

Abstract

In the general population, obesity is associated with significantly higher cardiovascular disease (CVD) and all-cause mortality risk compared to normal weight. Among patients with type 2 diabetes mellitus (T2DM), some studies reported significantly higher mortality risk for those with normal weight at the time of diagnosis compared to their obese counterparts – indicating the presence of the obesity paradox. However, a detailed exploration of the possible reasons for the obesity paradox in patients with T2DM has not been conducted.The clinical-epidemiological aim of this thesis was to conduct an extensive exploration of the potential role of weight change before the diagnosis of T2DM and ethnicity in the association between BMI and CVD / mortality risk in patients with T2DM, using a large nationally representative patient-level electronic medical record (EMR) database. Given the methodological and analytical challenges in using such databases to design and conduct epidemiological outcome studies, the methodological aims were to compare and generalise (1) statistical methodological approaches for the robust extraction of a disease cohort and (2) methods for imputation of missing longitudinal risk factor data.This thesis used the patient-level primary care EMR database from the United Kingdom –The Health Improvement Network (THIN) database. A robust methodological framework that incorporates several biostatistical methods was used to address the aims of this thesis. First, an extensive machine learning (ML) classification algorithm was used to identify and extract a cohort of patients with T2DM from the THIN database. Second, an exact matching algorithm was developed and used to match four non-diabetic controls to each patient with T2DM based on age, sex, and ethnicity. Longitudinal measurements of anthropometric, cardiovascular, and glycaemic risk factors were extracted and arranged in 6-monthly non-overlapping windows. Third, the predictive mean matching technique of multiple imputation was used to impute missing longitudinal cardiovascular and glycaemic risk factor data. These applied methodological tasks were conducted to ensure the ability to draw robust inferences on the epidemiological aims of this thesis, including the use of different study designs, inclusion, and exclusion criteria. Generalised linear model under general estimating equations setup, with unstructured covariance was used to evaluate body weight trajectories before and after diagnosis of T2DM while multivariate stratified Cox proportional hazards regression was used to assess the association of BMI at diagnosis with mortality risk in patients with T2DM.For large EMR databases like THIN (n=~11 million patients), the use of extensive data mining / ML algorithms are required to robustly identify patients with a disease of interest. Furthermore, multiple imputation of missing longitudinal risk factor data was a valid approach as the distributions of imputed data over 24 months post diagnosis of T2DM were similar longitudinally compared to that of the unimputed data. While patients with T2DM had a significantly higher mean BMI levels and prevalence of comorbidities at diagnosis compared to non-diabetic controls, similar prevalence of cardiovascular multi-morbidity was observed among White European, African-Caribbean, and South Asian patients who were normal weight at diagnosis.Weight trajectory analysis among patients with T2DM and no established comorbidities at diagnosis, showed that normal weight and overweight patients experienced a small but significant reduction in body weight six months before diagnosis, followed by significantly increasing trend post-diagnosis. For patients in all obese categories, consistently increasing body weight was observed six months before diagnosis followed by a decreasing trend after diagnosis. Furthermore, a paradoxical association of BMI with mortality risk was observed among patients who did not lose body weight before diagnosis – where normal weight patients had 35% significantly higher adjusted mortality risk compared with the grade 1 obese patients. However, among patients experiencing weight loss before diagnosis, BMI at diagnosis was not associated with mortality risk. The obesity paradox was further observed among White Europeans and South Asians where those with normal body weight at diagnosis were significantly more likely to die earlier by 0.6 years and by 2.5 years respectively, compared to their respective obese patients.The findings of this thesis add to the evidence base that patients with T2DM, who were normal weight at the time of clinical diagnosis have significantly higher mortality risk compared to those who were obese, and this may partially be driven by different cardiovascular and glycaemic risk profiles of different ethnic groups. Empirical results from this thesis suggest that there was no evidence of pre-existing latent or severe disease conditions being overrepresented in normal weight patients. In fact, dynamic changes in body weight before clinical diagnosis of T2DM were independent of pre-existing latent or severe disease conditions. The increased mortality risk in the normal weight group may reflect differences in the aetiology of diabetes in normal weight people and emphasises the importance of addressing risk factors for excess mortality in this group