Basic and clinical science posters: Epidemiology

A quantitative study to explore the relationship between ethnicity and renal dysfunction in a cohort of type 2 diabetes participants

IM Reece¹, J Karalliedde⁴, R Upsher³, K Ismail³, K Winkley²

¹Community Diabetes Team, Lambeth Diabetes Intermediate Care Team, London, UK, ²Nursing and Midwifery Research, King's College London, London, UK, ³Department of Psychological Medicine, King's College London, London, UK, ⁴School of Cardiovascular Medicine & Sciences, King's College London/GSTT, London, UK

Background Risk factors for diabetic kidney disease are poorly understood.

Aim Our aim was to determine the rate of progression and predictors of renal dysfunction in people of White and Black Caribbean and Black African ethnicity 6 years post-diagnosis of type 2 diabetes.

Methodology Adults with type 2 diabetes were recruited within 6 months of diagnosis to the south London diabetes (SOUL-D) study. Six-year follow-up was conducted using medical record data. Renal dysfunction was defined as: fall in estimated glomerular rate ≥3ml/min/year and/or an albumin creatinine ratio ≥3mg/mmol/l. Baseline data, including age, gender, HbA1c (%), macrovascular disease and triglycerides, were entered into logistic regression to determine prospective association with renal decline.

Results A subsample of 442 people from the SOUL-D cohort of White and Black ethnicity was followed up at 6 years. Results demonstrated that ethnicity (Black Caribbean: OR 0.51 (95% C.I. 0.27–0.95), White OR 0.67 (95% C.I. 0.37–1.21) compared with Black African participants), existing macrovascular disease (2.02 (95% C.I. 1.08–3.79) and baseline HbA1c OR 1.21 (95% C.I. 1.06–1.37) were associated with renal decline, after adjusting for baseline variables.

Discussion Black Caribbean compared with Black African ethnicity is associated with a slower risk of renal dysfunction, as was higher HbA1c around time of type 2 diabetes diagnosis and the presence of macrovascular complications. Although this study followed up a proportion of the original SOUL-D cohort, results suggest more active monitoring of renal decline in people of Black African ethnicity is important.

Performance of dementia specific risk predictors in Edinburgh Type 2 Diabetes Study

W Khalid¹, A Sluiman¹, GM Terrera², M Strachan³, J Price¹

¹Usher Institute, Medical College, University of Edinburgh, Edinburgh, UK, ²Centre for Clinical Brain Sciences & Centre for Dementia Prevention, University of Edinburgh, Edinburgh, UK, ³Western General Hospital, University of Edinburgh, Edinburgh, UK

Aim Diabetes is associated with 70% increased risk of dementia. Use of appropriate risk prediction scores in clinical settings could help to identify and manage diabetes patients at high risk for dementia. We evaluated the performance of individual and combined predictors of diabetes-specific dementia risk score (DSDRS) in the UK-based Edinburgh Type 2 Diabetes Study (ET2DS).

Methods ET2DS is a population-based prospective study of 1,066 men and women with type 2 diabetes aged 60–75 years, living in Lothian region of Scotland. The study collected data on known modifiable and potentially causal risk factors for dementia along with novel biomarkers related to diabetes specific cognitive impairment. Logistic regression analysis was used to evaluate the associations.

Results Over a period of 10 years of follow up, 100 people (incidence=14.04/1,000 person years) developed dementia. Age (OR, 1.09; 95% CI 1.04–1.15), male gender (1.6; 1.08–2.65), microvascular disease (1.56; 1.00–2.42) and depression (1.66; 1.05–2.60) were identified as significant predictors for development of dementia, whereas diabetic foot, cerebrovascular diseases, cardiovascular disease and acute metabolic events were not statistically significantly associated. The C-statistics of the full model combining all risk factors was 0.662 and Hosmer–Lemeshow showed good calibration of 4.64 (p-value 0.590).

Summary The risk prediction model showed relatively low acceptable discrimination ability in the ET2DS population. Our study suggests that there is a need to improve the predictive ability of the model and perform external validation. Predictive performance may be improved by adding APOE genotype, cognitive test results and/or novel biomarkers pertinent to diabetes.

Diabetes type as a computable phenotype in a large electronic healthcare database

LAK Blackbourn¹, T Caparrotta¹, SJ McGurnaghan¹, S Hatam¹, PM McKeigue², HM Colhoun¹

¹Institute of Genetics and Molecular Medicine, University of Edinburgh, Edinburgh, UK, ²Usher Institute of Population Health Sciences and Informatics, University of Edinburgh, Edinburgh, UK

Aims Ensuring the most correct possible assignation of diabetes type in electronic healthcare records (EHR) is fundamental to extracting maximum research potential. However, there will often be missing data in EHRs, especially for variables such as autoantibodies and c-peptide levels that precludes validation against a gold standard definition. Nonetheless reducing misclassification by cleaning against available data may be useful. Here, we aimed to quantify the level of re-classification from clinician assignation of diabetes type that can be achieved by cleaning against other fields in the EHR.

Methods Using SCI-Diabetes, containing the EHR of >99% of those with diabetes in Scotland (n=499,962), linked to other data we constructed an algorithm that re-assigned clinician-designated diabetes type given evidence of (i) secondary (SD) or (ii) monogenic diabetes (MD), (iii) conflicting time-to-insulin or (iv) conflicting duration of oral anti-diabetes medicines usage.

Results Of those clinically assigned as having type 1 diabetes (n=45,808), 11.2% had type re-assignation (10.5% to type 2 diabetes, 0.1% to MD, 0.7% to SD). Of those clinically assigned as having type 2 diabetes (n=431,759), 1.06% were re-assigned (0.8% to type 1, 0.04% to MD, 0.2% to SD). Re-assignations between type 1 and type 2 diabetes were mostly due to time-to-insulin following diagnosis. Although % misclassification was higher in type 1 than type 2 diabetes, the absolute numbers re-classified were similar – ˜4,000 in both types.

Conclusion Clinical assignation of diabetes type is imperfect where autoantibodies and c-peptide measurements are not available. However, substantial improvements can be obtained by carefully checking EHR against other available variables.

Acknowledgement Diabetes Medical Informatics and Epidemiology Group

Is preconception anxiety and depression associated with an increased risk for gestational diabetes? An analysis of the born in Bradford cohort

CA Wilson¹, G Santorelli², LM Howard¹, K Ismail³, RM Reynolds⁴, E Simonoff⁵, J Wright²

¹Section of Women's Mental Health, King's College London, London, UK, ²Bradford Institute for Health Research, Bradford, UK, ³Department of Psychological Medicine, King's College London, London, UK, ⁴Centre for Cardiovascular Science, University of Edinburgh, Edinburgh, UK, ⁵Department of Child and Adolescent Psychiatry, King's College London, London, UK

Objectives While there is now a well-established relationship between depression and type 2 diabetes, the relationship between gestational diabetes (GDM) and mental health is less well understood. There is now some emerging evidence for an association between preconception anxiety and depression and subsequent GDM but this has not yet been explored in a UK population. Thus, the objective was to examine this association within the Born in Bradford cohort.

Methods Diagnoses of anxiety and/or depression occurring prior to first pregnancy in 12,239 mothers with 13,539 pregnancies were obtained from linked primary care records. Diagnosis of GDM was from oral glucose tolerance test. Multivariable Poisson regression was performed within a generalised estimating equation framework to account for multiple pregnancy clustering. Multiple imputation by chained equations was used for missing data.

Results Unadjusted relative risk (RR) for GDM following exposure to preconception anxiety and depression was 0.79 (95% CI: 0.66, 0.94). However, this was attenuated on adjustment for maternal age, education, ethnicity and other obstetric complications (adjusted RR 0.97; 95% CI 0.81, 1.17). The main confounder was ethnicity. When the analysis was stratified by ethnicity, adjusted RR in Pakistani women was 0.95 (95% CI 0.73, 1.24) and in White British was 1.08 (95% CI 0.80, 1.45).

Conclusions The ethnic differences in risk exposed in this analysis of a bi-ethnic cohort demand further research in other populations and exploration of potential mechanisms. This may also need to be considered when tailoring interventions to reduce risk of obstetric morbidities such as GDM in specific populations.

SUPPORTING INFORMATION The conference poster for this abstract is available online in the Supporting Information section at the end of this page.

Logistic regression is equivalent to machine learning models to discriminate between type 1 and type 2 diabetes

LA Ferrat¹, JM Dennis¹, KR Owen^2,3, RA Oram¹, AG Jones¹, BM Shields¹, AL Lynam¹

¹Institute of Biomedical and Clinical Science, University of Exeter, Exeter, UK, ²Oxford Centre for Diabetes Endocrinology and Metabolism, University of Oxford, Oxford, UK, ³Oxford NIHR Biomedical Research Centre, University of Oxford, Oxford, UK

Background and objectives The use of machine learning to improve prognostic and diagnostic models has been increasing in medical research. We aimed to compare logistic regression and five optimised machine learning algorithms performance to predict if a patient diagnosed with diabetes has type 1 diabetes (vs type 2 diabetes).

Methods We used three pre-specified predictor variables (age, BMI and GADA islet-autoantibodies) using a UK cohort of adult participants (aged 18–50 years) with clinically diagnosed diabetes recruited from primary and secondary care (n=1,036, 13% type 1 diabetes). Discrimination performance (ROC AUC) and calibration of each model were compared in a separate external test dataset (n=549, 22% type 1 diabetes).

Results Average performance obtained in model training was similar in all models. Discrimination was good for all models (ROC AUC >= 0.94 with 5 fold-cross validation). In external validation, decreases in performance were observed in all models (ROC AUC >= 0.92). Logistic regression was the model with the best ROC AUC performances in external validation. Logistic regression, neural networks and support vector machine demonstrated a good calibration, superior to the other models.

Conclusion Logistic regression as well as optimised machine algorithms were performed to classify patients with type 1 and type 2 diabetes. This study highlights the utility of comparing traditional regression modelling to machine learning, particularly when using a small number of well understood, strong predictor variables.