Multi-modal characterisation of cardiac function and electrophysiology in type 2 diabetes: a UK Biobank cross-sectional study

Study design and population

The UK Biobank study is a multi-centre, prospective cohort study of over half a million adults recruited between the age of 40 and 69, and living in England, Scotland and Wales. It contains socio-demographic, lifestyle, and clinical information recorded at multiple timepoints since recruitment in 2006, and is linked to general practice primary care records and hospital episode statistics (HES).

Following a baseline assessment visit attended by all participants, about 50,000 participants selected at random were recalled for a multi-modal imaging assessment. This visit included a CMR scan performed on a 1.5 Tesla scanner (MAGNETOM Aera, Syngo Platform VD13A, Siemens Healthcare). Full details of the imaging protocol are available in [13]. A 12-lead resting ECG (CardioSoft ECG system, GE) was also recorded on the same day. Our primary analysis is a matched cross-sectional study examining the association between T2DM status and multi-modal cardiac biomarkers. As secondary analyses, we also carry out a case-control study quantifying the relationship between selected biomarkers in participants who did and did not develop CVD during follow-up, in two matched cohorts with and without T2DM.

Our study is reported in line with the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) statement.

Ethical considerations

General ethical approval was granted for UK Biobank studies by the United Kingdom’s National Health Service Research Ethics Service (11/NW/0382). Participants provided written informed consent for their data to be stored and used for research purposes. This study was conducted under UK Biobank Application Number 40161.

Cohort definition

Disease ascertainment

The selection process and cohort sizes are summarised in Figure 1. We defined our cohorts based on the clinical ascertainment of two disease phenotypes, namely T2DM and CVD. Relevant ICD-9 and ICD-10 codes and their corresponding UK Biobank fields are listed in Supplementary Table 1. T2DM was determined by one or more of the following criteria: date of a T2DM ICD code first reported, HES record corresponding to a T2DM ICD code, HbA1c ≥ 48 mmol/mol, response to a patient touchscreen questionnaire [14]. CVD was determined by one or more of the following criteria: date of a CVD ICD code first reported, HES record corresponding to a CVD ICD code, date of myocardial infarction (MI), date of ST-elevated MI, and date of non-ST-elevated MI. The last three data fields are UK Biobank-specific algorithmically defined diagnoses of MI summarising information contained in other fields; cases identified via those fields may overlap with other routes of ascertainment but were included for completeness.

• Download figure
• Open in new tab

Figure 1.

Cohort selection flowchart. CMR: cardiac magnetic resonance imaging, BMI: body mass index, CVD: cardiovascular disease, T2D: type 2 diabetes. Baseline refers to the time of recording of the ECG and CMR images. Follow-up refers to the period between baseline and censoring date (21/09/2021).

Exclusion criteria, exposure and outcomes

We excluded all participants with pre-existing CVD, and censored all participants with CVD diagnosed after 21/09/2021, the date of the latest CVD diagnosis identified at the time of analysis. The median follow-up period between date of assessment and date of first CVD diagnosis is 1.5 years (min 8 days, max 5.9 years). The breakdown of follow-up CVDs in each cohort is presented in Supplementary Figure 1.

The exposure cohort consists of eligible participants with known T2DM at time of assessment. Pair-matching was carried out to select controls with no known T2DM based on age, sex, body mass index (BMI), as well as binary occurrence of CVD during follow-up (i.e. with vs without a CVD diagnosis). We use BMI as a suitable matching variable as it is a commonly used marker of adiposity that modulates the ECG [15]. Euclidian distances were computed between the [age, BMI] vectors of each exposure participant and possible control participants. Each exposure participant was matched by a control participant of the same sex and follow-up CVD status with the lowest [age, BMI] Euclidian distance [16].

We defined outcomes as ECG and CMR biomarker values for all analyses apart from a secondary analysis considering CVD development, where we defined an outcome as a case of CVD during follow-up.

Baseline covariates

Demographic, lifestyle, and clinical characteristics

We included the following demographic, lifestyle, and clinical characteristics: age at assessment, sex, ethnicity, weight, height, smoking status, systolic blood pressure (SBP), diastolic blood pressure (DBP), medication. BMI was calculated using the ratio of weight in kilograms over squared height in meters. Information on medication was self-reported and divided into broad categories and/or using UK Biobank-specific individual medication codes recorded by nurses. We included cholesterol-lowering medication, blood pressure medication, insulin, metformin and rosiglitazone 1mg/metformin 500mg tablet. The following blood biochemistry measurements were included: high-density lipoprotein (HDL) cholesterol, total cholesterol, HbA1c, triglycerides, creatinine, C-reactive protein. These measurements were taken prior to the imaging assessment so we used the most recent one available. Estimated glomerular filtration rate (eGFR) was calculated using serum creatinine according to the CKD-EPI creatinine equation adjusted for age, sex and ethnicity [17]. Creatinine was converted to mg/dL by multiplying the given value in µmol/L by a factor of 0.0113 [18].

ECG and CMR-derived biomarkers

Resting ECG recordings were available in .xml file format. Each file included the raw ECG signal as well as automatically extracted key ECG markers. Wave amplitudes were available for all 12 leads, while other markers were available for lead I only. We used the following markers: ventricular rate, QRS duration, QTc interval, T wave offset, R-wave amplitude, S-wave amplitude, J point amplitude (equivalent to ST segment elevation or depression), and T wave amplitude. Pairwise Pearson correlation coefficients were calculated to assess inter-covariate correlations among T wave and J point amplitudes in different ECG leads (Figure 2). Two representative leads were retained, one limb lead (aVL) and one precordial lead (V3). We computed the Sokolow-Lyon index for left ventricular hypertrophy as the sum of the S amplitude in V1 and the maximum value of R amplitude in V5 or V6 [19].

• Download figure
• Open in new tab

Figure 2.

Matrix of Pearson correlation coefficients for 12-lead ECG amplitude measurements of T wave amplitude (a) and J point amplitude (b). Coefficients were computed on 2702 (a) and 3562 cases (b), respectively. These cases contained no missing data among the features of interest.

Some key CMR-derived biomarkers were directly available in the UK Biobank database. Manual analyses, algorithms and validation of these biomarkers are described in detail in [20,21]. We used the following in our study: left ventricular ejection fraction (LVEF), LV end-diastolic volume, LV end-systolic volume, LV stroke volume, cardiac output, LV myocardial mass, LV global average wall thickness.

Statistical analysis

All statistical analyses were performed using RStudio with R version 4.1.1.

Association of type 2 diabetes with ECG and CMR biomarkers

We used multivariate multiple linear regression to examine the association between T2DM and ECG and CMR biomarkers. A crude, unadjusted model (Model 0) was used to quantify the association of T2DM only with each biomarker. Three additional models were constructed incrementally to account for socio-demographic (Model 1), lifestyle (Model 2), and clinical covariates detailed in the previous section (Model 3; fully adjusted). Pairwise Pearson correlation coefficients were calculated to assess inter-covariate correlation for all potential model covariates. Coefficients >0.3 or <-0.3 determined the exclusion of the following covariates: eGFR, HbA1c, systolic blood pressure, HDL cholesterol, cholesterol-lowering medication (Supplementary Figure 2). Models were fitted using the Ordinary Least Squares method, therefore scaling of variables was not required. Regression coefficients, their 95% confidence intervals, and p-values are reported for each outcome variable and model in the tables.

Type 2 diabetes and occurrence of cardiovascular disease

We compared biomarkers between cases who did and did not develop CVD, in both cohorts with and without T2DM. We assessed the association between biomarkers with significant differences, and CVD development. We built two sets of logistic regression models, one for each cohort, and examined how these associations differ in each cohort. We followed the same covariate processing and sequential adjustment approach as described previously.

Subgroup analyses

We also examined the role of sex, age, BMI and ethnicity by comparing biomarkers in stratified subgroups. We investigated the effect of severity of disease on selected biomarkers by considering HbA1c as a proxy for hyperglycaemia. We built regression models as described previously, this time using HbA1c as a continuous exposure variable, and biomarkers that were associated with T2DM as outcomes. This analysis was performed within the T2DM cohort only, to avoid a dichotomous HbA1c distribution due to lower HbA1c levels in the control cohort.

Cohort description

Both the exposure and the control cohorts (T2DM vs no T2DM) consisted of 1,781 subjects matched by age, sex and BMI. Cohorts were mostly male (n=1133, 63.6%), had a median age of 67 years and a median BMI of 27.8 kg/m², which is considered overweight by the NHS (NHS Digital, 2022) (Table 1). Individuals with T2DM were more likely to be non-white (7.9% vs 2.7%) and to be current or previous smokers (43.5% vs 40.6%). They tended to have lower diastolic blood pressure, total cholesterol, and HDL cholesterol, while HbA1c, triglycerides, C-reactive protein, and eGFR were generally more elevated. Subjects with T2DM were also more likely to be on cholesterol-lowering drugs, anti-hypertensives, insulin and metformin. One patient in the control cohort reported taking insulin and metformin. This may be due to a lack of diabetes diagnosis at the time of visit, or an increase in HbA1c levels after the last measurement, leading to a missed positive classification of disease. Although rare, this case could also be due to the medication being used to treat another condition unrelated to diabetes, such as polycystic ovary syndrome [22].

Type 2 diabetes and ECG biomarkers

Compared to matched controls, ventricular rate in T2DM was higher (+5 bpm, median: 66 IQR: [59–74] vs 61 [55–68] bpm) and T wave offset was earlier (-12 ms, 842 [820–864] vs 854 [834–874] ms) (Table 2). T2DM was strongly associated with both biomarkers in all models (Table 3). This consistency is expected, given the known inverse correlation between T wave offset and ventricular rate. QRS duration was shorter (-2 ms, 86 [79–93] vs 88 [82–96] ms) and QTc interval was longer (+4 ms, 424 [408–440] vs 420 [405–436] ms) in the T2DM cohort (Table 2). There was a statistically significant association of T2DM with QTc interval in all models but the fully adjusted one (Table 3). However, the association of T2DM with QRS duration was significant only in the two last models adjusted for lifestyle and clinical factors (Model 2 and 3, Table 3). This suggests that the importance of this variable may increase relative to other covariates included in the later models. There were statistically significant differences in almost all wave amplitude biomarkers between the cohorts (Table 2). T2DM was associated with a flatter T wave amplitude in V3 in all models; a flatter T wave amplitude and less elevated J-point in aVL in all models but the fully adjusted one, and a less elevated J-point in aVL (Table 3). T2DM was associated with a lower Sokolow-Lyon index in all models but the fully adjusted one (Tables 2 and 3).

Type 2 diabetes and CMR-derived biomarkers

Left ventricular wall thickness was higher in the T2DM cohort (+2 mm, 6.1 (5.6-6.6) vs 5.9 (5.4-6.5) mm) (Table 2), and its association with T2DM was statistically significant in all models (Table 3). Left ventricular stroke volume was lower in the T2DM cohort (-6 ml, 72 (60-85) vs 78 (65-90) ml), as were end-diastolic and end-systolic volumes (Table 2). However, the association of T2DM was significant in all models only with stroke volume (Table 3). No statistically significant differences or associations were observed for LVEF nor cardiac output.

Type 2 diabetes and cardiovascular disease outcomes

T2DM individuals who went on to develop CVD (n=98) compared to those who did not (n=1683) were more likely to be male (70% vs 64%) and more likely to take metformin (57% vs 53%). In this cohort of subjects with diabetes, there were no changes observed in blood pressure, cholesterol, HbA1c, or eGFR between individuals who do and do not develop CVD. In the T2DM cohort, cases who developed CVD during follow-up (n=98) had a longer QRS duration (median: 88 ms, IQR: [82–98] vs 86 [80–94] ms, p=0.03) (Supplementary Table 2). There was a statistically significant association between QRS duration and CVD development in all models but the fully adjusted one (Supplementary Table 3). In subjects without T2DM, cases of CVD (n=98) had a higher Sokolow-Lyon index (23.4 [17.7-27.1] vs 20.1 [16.2-24.7] mm, p=0.005). We found a statistically significant association between the Sokolow-Lyon index and CVD development in all models (Supplementary Table 3). In both cohorts, cases with CVD had a higher left ventricular mass (T2DM: 100 [86–107] vs 91 [76–108] g, p=0.006; no T2DM: 101 [83–118] vs 92 [76–108] g, p<0.001) and a thicker left ventricular wall (T2DM: 6.3 [5.9-6.8] vs 6.1 [5.6-6.6] mm, p=0.006; no T2DM: 6.3 [5.6-6.9] vs 5.9 [5.4-6.4] mm, p<0.001) (Supplementary Table 2). Both variables showed a statistically significant association with CVD in all models (Supplementary Table 3).

Subgroup analyses

Sex, age, and body mass index

Similar trends were observed between biomarkers of subjects with and without T2DM, in females (n=648) and males (n=1133). Compared to those without T2DM, both females and males in the T2DM group exhibited increased ventricular rates, a prolonged QTc interval, an earlier T wave offset, flatter T wave amplitudes in leads V3 and aVL, a lower Sokolow-Lyon index, lower end-diastolic, end-systolic, and stroke volumes, and a thicker left ventricular wall (Supplementary Table 5). As expected, baseline biomarker values are different in males and females. Sex-specific results stratified by age and BMI are available in Supplementary Figures 3 and 4. No notable differences were observed across age groups. In different BMI groups, there was a gradual decrease in Sokolow-Lyon index with increasing BMI in both cohorts, likely explained by increased electrical impedance, and a steady increase in wall thickness with increasing BMI (Supplementary Figure 3e, 3h).

Ethnicity

Across the two cohorts, 141 of 189 (74.6%) of non-white participants had T2DM, compared to white participants of which only 1640 of 3373 (48.6%) had T2DM. In both white and non-white subgroups, those with T2DM showed a significant increase in ventricular rate and an earlier T wave offset (Supplementary Table 6). A significant QRS shortening, QTc prolongation, T wave amplitude flattening, Sokolow-Lyon score reduction, lower end-diastolic, end-systolic, and stroke volumes, and thinner left ventricular wall were only observed in white participants. In contrast, no significant changes were observed in the non-white subgroup for these biomarkers.

Association of HbA1c with ECG and CMR-derived biomarkers

In individuals with T2DM, we found a statistically significant association of higher levels of HbA1c with both a higher ventricular rate and an earlier T wave offset in all models (Supplementary Table 4). As noted earlier, heart rate and T wave offset are strongly correlated, so this consistent association is expected. Higher levels of HbA1c were also associated with a lower T wave amplitude in lead V3 and a lower stroke volume in all but the fully adjusted model. This suggests that, in addition to their association with binary T2DM status, variations in ventricular rate and T wave also have a significant association with changes in the glycaemic spectrum.

Type 2 diabetes is associated with a higher resting heart rate, longer QTc interval and reduced T wave amplitude

T2DM was associated with a higher resting heart rate. High heart rates have been associated to hypertension and metabolic conditions for many decades [23,24]. In T2DM subjects specifically, this is typically a consequence of cardiac autonomic neuropathy (CAN) [7]. CAN causes heart rate to increase in the early stages of disease and gradually decrease back to normal, but with reduced variability and an increased likelihood of arrhythmias [25]. This may explain why heart rates in our T2DM cohort lie mostly within the normal clinical range of 60-100 beats per minute. Most subjects with T2DM were on anti-hypertensive (n=1003, 56.3%) and cholesterol-lowering medication (n=1287, 72.3%), likely due to concomitant hypertension and hyperlipidaemia, or as a preventative measure for CVD [2]. UK Biobank medication categories considered here are likely to include statins and beta-blockers, which are known to reduce heart rate and may have a disproportionate effect in the T2DM cohort [26,27]. Additionally, we found a significant association between HbA1c and resting heart rate (Supplementary Table 4). This supports evidence linking dysregulations in glucose metabolism with increased sympathetic nervous system activity and severity of CAN [28–30].

T2DM was also associated with a longer QTc interval. QTc prolongation is a well-established risk factor for cardiac arrhythmias and is correlated with a higher CAN score in patients with T2DM [11]. In our T2DM cohort, 113 women and 159 men (n=272 of 1781 [15%]) met sex- and age-specific criteria for clinical QTc prolongation [31]. Diabetic cardiomyocytes have a local renin-angiotensin system; when activated, increases in angiotensin II causes a reduction of transient outward potassium current Ito, extending the action potential repolarisation phase and thus prolonging the QT interval [5,32]. In addition, high glucose decreases hERG channel expression, which in turn modulates the delayed rectifier potassium current IKr, another factor responsible for QT prolongation in diabetes [33,34]. Hyperglycaemia has also been linked with a longer QT in clinical studies of type 1 diabetes [35,36]. Our results, however, did not show this association with HbA1c; mechanisms in T2DM may differ.

T2DM subjects exhibited lower T wave amplitudes in leads V3 and aVL. BMI was included both as a matching variable and as a model covariate, to minimise the effect of body composition on electrical impedance and subsequent reductions in ECG amplitude measurements. Thus, our results suggest a plausible direct effect of T2DM on T wave amplitude. Several studies have already demonstrated the presence of T wave abnormalities in T2DM subjects using the Minnesota coding system, indicating inverted T waves [12,37]. Hyperinsulinemic hypoglycaemia, which is common in T2DM, may reduce T wave amplitude [38]. A recent ECG study showed that T wave amplitudes were reduced in individuals with T2DM [39]. However, some subjects had a previous history of myocardial infarction, which itself is likely to affect T wave morphology. A reduced T wave amplitude has also been associated with hypokalaemia [40]. In T2DM, hypokalaemia and impaired insulin secretion have a bidirectional link [41]. Thus, the lower T wave amplitudes observed may reflect underlying changes in potassium, insulin or glucose levels, or the presence of silent ischemia.

Type 2 diabetes is associated with increased left ventricular wall thickness and reduced stroke volume

In the T2DM cohort, the increase in heart rate (HR) was accompanied by a significant decrease in stroke volume (SV), while no change was observed in cardiac output (CO). Results suggest that the reduction in stroke volume in our T2DM cohort is most likely linked to a decreased end-diastolic volume. These results corroborate findings from a previous study examining CMR differences in a smaller cohort of 143 subjects with type 1 and 2 diabetes [8]. We also found that T2DM was associated with a thicker left ventricular wall, independently of blood pressure. This is consistent with previous studies that have demonstrated that patients with diabetes, hypertension, or both, tend to develop a thicker myocardial wall [42–44]. In later stages of disease, left ventricular hypertrophy is common in T2DM patients due to increased myocardial steatosis leading to hypertrophic signalling and concentric remodelling [45]. Coupled with a normal ejection fraction (>50%), both a lower stroke volume and thicker left ventricular wall can be indicators of heart failure with preserved ejection fraction (HFpEF), a frequent and serious complication of T2DM [46,47]. The changes observed may only represent early stages of adverse remodelling; measurements of left ventricular end-diastolic volume in our T2DM cohort fell within the normal sex-specific reference ranges that have been established for healthy adults without diabetes in the UK Biobank [20]. In the T2DM cohort, median wall thickness (6.1 mm) remained well below the clinical threshold for left ventricular hypertrophy (end-diastolic maximal wall thickness ≥ 15.0 mm) [48], and the absolute change compared to the control cohort was subtle (+0.2 mm).

Unexpectedly, the Sokolow-Lyon index was lower in the T2DM cohort, suggesting lower R and S wave amplitudes. The discordance between this index and left ventricular hypertrophy as assessed by CMR is interesting but plausible. In obese patients, QRS voltages are artificially reduced due to increased body fat causing electrical impedance [15]. Despite having similar BMI and left ventricular mass, subjects in different cohorts may have varying chest wall shapes, or additional myocardial fat deposition contributing to non-electrically active left ventricular mass, which may drive the observed difference in Sokolow-Lyon index.

Clinical outcomes: development of cardiovascular disease

When comparing subjects who did and did not develop future CVD in the T2DM cohort, no statistically significant differences were observed in blood pressure, cholesterol, HbA1c, or eGFR. These biomarkers are clinical risk factors used in the calculation of SCORE2-Diabetes, a 10-year T2DM-specific cardiovascular risk prediction score developed by a working group of the European Society of Cardiology [49]. Here, the lack of significant differences may be due to a shorter follow-up time and pathophysiological changes still too subtle to be reflected, or a small sample size (n=98 CVD cases).

Changes in ECG biomarkers between CVD cases and controls differed in the cohorts with and without T2DM (Supplementary Table 2). Surprisingly, we observed no significant changes in heart rate between cases and controls in both cohorts, despite the established association between a higher heart rate and adverse cardiovascular outcomes [50,51]. In the T2DM cohort, 668 subjects (38%) had a heart rate above 70 bpm, a threshold associated with a higher risk of cardiovascular events specifically in T2DM [52]. An increased QRS duration was the only ECG biomarker showing a significant association with the development of cardiovascular disease in the T2DM cohort. In the ACCORD trial, QRS duration was increased in patients with diabetes and incident heart failure (HF) compared to those without HF [53]. Additionally, a longer QRS complex was established as an independent predictor of cardiovascular events in middle- to older-aged men and is associated with all-cause mortality in T2DM [54,55]. In contrast, in the control cohort, the Sokolow-Lyon index was the only ECG biomarker showing a significant association with the development of future CVD. This is in line with previous studies linking left ventricular hypertrophy with adverse cardiovascular events [56–60]. The hypothesis behind the hypertrophic CVD phenotype is strengthened by the strong independent association of increased left ventricular mass and wall thickness with development of CVD, which was present in both cohorts.

Sex, age, body mass index and ethnicity-specific differences

Baseline biomarker values differed in males and females, which is expected [61–63]. The sex-specific subgroup analysis suggests that trends according to T2DM status were consistent across both sexes. The decrease in Sokolow-Lyon index observed across groups of increasing BMI is likely explained by increased electrical impedance due to a larger chest size.

In the subgroup of white ethnic background, trends in biomarker differences aligned with those of the full cohort. However, in non-white participants, we observed fewer significant differences. This lack of significance may be due partly to a smaller sample size (n=189). Indeed, we found a notable imbalance of participants in terms of ethnicity groups, with 1733 (97.3%) and 1640 (92.1%) of all participants with and without T2DM, respectively, being white. This is, however, reasonably representative of the national distribution for different ethnic groups in the UK population at the time of recruitment [64].

Strengths and limitations

This study is the first to provide a concurrent analysis of ECG and CMR-derived biomarkers in T2DM specifically, as opposed to diabetes of unspecified or combined type, and no prior history of CVD. Thanks to the range of data available in the UK Biobank, we were able to characterise this large cohort in depth by capturing demographical and clinical data of the participants studied. These data were included as covariates in our regression analyses, accounting for key confounders and mediators. The UK Biobank’s robustly validated and systematic data recording protocol ensured that data collection bias was minimised [64]. Our study also has several limitations. Firstly, statistical significance does not necessarily equate to clinical relevance. Slight changes in certain biomarkers may not be much larger than natural population variations. This is particularly relevant to subgroup analyses involving smaller sample sizes, which may result in imprecise estimates due to random error. This applies specifically to groups with CVD outcomes (n=98) and non-white ethnicity (n=289, of which n=48 without T2DM). Regardless of sample size, we strived to interpret all statistically significant results cautiously and within the context of existing knowledge established by previously published research. Secondly, the lack of precision on compounds present in broad medication categories hampers our ability to disentangle direct effects of T2DM from the effect of specific medications. Another limitation is the lack of widespread UK Biobank linkage to primary care records. As of 2024, only 45% of the UK Biobank cohort was linked to these records for general research purposes [65]. This directly impacts cohort size and composition, especially for chronic conditions like diabetes which tend to be diagnosed and recorded within a primary care setting as opposed to hospital admissions. Linkage for the entire UK Biobank cohort would increase statistical power and robustness of future population studies. We believe that matching participants by BMI was adequate for previously stated reasons, however we acknowledge that waist-to-hip ratio is a better indicator of general health and mortality [66]. Finally, we recognise that including multiple distinct conditions within the broad umbrella of CVD may dissolve opposing trends in certain markers. This could be tackled by focusing on a single condition, albeit with small samples sizes.