Predictive scores for identifying patients with type 2 diabetes mellitus at risk of acute myocardial infarction and sudden cardiac death

Abstract Introduction The present study evaluated the application of incorporating non‐linear J/U‐shaped relationships between mean HbA1c and cholesterol levels into risk scores for predicting acute myocardial infarction (AMI) and non‐AMI‐related sudden cardiac death (SCD) respectively, amongst patients with type 2 diabetes mellitus. Methods This was a territory‐wide cohort study of patients with type 2 diabetes mellitus above the age 40 and free from prior AMI and SCD, with or without prescriptions of anti‐diabetic agents between January 1st, 2009 to December 31st, 2009 at government‐funded hospitals and clinics in Hong Kong. Patients recruited were followed up until 31 December 2019 or their date of death. Risk scores were developed for predicting incident AMI and non‐AMI‐related SCD. The performance of conditional inference survival forest (CISF) model compared to that of random survival forests (RSF) model and multivariate Cox model. Results This study included 261 308 patients (age = 66.0 ± 11.8 years old, male = 47.6%, follow‐up duration = 3552 ± 1201 days, diabetes duration = 4.77 ± 2.29 years). Mean HbA1c and low high‐density lipoprotein‐cholesterol (HDL‐C) were significant predictors of AMI on multivariate Cox regression. Mean HbA1c was linearly associated with AMI, whilst HDL‐C was inversely associated with AMI. Mean HbA1c and total cholesterol were significant multivariate predictors with a J‐shaped relationship with non‐AMI‐related SCD. The AMI and SCD risk scores had an area under the curve (AUC) of 0.666 (95% confidence interval (CI) = [0.662, 0.669]) and 0.677 (95% CI = [0.673, 0.682]), respectively. CISF significantly improves prediction performance of both outcomes compared to RSF and multivariate Cox models. Conclusion A holistic combination of demographic, clinical and laboratory indices can be used for the risk stratification of patients with type 2 diabetes mellitus for AMI and SCD.


| INTRODUC TI ON
Type 2 diabetes mellitus is an increasingly prevalent disease burden across the globe due to ageing and lifestyle westernization, with numbers projected to increase by up to 439 million by 2030. 1 Diabetes mellitus is burdensome to the healthcare system for its chronic course and a multitude of possibly debilitating and lethal complications across different organ systems. Acute myocardial infarction (AMI) and sudden cardiac death (SCD) are major cardiovascular adverse outcomes in patients with type 2 diabetic mellitus. 2,3 Given the potentially lethal and debilitating nature of such cardiovascular adverse outcomes, many risk scores have been developed in hopes of identifying high-risk patients for early intervention and close monitoring. For example, the UKPDS Risk Engine is a type 2 diabetes mellitus-specific risk score based on the United Kingdom Prospective Diabetes Study (UKPDS) for ischaemic heart disease. 4 The Reynolds Risk Score was developed to assess female cardiovascular risk, and the China-PAR project was devised to target the Chinese population specifically. 5,6 However, typically these risk scores involving HbA1c and lipid level predicted for composite outcomes of major cardiovascular adverse outcomes or cardiovascular mortality, which did not account for the difference in pathogenesis and prognosis between acute coronary syndrome and lethal ventricular arrhythmias. Furthermore, recent studies reported that HbA1c and lipid levels, which were often accounted for in these risk scores, have J/U-shaped relationships with adverse outcomes. [7][8][9][10] Therefore, updated risk scores that incorporate these new findings for predictions of specific cardiovascular adverse outcomes were warranted for personalized management.
The present study evaluated the application of incorporating non-linear J/U-shaped relationships between both mean HbA1c and cholesterol levels into risk scores for predicting AMI and non-AMIrelated SCD respectively, amongst type 2 diabetes mellitus patients.
A conditional inference survival forests (CISF) model was used for time-to-event survival data analysis in predicting AMI and non-AMI SCD. 11,12 2 | ME THODS

| Study design
The present study has been approved by The Joint Chinese University

| Data extraction
The primary outcome of the present study, the time to the initial AMI and non-AMI-related SCD episode, is defined as days from 1st January 2009 to the date of initial AMI/ non-AMI-related SCD or the end of the follow-up period (31 December 2019). A SCD episode is defined as an episode of sustained ventricular tachycardia, ventricular fibrillation or non-specific cardiac arrest. This includes episodes that were aborted (sudden cardiac arrest) and episodes that resulted in death. SCD episodes with AMI within a week before or after the SCD episode were considered AMI-related and thus excluded. The number of AMI and non-AMI-related SCD episodes during the follow-up period was extracted as well. Other clinical characteristics, including demographic details (age and sex), diabetes duration, pre-existing comorbidities, anti-diabetic agents, and cardiovascular agents prescribed, and all-cause mortality, were also extracted. The onset of diabetes is determined by fulfilment of the following criteria, whichever is the earliest: 1) earliest record of type 2 diabetes mellitus-related ICD-9 codes; 2) earliest record of HbA1c >6.5%; 3) earliest record of fasting blood glucose (FBG) >7 mmol/L. The following pre-existing comorbidities were identified using ICD-9 codes (Table S1) December 2008 were also calculated.

| Statistical analysis
The annualized rate and mean event frequency were calculated for the primary outcomes. The annualized rate was calculated by dividing the total number of episodes across the cohort by the number of patientyears follow-up. The mean event annual frequency was calculated by averaging the individual mean number of episodes per year throughout follow-up amongst those who experienced the event. Univariate Cox regression was used to identify predictors for incident episodes of both AMI and non-AMI-related SCD. Patients with AMI prior to non-AMI-related SCD were excluded for the SCD analysis. Hazard ratio (HR), 95% confidence interval (CI) and P value were reported for the Cox regression. Univariate predictors with P < 0.10 were entered into a multivariate model. Significant predictors were then selected into predictive scores. The multivariate Cox regression was then repeated with only the significant predictors to obtain the HR for adjustments for the score. For variables with HR between 0.67 and 1.5, a score of 1 was assigned, otherwise a score of 2 was assigned.
To examine the potential incorporation of the J/U-shaped relationship reported between glycaemic/cholesterol profile and   (Figure 1, top and middle panels).

| Acute myocardial infarction prediction
After adjusting for the multivariate HR of the included parameters (Table S4), a score-based system was developed (Appendix S1). On

| Sudden cardiac death prediction
For risk stratification of SCD, 0.822% (n = 2149) patients were excluded because of AMI occurring before the SCD episode, or the SCD was associated with AMI. For this excluded subset of patients, only triglyceride levels were predictive of SCD (Appendix S1  (Figure 2, top and middle panels). Therefore, the cut-offs for mean HbA1c and total cholesterol were adjusted accordingly. The multivariate HR that the marks assigned in the score are shown in Table S7. None of the variables had HRs beyond the ranges of 0.67-

| Machine learning survival analysis
A CISF model was developed to predict AMI and SCD based on the baseline clinical variables. Optimal tree number of CISF model to predict AMI was set as 700 to predict AMI, while the number was set as 600 to predict SCD, based on the fivefold cross-validation parameter selection results as shown in Figure S1. Survival curves to predict AMI and non-AMI-related sudden cardiac death were generated using the CISF model ( Figure S2). Variable importance values and relative importance values of variables to predict AMI and non-AMI-related SCD are presented in Table 3. Creatinine and age were ranked as the most important predictors of AMI, followed by baseline anaemia, mean HbA1c, triglyceride, male sex, hypertension and IHD ( Figure S3, top panel). For non-AMI-related SCD, age and creatinine were the most important predictors, followed by baseline anaemia, mean HbA1c, HF, male sex, total cholesterol, AF, ophthalmological diabetic complication ( Figure S3,

| DISCUSS ION
There are several major findings from the present study: 1) a combination of clinical and laboratory parameters can be used to predict AMI and SCD amongst patients with type 2 diabetes mellitus; 2) J/U-shaped relationships were not presented consistently across On a separate note, the U-shaped relationship between mean HbA1c and SCD may be explained by the increased arrhythmic potential during both persistent hyperglycaemia and hypoglycaemia.
Under chronic hyperglycaemia, persistently increased activation of calcium channels, and increased oxidative stress can induce arrhythmogenesis. [27][28][29][30] By contrast, hypoglycaemia is a well-known trigger for ventricular tachyarrhythmia and is associated with delayed repolarization and altered repolarization gradients. 31 [44][45][46][47][48] However, RSF model has been criticized for the bias due to favouring covariates with many split-points. 49 In our study, the CISF model was used for time-toevent survival data analysis in predicting AMI and non-AMI SCD, 11,12 which were shown to shown superior predictive performance compared to RSF and multivariate Cox models.

| Limitations
Several limitations should be noted for the present study. First of all, given its observational, data-based nature, it is susceptible to under-coding and coding error, with an inability to establish causal relationships. In addition, the large number of patients included in the analysis drove the high statistical significance but low hazard ratio in some predictive parameters. Thus, findings of these parameters may be driven by the statistical power of the analysis and may have limited clinical significance. Furthermore, duration of diabetes was not adjusted for, given the possible competing variable of time from baseline till outcome onset. This is also to avoid interference of inaccuracy in diabetic duration because of a lack of data beyond a decade prior to baseline. Additionally, the effect of medications was not accounted for due to the potential drug-drug interactions and effect upon the laboratory markers, which would greatly complicate the analysis. Finally, data on other cardiovascular health predictors, such as smoking status, alcohol use and family history of cardiac conditions, were unavailable due to limitations of our administrative database of not converting them into structured data for extraction.

| CON CLUS ION
A holistic combination of demographic, clinical and laboratory indices can be used for the risk stratification of patients with type 2 diabetes mellitus against AMI and SCD. Cause-specific analysis should be applied to further examine the relationship between both mean HbA1c and lipid parameters against different cardiovascular adverse outcomes. The application of machine-learning techniques can improve the sensitivity and specificity of risk prediction by identifying the latent interactions between risk variables.

CO N FLI C T S O F I NTE R E S T
None.

AUTH O R S' CO NTR I B UTI O N S
SL, JZ and CG involved in data analysis, data interpretation, statistical analysis, manuscript drafting and critical revision of manuscript.
WTW, ICKW, TL, WKKW and KJ involved in project planning, data acquisition, data interpretation and critical revision of manuscript.
QZ and GT involved in study conception, study supervision, project planning, data interpretation, statistical analysis, manuscript drafting and critical revision of manuscript.

DATA AVA I L A B I L I T Y S TAT E M E N T
The deidentified data set of this study has been deposited in Zenodo (https://zenodo.org/recor d/4382440) in accordance with university's policies.