Development and validation of a risk score for hospitalization for heart failure in patients with Type 2 Diabetes Mellitus

Besides coronary heart disease (CHD), diabetes is another major cause for hospital admissions due to heart failure (HF), which contributes to major morbidity and premature mortality in people with diabetes [1]. Subjects with Type 2 diabetes and impaired glucose regulation have 2.8-fold and 1.7-fold risk of developing HF respectively, when compared to individuals with normoglycemia [2]. The Framingham Study [3] and the United Kingdom Prospective Diabetes Study (UKPDS) developed risk scores or engines to predict CHD-related events [4] and stroke [5]. Based on the Hong Kong Diabetes Registry, our group has developed and validated risk scores for predicting end-stage renal disease [6, 7], stroke [8], coronary heart disease [9] and all-cause mortality [10]. These risk equations may enable risk stratification for more effective preventive strategies in Chinese patients with type 2 diabetes. Notwithstanding the importance of HF in type 2 diabetes, the predictors for HF have not been fully explored.

The Hong Kong Diabetes Registry was established in 1995 as a quality assurance and continuous improvement tool with particular focus on risk stratification, treatment to targets and patient empowerment. In the present analysis, we aimed to develop and validate a risk score for predicting HF that needed hospitalization.

In this analysis, we demonstrated that HF that needed hospitalisation can be predicted with good accuracy in type 2 diabetes. Using 7 commonly collected variables: age, sex, BMI, HbA_1c, ACR and blood Hb, the risk score can provide absolute risks of hospitalisation for HF, which depends on whether CHD event/s occurs or not in a 5-year period. The risk score has achieved an excellent discriminatory power of 0.920 for predicting hospitalisation due to HF over a 5-year period in type 2 diabetic patients while maintaining adequate calibration.

The discrimination of a risk score determines the efficacy and thus usefulness of a risk score. Several risk scores have been published for predicting hard CHD events, i.e. nonfatal myocardial infarction and fatal CHD events, in both general and type 2 diabetic populations. The Framingham CHD risk score has an C-statistics of 0.79 for men and 0.83 for women in their original population and 0.63–0.75 for men and 0.66–0.83 for women in other USA populations [20]. Liu et al [21]validated the Framingham risk score in a general Chinese population and the re-calibrated risk score had an aROC of 0.71 for Chinese men and 0.74 for Chinese women. The UKPDS risk engine for hard CHD endpoints did not report any measures of discrimination in the original report [4]. A validation study in a UK type 2 diabetic population by Guzder et al [22] reported that the UKPDS CHD risk engine had a discrimination of 0.67 for C-statistics. Donnan et al [23] developed a risk score for major CHD in type 2 diabetes population with a discrimination score of 0.71 in the original dataset and 0.69 in an independent dataset. Another risk score which predicted cardiovascular events and mortality using clinical predictors and baseline dobutamine stress echocardiography reported C-statistics values of 0.60 and 0.64 respectively [24]. Thus, a discriminatory power of C index 0.85 and aROC 0.92 of the current HF-related hospitalisation risk score indicates an excellent performance in distinguishing between those who would need hospitalisation for HF from their counterparts who would not. Comparing this risk score with other risk scores that were derived from the same cohort, the risk score has better discriminatory power than CHD (aROC: 0.74) [9], stroke (aROC: 0.77) [8] and all-cause mortality (aROC: 0.85) risk scores [10] but poorer discriminatory power than end stage renal disease risk score (aROC > 0.94)[6, 7]. Similarly, the calibration is also adequate, better than the CHD risk score[9]but poorer than the all-cause mortality risk score [10].

Besides the CHD status on follow-up, the risk scores included as few as 6 baseline variables: age, BMI, HbA_1c, ACR, blood Hb and sex. Most of these parameters have been verified to be associated with HF in other studies. HbA_1c has been repeatedly found to be associated with increased risk of HF [1, 25]. Low hematocrit and Hb levels have recently emerged as a major promoter of cardiovascular morbidity and mortality [26]. In agreement with Horwich et al [27] who reported association of anemia with early mortality in patients with advanced HF, we found that blood Hb had a strong independent predictive effect in the present cohort. Low Hb level can cause HF via many mechanisms including tissue hypoxia and left ventricular failure [28]. Several studies have now confirmed the strong risk association between albuminuria and HF in patients with type 2 diabetes [29‐31]. Hockensmith et al [30] reported that baseline albuminuria was independently associated with 5.4 fold increased odds ratio of hospitalizations due to HF (95% CI: 2.3–12.5). In the Angiotensin II Antagonist Losartan Trial which recruited 1513 type 2 diabetic patients with nephropathy, high baseline albuminuria was associated with a 2.70-fold risk for HF compared with patients with low levels of albuminuria and that reduction of ACR at 6 months post treatment including the use of ARB predicted reduced risk of new onset of HF [32]. A study reported that BMI was not a significant predictor for mortality in patients with advanced HF [33]. In our analysis, BMI was also not a significant predictor of HF on univariable analysis but became significant after adjustment for Hb. Close associations between high blood Hb, hyperviscosity syndrome and obesity [34] may mask the true increased risk of high BMI if blood Hb was not adjusted.

The risk score used future CHD events as a "predictor". The future event is unknown at the time of prediction. Nevertheless, the CHD absolute risk can be estimated using the CHD risk score developed from the same cohort[9], using age, sex, current smoking status, duration of diabetes, non high density lipoprotein cholesterol, ACR and eGFR. In the UKPDS, both CHD and stroke risk engines were used in the case mortality risk engine to estimate the absolute risk of case mortality of patients who subsequently developed CHD or stroke event [35]. Similarly, our CHD risk score can also be used in the risk score to estimate the absolute risk of hospitalization for HF.

First, similar to another study [29], we used the definition of hospitalisation for HF rather than HF itself. The possibility underestimating severe HF incidence using this definition should be very small since Hong Kong Government maintains a heavily subsidised health care system and severe HF without being hospitalised are most unlikely. Second, although the registry is not population-based, due to the lack of a comprehensive health insurance policy and integrated primary health care system in Hong Kong, majority of patients in patients with chronic diseases such as diabetes and heart disease are managed in public hospitals. In 2000, Department of Health of Hong Kong conducted a survey and reported that over 90% of patients diagnosed with diabetes were managed in the public health sector [10]. Third, our cohort of type 2 diabetic patients had a wide range of disease duration which may theoretically widen the 95% CI of parameter estimations. On the other hand, it is the heterogeneity that extends usefulness of this risk score from predicting HF hospitalization at diagnosis to predicting HF hospitalisation during long-term care of patients with type 2 diabetes, i.e. the risk score can be used periodically during chronic care management.

To our knowledge, this is the first risk score developed to predict HF events in a type 2 diabetic cohort. Given the rising prevalence of type 2 diabetes [36] and the high risk of morbidity and mortality of HF, especially in patients with diabetes [1], a risk score with such excellent performance, should enable clinicians to identify high risk subjects for intensive monitoring and treatment. Hence, validation of the risk score in other populations will be of public health importance.