Impact of Diabetes Mellitus on Heart Failure Patients: Insights from a Comprehensive Analysis and Machine Learning Model Using the Jordanian Heart Failure Registry

Introduction

Heart failure (HF) is a complex clinical syndrome arising from any structural or functional cardiac disorder, and the structural base for heart failure development is systolic or diastolic myocardial dysfunction. It represents the end stage of various diseases affecting the heart’s components and metabolic cardiovascular activities, mainly targeting the left ventricle.¹ Typical clinical picture of heart failure includes breathlessness, ankle swelling, fatigue, and signs such as elevated jugular venous pressure, pulmonary crackles, and peripheral edema. Heart failure can be distinguished because of ischemic or non-ischemic damage to the heart muscle. The risk factors for ischemic cardiomyopathy include typical risk factors for the development of atherosclerosis, such as diabetes, hypertension, dyslipidemia, nicotine addiction, obesity, and low physical activity. The progression to HF involves ventricular dilatation and remodeling, leading to decreased cardiac output and increased intracardiac pressures, affecting more than 26 million people globally. The risk factors for HF, such as hypertension, coronary artery disease, obesity, and type 2 diabetes mellitus (DM), either alone or in combination with dyslipidemia, hypertension, and obesity, are particularly pertinent.^2–4 DM precipitate and worsens the course of HF because of the buildup of advanced glycation end products, increased oxidative stress, compromised inflammatory responses, deterioration of intracellular calcium levels, alterations in microRNA expression, as well as the advancement of atherosclerosis and coronary artery disease.⁵ Clinically, studies have demonstrated that patients with concomitant DM and HF have higher mortality.^6,7 Given the alarming rise in global and regional diabetes projections, understanding this relationship is crucial. Notably, prior research has predominantly focused on Western populations, with limited data available on the intersection of HF and DM in Middle Eastern settings, particularly in Jordan. Due to that, this study primarily aimed at comparing the clinical characteristics and outcomes of HF patients with and without diabetes mellitus (DM) in the Jordanian context. This investigation will provide vital insights into how DM influences the presentation and prognosis of HF in this specific population. Additionally, we aim to employ machine learning techniques to predict mortality among these patients, offering a novel approach to evaluating their clinical trajectories. Secondary objectives include analyzing the prevalence and impact of DM on hospitalization rates and lengths of stay among HF patients, evaluating the effect of DM on mortality rates and clinical complications, and assessing the healthcare utilization and economic burdens associated with HF in the context of DM. These objectives are crucial for developing targeted interventions that enhance patient care and manage the public health impact of heart failure more effectively in Jordan.

Methods

Study Design and Setting

Data for this study were obtained from the Jordanian Heart Failure Registry (JoHFR), which includes records of patients with acute and chronic heart failure seen in cardiology clinics and hospitals across Jordan from July 1st, 2021, to February 28th, 2023. This comprehensive registry facilitated longitudinal follow-up at 3, 6, 9, and 12 months to document changes in laboratory results, emergence of complications, and modifications to treatment plans.

Ethical Consideration

The study received ethical approval from each participating center’s Institutional Review Board (IRB) and was registered at clinicaltrials.gov (NCT04829591). Following the ethical guidelines and standards outlined in the Declaration of Helsinki, we hereby confirm that our study fully complies with these principles. The Research Committee of the Faculty of Medicine and the Institutional Review Board at the Specialty Hospital approved the study, and the Institutional Review Board provided the ethical approval. The ethics committee approved a waiver of consent from the patients because the study did not include any therapeutic intervention and the outcomes planned are routinely registered in patients with heart failure.

Data Collection and Variables Measurement

An online form designed to collect patient data was designed for completion by healthcare professionals. The form was structured into 10 sections to capture a range of information, including medical history, heart failure status, laboratory tests and procedures performed by patients, treatment outcomes, mechanical ventilation requirement, hospital length of stay, morbidity, or mortality occurrence. The primary aim of data collection was to explore the outcomes in patients with heart failure overall and then in those with or without DM. Patients are further categorized, compared with and without DM, and further stratified as having heart failure with preserved ejection fraction (HFpEF) or heart failure with reduced ejection fraction (HFrEF). This allowed for a more granular analysis of the effect of DM on patient outcomes across the spectrum of HFpEF and HFrEF. Key variables collected included demographics (sex, age), clinical presentation, and HF characteristics (BMI, current smoking status, alcohol use, symptoms [fatigue and dyspnea], type of HF [HFpEF vs HFrEF]), comorbidities (HTN, dyslipidemia, CKD, atrial fibrillation, ASCVD), laboratory findings, outcomes from treatments (need for mechanical ventilation, number of admissions, length of stay, 30-day mortality), and echo results (left ventricle ejection fraction (LV EF), pulmonary artery systolic pressure [PASP], left ventricular hypertrophy [LVH], and left atrial enlargement [LAE]).

Data Analysis

Initial data entry was conducted using Microsoft Office Excel 2019. The analysis of the data was performed using IBM SPSS version 25 (IBM Corp., Armonk, N.Y., USA). For continuous variables, central tendency and dispersion were assessed using the mean and standard deviation when data followed a normal distribution, as verified by Shapiro–Wilk tests. For non-normally distributed data, the median and interquartile ranges were used. Continuous variables were compared using t-tests, and categorical variables were analyzed using chi-square tests. A p-value of less than 0.05 was considered statistically significant. Missing data were handled by multiple imputations using the “mice” package in R. To ensure a robust analysis in the presence of potential data gaps, we created five imputed datasets using Predictive Mean Matching (PMM).

Machine Learning Analysis

Machine learning was employed to predict mortality in this study, utilizing Python 3.8 for all computational tasks on a MacBook equipped with an M1 processor and 16 GB RAM. For handling missing values, median imputation was used for numerical columns and the most frequent value for categorical columns, with data manipulation facilitated by the Pandas library, version 1.2.4. To ensure comparability of predictive features, all numerical data were standardized using the StandardScaler from the Scikit-learn library, version 0.24.2. Feature selection was conducted using Recursive Feature Elimination with Cross-Validation (RFECV) with a RandomForestClassifier, focusing on isolating the most relevant features for mortality prediction. The dataset was divided into training, validation, and test sets, with 40% reserved for validation and testing. For handling imbalanced datasets, the Synthetic Minority Over-sampling Technique and Edited Nearest Neighbors (SMOTEENN) were applied using the Imbalanced-learn library, version 0.8.0.⁸ A grid search with five-fold cross-validation on the resampled training data was conducted to identify the optimal set of hyperparameters, after which the final model was trained. We compared the predictive capabilities of four different machine learning models: Random Forest Classifier (RFC),⁹ Logistic Regression (LR),¹⁰ Support Vector Machine (SVM),¹¹ and eXtreme Gradient Boosting (XGBoost).¹² These models were evaluated on their performance metrics such as accuracy, specificity, sensitivity, and Area Under the ROC Curve (AUC), employing numerical operations facilitated by the Numpy library, version 1.20.3. Permutation Feature Importance was utilized to assess the impact of each feature on our model’s predictions by measuring changes in accuracy when feature values were randomly shuffled. This technique, integrated through Scikit-learn’s feature importance functionality, provides an intuitive means to understand the relevance of each feature within our model’s predictive framework, ensuring a comprehensive evaluation of the data’s influence on patient outcomes.

Results

Clinical Characteristics and Laboratory Variables

This study analyzed the clinical characteristics and laboratory variables of heart failure patients, where the registry included 2151 patients, of whom 2007 met the inclusion criteria, including 1388 patients with diabetes mellitus (DM) and 619 patients without DM (No DM). This comparison revealed statistically significant differences in several clinical characteristics, Table 1. Notably, the prevalence of DM was associated with an increased incidence of males (p < 0.001), patients in the age group ≥70 years (p < 0.001), and those with a BMI categorization as obese (p < 0.001). The heart failure type was also significantly associated with DM, with a higher proportion of HFpEF observed in the DM group (p = 0.006). Mechanical ventilation and the number of hospital admissions were not significantly different between the groups, indicating similar acute care needs, irrespective of DM status. Laboratory findings highlighted the impact of DM on lipid profiles, with DM patients showing a higher incidence of abnormal cholesterol and LDL levels as detailed in Table 2 (p = 0.008 and p = 0.003, respectively). Hemoglobin levels were also notably different, with DM patients more frequently presenting with levels below the defined normal range (p < 0.001). Kidney function test results were significant for the Estimated Glomerular Filtration Rate (eGFR), showing more advanced stages of renal impairment in the DM group (p = 0.006).

Table 1 Clinical Characteristics of Diabetic and Non-Diabetic Patients with Heart Failure

Table 2 Laboratory Variables of Diabetes and Non-Diabetes Among Patients with Heart Failure

Comparative Analysis by Diabetes Status and Ejection Fraction

When categorized by diabetes status and ejection fraction, the study identified significant differences in sex distribution, age, and prevalence of hypertension among the groups as presented in Table 3 (p = 0.011, p = 0.002, and p = 0.020, respectively). There was a notable difference in the prevalence of atrial fibrillation between the No DM with HFrEF group and the DM with HFrEF group (p = 0.040). However, no significant differences were observed in the history of implanted devices or structural heart disease across different categories. Additionally, mortality varied significantly between the DM with HFrEF and the No DM with HFrEF groups (p = 0.044).

Table 3 Comparative Analysis of Clinical Characteristics and Outcomes Among Patients with Heart Failure Categorized by Diabetes Mellitus Status and Ejection Fraction

Mortality Prediction Models

The performance metrics of the mortality prediction algorithms are summarized in Table 4. The Random Forest Classifier emerged as the top performer with the highest accuracy of 90.02% and an AUC of 80.51%, suggesting it is the most effective model for predicting mortality among heart failure patients (Figure 1). Despite its lower sensitivity at 32.56%, its specificity of 96.39% indicates excellent capability in correctly identifying patients who will not experience the event (mortality), making it highly reliable in negative predictions. The Logistic Regression model, known for its interpretability, also demonstrated good performance with a sensitivity of 72.09%, the highest among the models, which makes it particularly useful in identifying high-risk patients correctly. It has a moderate specificity of 73.97% and an AUC of 79.15%, indicating decent overall performance. The Support Vector Machine showed a balanced profile with an accuracy of 80.74% and a relatively lower AUC of 73.65% compared to other models. It has a sensitivity of 46.51% and a specificity of 84.54%, positioning it as a solid choice for predicting mortality with reasonable confidence in identifying true negatives. Lastly, the eXtreme Gradient Boosting model matched the Random Forest in accuracy at 90.02% and presented an AUC of 78.21%. It has a sensitivity of 39.53% and a very high specificity of 95.62%, like the Random Forest, highlighting its strength in specificity but lower performance in sensitivity.

Table 4 Performance Metrics of Proposed Algorithms for Mortality Prediction

Figure 1 The dashed line represents the chance discrimination level. Each model’s AUC was noted, with the Random Forest Classifier and Logistic Regression outperforming the SVM and XGBoost models marginally.

Permutation Feature Importance

Permutation feature importance analysis as detailed in Figure 2, identified creatinine >115 µmol/L as the variable with the highest mean decrease in model performance when omitted, underscoring its importance in mortality prediction in heart failure patients. The length of hospital stays and requirement for mechanical ventilation were also identified as significant predictors, followed by chronic kidney disease, dyslipidemia, sex (male), and Blood Urea Nitrogen >20 mg/dL.

Figure 2 The error bars represent the standard deviation of the permutation importance over multiple shuffles.

Discussion

Our comprehensive analysis of 2007 heart failure (HF) patients, distinguishing between those with diabetes mellitus (DM) and those without DM, provides critical insights into the clinical and laboratory characteristics that define these cohorts. This detailed exploration reinforces the complex relationship between DM and HF. The prevalence of DM in our HF cohort underscores a notable intersection, aligning with the findings of previous studies that suggest a significant overlap between these conditions.^13,14 The demographic findings from our study—highlighting an increased incidence of HF among males, older individuals, and those classified as obese within the DM cohort mirror global trends and emphasize the multifaceted risk factors contributing to HF in the DM population.^15,16 These trends are critical for clinicians to consider, as they suggest that targeted screening and intervention strategies could significantly benefit these high-risk groups. Targeted screening involves systematically identifying individuals based on specific risk factors, such as diabetes, age, and obesity, that predispose them to heart failure. Subsequent interventions are then tailored to address these specific risks, potentially including more aggressive management of diabetes, lifestyle modifications, and closer monitoring for heart failure symptoms, to improve outcomes and prevent disease progression. Our findings on the higher prevalence of HF with preserved ejection fraction (HFpEF) among patients with DM contribute to an evolving narrative in the literature, suggesting distinct pathophysiological mechanisms in HF development among diabetic patients.^17,18 This is further supported by our laboratory findings, where abnormal cholesterol and LDL levels, alongside reduced eGFR, highlight the systemic impact of DM on cardiovascular health. These findings echo the significance of comprehensive cardiovascular risk management in patients with DM, as evidenced by multiple studies.^19,20 The mortality prediction models evaluated in our study, particularly the Random Forest Classifier, demonstrated high accuracy and area under the curve (AUC) metrics, with specificity and sensitivity analyses offering nuanced insights into their clinical utility. The logistic regression model’s 72.09% sensitivity indicates its strength in correctly identifying patients at higher risk of death, a crucial capability for proactive patient management. The specificity of these models, though not explicitly stated, complements this sensitivity by identifying those not at risk and minimizing unnecessary interventions. The predictive power of these models, especially when considering variables like creatinine levels and the requirement for mechanical ventilation, is supported by the broader literature. Studies have consistently highlighted the role of kidney function markers in predicting outcomes in HF patients, with creatinine serving as a pivotal indicator of cardiovascular risk.^21,22 The inclusion of mechanical ventilation requirements reflects the severity of acute exacerbations, aligning with research suggesting its predictive value in HF mortality.²³ The specificity and sensitivity of our mortality prediction models have significant clinical implications. The ability to accurately predict mortality risk in HF patients, particularly those complicated by DM, can guide clinicians in prioritizing interventions for those at greatest risk. This approach not only enhances patient care but also optimizes resource allocation within healthcare systems. The high accuracy and AUC of the Random Forest Classifier suggest it may be best suited for integration into clinical decision-making processes, offering a robust tool for risk stratification and management planning. Further research is needed to refine these models, particularly by expanding the variables included in the analysis to encompass a wider range of clinical and social determinants of health. Future studies should consider integrating socio-economic factors such as education, income, and employment status, genetic markers linked to heart failure and diabetes mellitus, detailed dietary and lifestyle data, inflammatory markers like C-reactive protein, and data on medication adherence. These additions will enhance the models’ capacity to account for the complex interactions between clinical outcomes and broader social and biological factors, potentially improving the predictive accuracy and clinical usefulness of our findings. Our study should be interpreted with caution in the context of several limitations. First, our analysis is based on single-country data from Jordan, which may limit the generalizability of the findings to populations with different demographic and health system characteristics. Secondly, the study utilizes registry data, which, while comprehensive, is prone to potential biases such as missing information and reporting inaccuracies. Furthermore, not all desirable variables were consistently recorded or available for all patients, possibly affecting the robustness of our conclusions. Additionally, the relatively short follow-up duration in some cases may not fully capture long-term outcomes and survival predictors. Another significant limitation is that while HbA1c values were used to confirm the presence of diabetes mellitus, we did not collect or analyze these values to assess the severity of diabetes across all centers, which precluded a detailed examination of the impact of diabetes severity on heart failure outcomes. Lastly, while our models showed promising results within our dataset, their performance has not yet been validated externally, a crucial step to assess their real-world applicability and reliability. Future studies should aim to validate these models in diverse populations and settings, ensuring their applicability across different healthcare systems and patient demographics.

Conclusion

This study reveals notable differences between diabetic and non-diabetic heart failure patients in Jordan, showing poorer outcomes for those with diabetes, particularly higher prevalence of HFpEF and worse renal and lipid profiles. High creatinine levels, longer hospital stays, and mechanical ventilation needs were significant mortality predictors. The utility of our machine learning models, especially the Random Forest Classifier and Logistic Regression, varied. The Random Forest demonstrated high accuracy and specificity, ideal for minimizing false positives, while Logistic Regression, with higher sensitivity, proved better at identifying high-risk patients. These models, although promising in mortality prediction, should be complementarily used with clinical assessments to enhance decision-making in patient care. Further validation and refinement of these models are crucial for improving their accuracy and real-world applicability.