A prediction study on the occurrence risk of heart disease in older hypertensive patients based on machine learning | BMC Geriatrics

In recent years, machine learning models have been extensively applied to predict the short-term or long-term risk of various diseases, the risk of complications, analysis of risk factors, and mortality analysis [11,12,13,14,15]. This demonstrates the advantages of machine learning models in predicting various diseases. Particularly, there is an increasing application of these models in predicting the risk of heart disease, which is crucial for primary prevention. They play a foundational role in heart disease risk prevention and clinical decision-making. Numerous heart disease risk prediction models based on machine learning have shown good performance, but the focus of each study varies and each has its own limitations [16, 17]. Multiple studies on heart disease risk prediction models indicate that hypertension is one of the significant factors in predicting the risk of heart disease and cardiovascular events [18,19,20]. However, there is a scarcity of machine learning prediction models specifically targeting the risk of heart disease in hypertensive patients, which is an aspect worthy of attention. While some prediction models in this domain have demonstrated promising outcomes, they are predominantly based on cross-sectional studies. Consequently, these models are confined to short-term risk prediction for current heart disease occurrence in hypertensive patients, lacking the longitudinal depth necessary for long-term risk forecasting and thereby presenting significant limitations [17, 21]. Our study, conducted over a 7-year follow-up period, specifically targeted hypertensive patients, thereby contributing to the prediction of long-term cardiovascular disease risk and providing a basis for early risk identification, thus offering distinct advantages. Moreover, while many current machine learning-based prediction models primarily focus on cardiovascular diseases such as atherosclerosis and coronary heart disease, they often overlook other types of cardiac conditions (such as heart failure, rheumatic heart disease, etc.) and their complications. However, it is evident that hypertension is not solely a significant risk factor for coronary heart disease. Therefore, our study simultaneously addresses various cardiac diseases, including coronary heart disease, heart failure, rheumatic heart disease, among others, resulting in a broader scope of prediction. Finally, our focus is on the older population. In an era marked by the continuous aging of society, the application of machine learning in geriatric diseases is becoming increasingly widespread [22]. However, there is a notable scarcity of research on the prediction of long-term cardiac risk in older hypertensive patients, particularly those in underserved communities. This study addresses precisely these focal issues.

Dyslipidemia has become a prevalent condition in the elderly, posing a significant risk for the occurrence and progression of cardiovascular diseases. Elevated levels of total cholesterol (TC), triglycerides (TG), and low-density lipoprotein cholesterol (LDL-C) are characteristic features of these diseases [2, 23]. The predictive results of our study suggest that hypertensive elderly patients with dyslipidemia have a greater risk of developing heart disease. Factors such as hypertension, diabetes, and dyslipidemia can mediate abnormal platelet activation, favoring pathological thrombus formation and cardiovascular diseases, making them common risk factors for the development of cardiovascular diseases [24]. Moreover, lifestyle factors and dietary habits in hypertensive patients often resemble those in individuals with dyslipidemia [25]. Recent research indicates that triglycerides are the main lipid indicator most likely to increase systolic and diastolic blood pressure, with a particularly strong correlation between elevated triglycerides in small high-density lipoprotein (HDL) and increased blood pressure [26]. As blood pressure, lipid, and glucose levels are often latent, vascular diseases, myocardial infarctions, strokes, and other severe events have likely occurred by the time they are detected in this population. Therefore, for older hypertensive patients, dyslipidemia is a robust factor for predicting a higher risk of future heart disease. Timely detection and management of lipid levels can help reduce the risk of heart disease occurrence and prevent serious cardiovascular events.

Patients with chronic lung diseases may experience increased pulmonary circulation resistance, leading to pulmonary arterial hypertension and affecting the right heart system. Hypertension itself can contribute to cardiac remodeling, which further promotes the occurrence of diseases such as heart failure. Therefore, for older hypertensive patients, timely prevention, identification, and treatment of chronic lung diseases can help reduce the risk of future heart diseases. Rheumatic heart disease (RHD) is an acquired valvular disease that begins with untreated streptococcal pharyngeal infection, characterized by valve regurgitation and/or stenosis, often associated with complications such as arrhythmias, systemic embolism, infective endocarditis, pulmonary hypertension, heart failure, and death [27]. It remains a significant cause of cardiovascular disease-related deaths in developing countries [28]. Acute rheumatic fever (ARF) and RHD are important determinants of global cardiovascular disease incidence and mortality [27]. Additionally, in rheumatic diseases characterized by severe systemic inflammation, there is often an increased incidence and mortality of atherosclerosis and cardiovascular diseases [29].

Elderly patients with baseline hypertension will experience long-term damage to the target organ, the heart. This makes the occurrence of various heart diseases more likely, not limited to rheumatic heart disease. According to our predictive results, elderly individuals with hypertension who later develop arthritis or rheumatic diseases are more likely to develop heart disease. Therefore, in older hypertensive patients, particularly those with arthritis or rheumatic diseases, it is essential to strengthen prevention, identification, and management to reduce the risk of heart disease occurrence.

In the prediction using the full variable set, the predictive abilities of all three machine learning methods were relatively weak, with the highest AUC being 0.60 for LR. After feature engineering, a smaller set of more valuable variables was selected, and at this point, the performance of both XGBoost and DNN models showed significant improvement, reaching 0.64 and 0.67, respectively. This suggests that feature engineering not only reduces the dimensionality of the predictive data but also helps enhance the predictive performance of the models. It is noteworthy that during the feature selection process, there were inconsistencies between the selected features and the identified risk factors, such as gender, smoking, and drinking. These factors are known risk factors for heart disease but did not enter the feature selection results. Some studies suggest that factors influencing a disease may not necessarily contribute significantly to its prediction [30]. In other words, factors with a substantial impact on the disease may have a minimal contribution to prediction, highlighting an important difference between predictive research and causal inference.

This study also conducted discrimination analysis, calibration analysis, and clinical utility analysis on the predictive models. A good discrimination model should be able to differentiate between individuals with high and low future disease risks, often evaluated using AUC. Calibration reflects the consistency between predicted risk and actual risk. A systematic review of cardiovascular disease prediction models in 2015 found that 63% of studies reported the discrimination of predictive models, but only 36% reported model calibration, leading to variations in the quality of predictive models [31]. Currently, more and more predictive research recommends reporting both discrimination and calibration to achieve a more scientific and objective evaluation. Additionally, decision curve analysis can guide clinical application by determining the appropriate model threshold based on gains [32]. This study found differences in predictive model gains at different threshold ranges.

Through feature selection, this study identified seven important predictive variables: age, waist-to-height ratio, drinking, diabetes, dyslipidemia, lung diseases, and arthritis. Except for age, which is an immutable factor, the other factors can be improved through changes in diet, increased physical exercise, and weight control. Specifically, leveraging epidemiological data can aid in the early identification of heart disease risk in older hypertensive patients. For low-risk individuals, maintaining healthy lifestyle habits is recommended. For high-risk individuals, changing unhealthy habits, undergoing regular check-ups, and seeking medical examination and treatment for severe cases are advised.

The strengths of this study may include the large-scale community cohort survey, which provides representative and cost-effective data. The adoption of a high-risk population strategy, combined with epidemiological data, facilitates early screening and improves intervention efficiency. Moreover, the use of feature engineering techniques helps reduce the dimensionality of predictive variables, select better predictive models, and facilitates practical application. Finally, a comprehensive evaluation of models using discrimination, calibration, and clinical decision curve analysis guides model selection from multiple perspectives. However, the study only included demographic, lifestyle, and disease history characteristics, and the sample size was relatively small, unable to fully demonstrate the advantages of machine learning algorithms in handling big data and multidimensional features. Furthermore, the study only conducted internal validation and did not perform external validation in a more extensive population, limiting the generalizability of the research. Overall, the results of the study provide valuable insights into predicting the future risk of heart disease in the elderly hypertensive population and early risk identification.