Predictive modeling of preoperative acute heart failure in older adults with hypertension: a dual perspective of SHAP values and interaction analysis

Abstract

Background

In older adults with hypertension, hip fractures accompanied by preoperative acute heart failure significantly elevate surgical risks and adverse outcomes, necessitating timely identification and management to improve patient outcomes.

Research objective

This study aims to enhance the early recognition of acute heart failure in older hypertensive adults prior to hip fracture surgery by developing a predictive model using logistic regression (LR) and machine learning methods, optimizing preoperative assessment and management.

Methods

Employing a retrospective study design, we analyzed hypertensive older adults who underwent hip fracture surgery at Hebei Medical University Third Hospital from January 2018 to December 2022. Predictive models were constructed using LASSO regression and multivariable logistic regression, evaluated via nomogram charts. Five additional machine learning methods were utilized, with variable importance assessed using SHAP values and the impact of key variables evaluated through multivariate correlation analysis and interaction effects.

Results

The study included 1,370 patients. LASSO regression selected 18 key variables, including sex, age, coronary heart disease, pulmonary infection, ventricular arrhythmias, acute myocardial infarction, and anemia. The logistic regression model demonstrated robust performance with an AUC of 0.753. Although other models outperformed it in sensitivity and F1 score, logistic regression’s discriminative ability was significant for clinical decision-making. The Gradient Boosting Machine model, notable for a sensitivity of 95.2%, indicated substantial capability in identifying patients at risk, crucial for reducing missed diagnoses.

Conclusion

We developed and compared efficacy of predictive models using logistic regression and machine learning, interpreting them with SHAP values and analyzing key variable interactions. This offers a scientific basis for assessing preoperative heart failure risk in older adults with hypertension and hip fractures, providing significant guidance for individualized treatment strategies and underscoring the value of applying machine learning in clinical settings.

Introduction

In the elderly population, hip fractures are a significant public health issue due to their high incidence and severe impact on the quality of life of patients [1]. Epidemiological studies have shown that the incidence of hip fractures increases significantly with age, especially in individuals over 65 years old. This population often suffers from multiple chronic conditions, such as osteoporosis, hypertension, and diabetes, all of which can increase the risk of fractures [2, 3]. This population often suffers from multiple chronic conditions, such as osteoporosis, hypertension, and diabetes, all of which can increase the risk of fractures. Hip fractures can cause immediate pain and functional impairment in patients and may lead to a range of long-term health issues, including reduced mobility, decreased quality of life, and the onset of psychological conditions such as depression [4]. Moreover, hip fractures are associated with a higher mortality rate, with studies indicating that the one-year mortality rate after a hip fracture can be as high as 10% [5]. Therefore, preventing hip fractures and improving post-fracture management are crucial for enhancing the quality of life of elderly individuals and reducing the burden on healthcare resources.

In older adults with hypertension, the combination of hip fractures and preoperative acute heart failure presents a complex clinical challenge with significant impacts on treatment and prognosis. Hypertension, as a chronic cardiovascular condition, can cause gradual changes in cardiac structure and function, including left ventricular hypertrophy, diastolic dysfunction, and ultimately, heart failure [6, 7]. If left uncontrolled, hypertension can also result in other serious complications such as coronary artery disease, stroke, and kidney disease, especially prominent in the elderly population [8]. When these patients suffer hip fractures, the acute pain, stress response, and potential surgical interventions significantly increase the cardiac burden, exacerbating existing cardiovascular conditions [9, 10]. Preoperative acute heart failure in these patients is a marker of poor prognosis. First, it indicates that the patient’s cardiac function is severely compromised and unable to effectively handle the additional stress of surgery [11, 12]. It significantly increases perioperative risks, including arrhythmias, cardiac ischemia, or even cardiac arrest, and can lead to a slow postoperative recovery, further complicating the patient’s condition [13]. Additionally, acute heart failure can lead to surgical delays or cancellations, prolonging hospital stays, increasing the risk of hospital-acquired infections, and potentially leading to complications at the fracture site, such as thrombosis and pulmonary embolism [14,15,16]. These factors collectively significantly increase the risk of mortality. Studies have shown that optimizing cardiac status preoperatively, especially in hypertensive patients who already show signs of heart failure, is key to improving surgical safety and outcomes [17]. Thus, for older hypertensive adults with hip fractures, a comprehensive preoperative assessment of cardiac function and implementing targeted treatment measures are essential.

In the preoperative risk assessment of older hypertensive adults with hip fractures, a critical indicator is the accurate measurement of cardiac stress responses. Cardiac biomarkers, such as B-type Natriuretic Peptide (BNP) and N-terminal pro B-type Natriuretic Peptide (NT-proBNP), are widely recognized for their effectiveness in predicting the risk of heart failure, particularly in patients with a history of cardiovascular disease [18]. However, our research has found that most surgeons do not routinely test for these cardiac markers during preoperative assessments, leading to missed cases of patients who have already developed acute heart failure before surgery. This oversight increases the risk of surgical and postoperative complications, negatively impacting patient recovery and long-term prognosis.

To bridge this crucial gap, our study introduces an innovative predictive model that leverages various machine learning techniques to enhance the precision and reliability of preoperative acute heart failure assessments in this patient group. Machine learning methods exhibit significant advantages over traditional statistical approaches in handling complex medical data. These methods can automatically identify patterns and correlations in large medical datasets, especially excelling in dealing with nonlinear relationships, high-dimensional data, and large volumes of unlabeled data [19, 20]. Additionally, machine learning provides more precise predictive models and enables personalized risk assessments [21,22,23].

Although machine learning has been successfully applied in various medical fields, research on its use for preoperative risk assessment of acute heart failure in older hypertensive adults with hip fractures is still relatively scarce. Therefore, we have developed an innovative predictive model using different machine learning techniques to improve the assessment of preoperative acute heart failure risk in older hypertensive adults with hip fractures, enhancing the accuracy and reliability of predictions. By utilizing SHAP (SHapley Additive exPlanations) values, we can identify and explain which variables in the model significantly impact the predictions, further enhancing the model’s interpretability and transparency.

Additionally, through multivariate correlation analysis and interaction effect analysis, we assessed the impact of key variables on the risk of preoperative acute heart failure. To facilitate the clinical application of this predictive model, we also developed a web-based calculator tool. This tool allows physicians to quickly input clinical data for a patient and instantly receive an assessment of their risk of developing acute heart failure before surgery, enabling more precise clinical decision-making. This not only improves the efficiency and convenience of assessments but also significantly enhances the personalization of patient management, helping to reduce preoperative risks and improve surgical outcomes.

Materials and methods

Study design and patients

This retrospective study involved in patients who underwent hip surgery at the Department of Geriatric Orthopedics, Hebei Medical University Third Hospital, from January 2018 to December 2022. Inclusion criteria were: (1) aged 65 years or older; (2) diagnosed with a history of hypertension; (3) hip fracture confirmed by radiographic imaging such as X-rays; (4) patients with complete medical records, laboratory test results, and other necessary medical documents. Exclusion criteria included: (1) non-hypertensive patients; (2) patients not meeting the diagnostic criteria for hip fractures; (3) those lacking complete medical records, laboratory test results, or other necessary medical documentation.

Ethical statement

This research involved a retrospective analysis of existing case data, where all patient data collection and analysis were anonymized to ensure privacy protection. Furthermore, this study received approval and support from the Institutional Review Board of Hebei Medical University Third Hospital (Approval No. 2021-087-1).

Disease definition

Hypertension

According to the guidelines of the World Health Organization (WHO) and the International Society of Hypertension (ISH), hypertension is defined as having blood pressure measurements that consistently reach or exceed 140/90 millimeters of mercury (mmHg). Specifically, a diagnosis of hypertension can be made when the systolic pressure (pressure during heart contraction) reaches or exceeds 140mmHg, and the diastolic pressure (pressure during heart relaxation) reaches or exceeds 90mmHg [24]. The diagnosis of hypertension typically requires multiple blood pressure readings at different times to ensure accuracy.

Heart failure

Heart failure is a clinical syndrome defined by the heart’s inability to pump blood effectively due to structural or functional impairments, meeting the body’s needs. Symptoms of heart failure include dyspnea, unusual fatigue, and leg swelling, which typically worsen with activity and alleviate at rest [25]. The diagnosis of heart failure primarily relies on the patient’s clinical symptoms, physical signs, electrocardiograms, echocardiography, and other cardiac function tests. In laboratory tests, the diagnostic threshold for BNP is ≥ 300pg/mL, applicable across all age groups to guide the diagnosis of acute heart failure. For NT-proBNP, the following age-specific criteria are used: <55 years: NT-proBNP > 450pg/mL; ages 55 to 75 years: NT-proBNP > 900pg/mL; >75 years: NT-proBNP > 1800pg/mL [18].

Definition of outcome

The primary outcome of this study is defined as preoperative acute heart failure, which is identified from the time of hospital admission until just before undergoing surgery. Preoperative acute heart failure is diagnosed based on elevated levels of specific cardiac biomarkers, which are critical in assessing heart function under stress conditions. BNP and NT-proBNP are utilized as the principal diagnostic indicators. These biomarkers are known for their sensitivity and specificity in reflecting ventricular dilation and stress, making them reliable tools for this predictive model.

Data collection

In this retrospective study, we collected data from older hypertensive adults who underwent hip fracture surgery in the Department of Geriatric Orthopedics at Hebei Medical University Third Hospital from January 2018 to December 2022. The data collection process relied on the hospital’s electronic medical record system, and we systematically organized and analyzed patients’ preoperative information to explore various factors influencing the risk of preoperative heart failure in older adults with hypertension and hip fractures. The collected data encompassed a wide range, including basic demographic characteristics (such as gender and age), admission dates, and a series of medical conditions that might affect the risk of heart failure, including past cerebrovascular accidents, coronary artery disease, diabetes, chronic obstructive pulmonary disease, cancer, and non-ventricular arrhythmias. New complications developed after admission mainly included pulmonary infections, acute myocardial infarction, acute cerebrovascular events, stress hyperglycemia, stress ulcers, urinary tract infections, etc. Laboratory tests primarily assessed conditions like anemia, electrolyte imbalances (such as hypokalemia, hyponatremia), and hypoalbuminemia. All patients underwent a duplex ultrasound of the lower limbs to assess the status of venous thrombosis in the legs.

Model development

In this study, we employed logistic regression techniques to predict the likelihood of acute heart failure in older adults with hypertension prior to undergoing hip fracture surgery. We divided the collected data into a 70% training set and a 30% validation set, a split designed to optimize model training quality and ensure independent evaluation. To further validate the robustness and reliability of our logistic regression model, we applied 10-fold cross-validation, enabling us to assess the model’s performance more comprehensively across different subsets of data. To identify variables closely associated with acute heart failure, we initially used LASSO regression for variable selection. These selected variables were then utilized to construct a multivariate logistic regression model, ensuring that only statistically significant factors were included in the final model. To enhance the interpretability of the model, we further utilized nomograms to display the specific contributions of each variable to the prediction outcomes.

In this research, we constructed predictive models using five different machine learning methods, including Random Forest, Support Vector Machine, Adaptive Boosting (AdaBoost), Extreme Gradient Boosting (XGBoost), and Gradient Boosting Machine (GBM). To precisely assess these models’ capabilities with unseen data, we applied five-fold cross-validation. Given the issue of class imbalance in the dataset, we employed the Borderline SMOTE technique for oversampling the minority class samples, helping to balance the sample distribution and enhance model performance.

In terms of model interpretability and evaluation, we not only measured the models’ discriminative abilities by calculating their AUC values but also tested their calibration using the Hosmer-Lemeshow test. Additionally, decision curve analysis was used to assess the net benefits at various thresholds, thoroughly evaluating the models’ effectiveness in real medical decision-making scenarios. We delved deeper into understanding and identifying the significant impacts of variables within the model using SHAP value analysis. SHAP values provide an interpretable perspective for each prediction, explaining how each feature influences the outcome. This method not only increases the transparency of the models but also enhances their interpretability, helping to understand the decision-making process of the models. We also conducted multivariate correlation analysis and analyzed interaction effects among variables. These analyses helped us evaluate how key variables interact with each other and their impact on the risk of preoperative acute heart failure. By identifying these key interrelationships, we can better understand how different factors collectively influence the risk of heart failure.

Statistical analysis

In this study, we explored the associations between various risk factors and acute heart failure in older hypertensive adults with hip fractures. We first conducted descriptive statistical analyses of participants’ basic information. For continuous variables, we used the Kolmogorov-Smirnov test to assess their distribution for normality. This test was chosen because of its applicability to any sample size and its sensitivity to detect deviations from normality compared to other tests such as the Shapiro-Wilk test, which may perform better for small sample sizes but less so for larger datasets like ours. Variables following a normal distribution were described using mean ± standard deviation, while those not following a normal distribution were described using median and interquartile ranges, offering a more robust measure of central tendency and variability in the presence of outliers. For categorical variables, we presented frequencies and percentages. Furthermore, to assess collinearity among variables, we calculated the Variance Inflation Factor (VIF) and tolerance, where a VIF value below 5 and a tolerance greater than 0.1 generally indicate no significant collinearity among variables. This metric helps ensure that our regression models are not unduly influenced by multicollinearity, which can distort the estimated relationships between predictors and the outcome. If the expected frequencies for categorical variables were low, we utilized Fisher’s exact test to ensure accurate significance testing. This test provides a precise method for determining the likelihood of observing a given set of frequencies among categorical variables, suitable especially for small sample sizes. Statistical analyses were performed using SPSS 24.0 software and R language, with all analyses conducted at a statistical significance level of p value < 0.05.

Results

Patient baseline characteristics

Between January 2018 and December 2022, we included a total of 4170 older adults with hip fractures. After screening, 2800 patients were excluded, resulting in 1370 patients being included in the study. The exclusions comprised 1211 non-hypertensive patients, 1127 patients without hip fractures, 328 non-surgical patients, and 134 cases with incomplete data (Fig. 1).

Fig. 1

The patient flow chart in our study

Table 1 presents the baseline clinical characteristics of older hypertensive adults with hip fractures, divided into acute heart failure and non-acute heart failure groups. Overall, the average age was 79.0 ± 7.4, with 382 males (27.9%) and 988 females (72.1%). Of these, 470 patients (34.3%) experienced acute heart failure preoperatively. There were significant differences between the groups in terms of gender distribution and age categories (< 75 years and ≥ 75 years). The prevalence of coronary artery disease and chronic obstructive pulmonary disease was significantly higher in the acute heart failure group (p < 0.05). Additionally, preoperative complications such as pulmonary infections, ventricular arrhythmias, and acute myocardial infarction were also more prevalent in the acute heart failure group, showing statistically significant differences (p < 0.05).

Table 1 Baseline clinical characteristics of older hypertensive adults classified by acute heart failure

Univariate analysis of laboratory data and ultrasound examinations

Table 2 shows the characteristics of preoperative laboratory tests and lower limb venous ultrasound examinations for older hypertensive adults with hip fractures. The incidences of anemia, hypokalemia, and hyponatremia were significantly higher in the acute heart failure group, with notable statistical differences (p < 0.05). However, there were no significant differences between the groups in the incidence of lower limb venous thrombosis.

Table 2 Laboratory and ultrasound examination results of older hypertensive adults classified by acute heart failure

Development and validation of the nomogram

We randomly assigned 1,370 patients to the training set (962 people) and the testing set (408 people) based on a 70/30 split using R language. Initially, we used the LASSO regression method to select variables from the training set, retaining 18 out of 21 variables (Fig. 2A and B). Further multivariable logistic regression analysis identified gender, age, coronary artery disease, pulmonary infection, ventricular arrhythmias, acute myocardial infarction, and anemia as independent risk factors for acute heart failure before surgery in older hypertensive adults (Table 3; Fig. 3A). Based on this, we have developed a nomogram model designed to predict the risk of acute heart failure in these patients preoperatively (Fig. 3B). Additionally, this tool can be accessed and utilized online at https://longmao.shinyapps.io/dynnomapp/ (Fig. 3C). The prediction formula for the model is Logit(P) = − 1.820 − 0.894 × Sex + 0.649 × Age + 0.577 × Coronary heart disease + 1.036 × Pulmonary infection + 0.688 × Ventricular arrhythmia + 2.412 × Acute myocardial infarction + 1.125 × Anemia. Additionally, we confirmed the independence of variables within the model by calculating the Variance Inflation Factor (VIF), finding all VIF values significantly below the common threshold of 5, indicating low multicollinearity and demonstrating the model’s stability and reliability. Specific VIF values included: Gender 1.02, Age 1.01, Coronary heart disease 1.01, Pulmonary infection 1.02, Ventricular arrhythmia 1.02, Acute myocardial infarction 1.02, Anemia 1.01.

Fig. 2

Data statistics and clinical feature selection using the LASSO binary logistic regression model. (A) LASSO coefficient profiles of the 21 features. A coefficient profile plot was produced against the log(lambda) sequence. (B) Optimal parameter (lambda) selection in the LASSO model used fivefold cross-validation via minimum criteria. The partial likelihood deviance (binomial deviance) curve was plotted versus log(lambda)

Fig. 3

A. Forest plot showing the relationship between risk factors and the occurrence of preoperative acute heart failure in geriatric hypertensive patients with hip fracture. B. A nomogram designed to assess the risk of preoperative acute heart failure in older adults with hypertension prior to hip fracture surgery. The nomogram was developed using a training set and includes parameters such as age, gender, coronary heart disease, pulmonary infection, ventricular arrhythmia, acute myocardial infarction, and anemia. C. An online dynamic nomogram accessible at https://longmao.shinyapps.io/dynnomapp/, depicting an example for predicting the probability of preoperative acute heart failure in older adults with hypertension prior to hip fracture surgery for a female, aged 75 years or older, with comorbid pulmonary infection and acute myocardial infarction

Table 3 Prediction factors of preoperative acute heart failure in geriatric hypertensive patients with hip fracture

By utilizing the Bootstrap resampling method to assess the nomogram 1000 times, we observed that the model’s calibration curve deviates only slightly from the ideal linear relationship, indicating a high consistency between model predictions and observed outcomes (Fig. 4). We constructed a ROC curve and calculated the AUC (Area Under the Receiver Operating Characteristic Curve) for the training set, where the nomogram AUC was 0.745, significantly outperforming other variables (Fig. 5A). The AUCs for the training and testing sets were 0.745 (confidence interval: 0.713–0.777) and 0.744 (confidence interval: 0.694–0.795) respectively (see Fig. 5B), demonstrating good accuracy and stability in internal validation. Additionally, we employed 10-fold cross-validation to further evaluate the model’s robustness and reliability, the AUC from the 10-fold cross-validation was 0.735 (confidence interval: 0.714–0.7693). Additionally, the adjusted C-statistic reached 0.745, further confirming the model’s effectiveness. Decision Curve Analysis (DCA) illustrated the practical application value of our model in predicting acute heart failure following hip fractures in older hypertensive adults. In the training set, the predicted probability range of the model was from 7 to 97% (Fig. 6A), while in the testing set, it ranged from 12 to 90% (Fig. 6B). These data reveal the potential effectiveness of the model in a clinical setting, especially in formulating targeted preventive measures and treatment strategies. The Clinical Impact Curves (CIC) showed the impact of the predictive model on patient numbers at different thresholds (Fig. 6C and D). By applying this model, clinicians can more accurately assess the risk of acute heart failure following hip fractures in older hypertensive adults, thus providing optimized management and treatment options. Integrative use of this predictive model for intervention could significantly enhance patient management efficiency and is expected to promote overall health improvement. This underscores the importance of utilizing precise data-driven decisions in clinical practice.

Fig. 4

Calibration curves of the acute heart failure nomogram prediction in the cohort. Panel A shows the calibration curve for the training dataset. Panel B shows the curve for the test dataset

Fig. 5

Analysis of the ROC curve for the predictive values of preoperative acute heart failure conditions. A presents the ROC curves for various factors that predict acute heart conditions before surgery in older hypertensive adults with hip fractures. The lines represent the predictive accuracy of factors like age, sex, and specific medical conditions, with the overall nomogram achieving an AUC of 0.745, showing good prediction power. B compares the nomogram’s predictive performance on training and validation sets. The training set ROC curve has an AUC of 0.745, while the validation set’s curve has a similar AUC of 0.744, indicating the model’s consistent accuracy in predicting acute heart failure across different data samples

Fig. 6

Decision curve analysis (DCA) and Clinical Impact Curves (CIC) for the acute heart failure nomogram. A and B are Decision Curve Analysis graphs for the nomogram, with the net benefit plotted on the y-axis against different threshold probabilities on the x-axis. Panel A represents the DCA for the training set, and Panel B represents the DCA for the testing set. The blue line represents the nomogram’s net benefit compared to treating all or no patients, indicated by the grey lines. The value of the nomogram is in its ability to balance true and false positives when predicting acute heart failure. Figures C and D show the number of patients identified as high-risk by the nomogram (solid purple line) and the actual patients with heart failure events (dashed red line). These graphs help in understanding the nomogram’s effectiveness and its implications for clinical decision-making at various risk thresholds

Development of predictive models using machine learning methods

Before training machine learning models, all raw data underwent preprocessing steps such as data cleaning and transformation to ensure integrity and quality during the training phase. These steps ensured that the data was effectively recognized and analyzed by machine learning algorithms. In the importance assessment, the highest scoring features included acute myocardial infarction, ventricular arrhythmias, pulmonary infection, and anemia (Fig. 7A; Table 4). Additionally, an analysis of correlations among features was conducted (Fig. 7B). These preprocessing and analytical steps are crucial for ensuring the accuracy of model analyses.

Fig. 7

Variable importance and correlation matrix from preprocessed data in machine learning model analysis. A, presents a correlation heatmap where the intensity of the colors represents the strength of the relationship between clinical variables: red for a strong positive relationship, blue for a strong negative relationship, and white for no significant relationship. This provides a visual depiction of how various factors are interrelated. B, illustrates the significance of different predictors in the model, highlighting Acute Myocardial Infarction, Ventricular Arrhythmia, Coronary heart disease, and Anemia as the top factors contributing to the model’s predictive ability

Table 4 The exact data of importance of all the variables

Using various machine learning methods for model evaluation, the AUC values obtained were as follows: AdaBoost 0.714 (0.662–0.766), XGBoost 0.715 (0.661–0.770), GBM 0.749 (0.698–0.799), RF 0.710 (0.656–0.763), and SVM 0.693 (0.639–0.747), with GBM demonstrating the highest AUC in the models (Fig. 8). Accuracy, sensitivity, precision, and F1 scores were also calculated for LR and the various models. It was found that, in terms of accuracy, GBM and LR performed similarly, both at 73.5% and 73.1% respectively, while SVM had the lowest accuracy at 68.9%. In terms of sensitivity, the GBM model excelled, reaching 95.2%, significantly higher than the other models, with the LR model scoring the lowest at 79.7%. For precision, XGBoost and GBM were relatively close, at 76.0% and 73.2% respectively. The F1 score, reflecting a balance between precision and sensitivity, showed GBM leading at 82.8%, while LR had the lowest F1 score at 42.1% among the six models (Table 5).

Fig. 8

Receiver Operating Characteristic (ROC) curves for various machine learning models in the evaluation of the dataset. The curves compare the sensitivity (true positive rate) and 1 - specificity (false positive rate) across different thresholds for Random Forest (RF), Support Vector Machine (SVM), AdaBoost, Extreme Gradient Boosting (XGBoost), and Gradient Boosting Machine (GBM). Area Under the Curve (AUC) values are displayed in the legend, with GBM showing the highest AUC of 0.749

Table 5 Comparison of parameters in the acute heart failure model before surgery in older hypertensive adults with hip fracture

In studying risk factors for acute heart failure prior to hip fracture surgery in older hypertensive adults, SHAP values were used to interpret the predictions of the machine learning models. The Feature Importance Plot showed the impact of various features on the predictive model for preoperative acute heart failure in older hypertensive adults with hip fractures (Fig. 9A and B). The bar graph displayed the mean absolute impact of each feature on the prediction outcomes, with anemia, acute myocardial infarction, and pulmonary infection being the most influential features, indicating a close association with the risk of acute heart failure. Chronic obstructive pulmonary disease had the smallest average impact. This visualization clearly displays the relationship between features and the predictive goal. The swarm plot provides the specific distribution of SHAP values for each feature, where each dot represents the SHAP value for an individual prediction instance, and colors range from light purple to yellow, representing feature values from low to high. By observing the distribution of dots, we can see whether the contributions of certain features to the model predictions are positive or negative and the consistency of this impact. For instance, the dots for anemia are predominantly in the positive region, indicating that this feature typically increases the predicted probability of heart failure risk.

Fig. 9

SHAP Value analysis for predictive modeling of preoperative acute heart failure in geriatric hypertensive patients with hip fracture. Plot A illustrates the individual contribution of clinical features to a model predicting acute heart failure. Each point’s position reflects how much that feature changes the risk prediction, with color indicating the feature’s value. Plot B ranks the clinical features by their importance in predicting heart failure, with longer bars representing greater influence on the model’s output

Through the constructed multivariate dependency and interaction effect graphs (Fig. 10A and B), we conducted a comprehensive analysis of the risk factors for acute heart failure before hip fracture surgery in older hypertensive adults. The interactions between features such as anemia, acute myocardial infarction, pulmonary infection, and sex, and their impact on predicting heart failure risk were revealed. In the multivariate dependency graphs, we observed that the SHAP values for anemia increase with higher coronary artery disease values, suggesting a positive association between them. In the analysis of acute myocardial infarction, despite considering chronic obstructive pulmonary disease values, this feature’s contribution to predicting heart failure risk remained significant. The interaction effect graphs provide a quantified display of how each feature impacts the risk of heart failure. For instance, the probability of heart failure significantly increases with acute myocardial infarction compared to when no chronic obstructive pulmonary disease is present, while the presence of chronic obstructive pulmonary disease seems to mitigate this increased risk. The analysis of the gender feature shows that females have a higher predicted probability of heart failure than males, highlighting gender as a non-negligible risk assessment factor. The presence of hypoalbuminemia is also associated with a higher risk of heart failure. These graphical analyses provide robust evidence for assessing the risk of acute heart failure in older hypertensive adults and highlight the complex interactions between features. Through this in-depth analysis, we can more accurately identify which patients are most likely to be affected by these compound risk factors and thus develop more effective prevention and management strategies.

Fig. 10

Multivariate interdependent diagram and interaction effects of preoperative acute heart failure characteristics in older hypertensive adults with hip fracture. A. The scatter plots in Figure A illustrate the impact of medical conditions—like anemia, myocardial infarction, lung infections, and gender—on heart failure risk prediction. Each point shows how strongly each condition influences the risk, with colors from purple to yellow indicating the intensity of the condition. B. Showing interactive effect graphs detailing the probabilities of heart failure based on conditions such as anemia or myocardial infarction. Each bar indicates the likelihood of heart failure when a condition is present or absent, with vertical lines representing the predictive confidence interval for each scenario. These graphs effectively illustrate how the interaction between various health conditions can influence the risk assessment for heart failure

Discussion

In the elderly population, hypertension is a common condition, and prolonged high blood pressure can lead to continuous high cardiac load, potentially causing structural and functional changes in the heart [26]. This often results in left ventricular hypertrophy and diastolic dysfunction among hypertensive patients. When these individuals encounter physical stressors such as fractures, their hearts may struggle to effectively handle the sudden changes in hemodynamic demands, leading to a heightened risk of acute heart failure. Specifically, long-standing hypertension can cause arteriosclerosis and reduced arterial elasticity. The heart must exert more energy to overcome this increased peripheral resistance, adding to the cardiac workload [27, 28]. During acute events like fractures, stress responses increase sympathetic nervous activity, further exacerbating cardiac load. If there is pre-existing cardiac impairment, this sudden increased load can lead to the heart’s inability to maintain normal pumping functions, resulting in acute heart failure [29,30,31]. Moreover, hypertension is closely associated with various cardiac pathologies, including coronary artery disease and myocardial infarction, which can accelerate the development of heart failure if blood pressure is not effectively controlled. Therefore, in clinical practice, close monitoring and management of older hypertensive adults before surgery are crucial to prevent and mitigate the risk of acute heart failure. Our study found that approximately 34.3% of older adults with hypertension and hip fractures experienced acute heart failure preoperatively. Consequently, we developed a comprehensive predictive model, including logistic regression and five machine learning algorithms, specifically to assess the risk of acute heart failure before hip fracture surgery in older hypertensive adults. Through SHAP analysis of feature vectors, we enhanced the understanding of feature importance, elucidated interactions between different features, and significantly increased the model’s transparency and explanatory power. Additionally, we provided an online nomogram tool to help clinicians more precisely identify high-risk patients preoperatively, thereby enabling targeted preventive and therapeutic measures to improve overall patient outcomes (https://longmao.shinyapps.io/dynnomapp/). This approach allows physicians to conduct more detailed risk assessments based on model outputs, optimizing patient management strategies.

In the era of big data, using machine learning models to predict clinical events has become particularly important. For instance, models such as Gradient Boosting Machines (GBM) and XGBoost can effectively predict the risk of acute heart failure during the perioperative period for older adults [32]. These models learn from and identify complex risk patterns in vast amounts of patient data, thus providing clinicians with precise risk assessments to improve patient management and outcomes. This study demonstrates the capability of various machine learning models to predict acute heart failure prior to hip fracture surgery in older hypertensive adults. The logistic regression (LR) model, achieving an AUC value of 0.753, plays a pivotal role in clinical decision-making despite exhibiting lower sensitivity and F1 scores relative to more complex models. Its significant utility is derived from several key attributes: (1) Interpretability: Logistic regression provides a clear and interpretable model, where each variable’s contribution to the prediction outcome is explicitly defined through coefficients. This transparency is crucial in clinical settings where understanding the influence of each predictor is necessary for decision-making. (2) Calibration: The logistic regression model exhibits good calibration capabilities. This means the predicted probabilities of acute heart failure closely align with the actual observed probabilities, which is critical for risk stratification and clinical decision-making. (3) Clinical Utility: In scenarios where clinicians need a quick and straightforward assessment, the simplicity of the logistic regression model can be more practically applied than more complex machine learning models. This model can be used to provide a baseline risk assessment, which can be particularly useful in resource-limited settings. (4) Foundation for Further Analysis: The logistic regression model often serves as a baseline for comparison with more complex models. The performance of logistic regression, though fundamental, distinctly illuminates the necessity of employing more advanced machine learning models. This approach methodically highlights the limitations of logistic regression in managing intricate interactions and voluminous datasets, thereby offering a robust justification for the adoption of these sophisticated models to significantly improve predictive accuracy in complex scenarios. The logistic regression model, therefore, despite its comparative limitations in sensitivity and F1 score, continues to offer substantial discriminative power that is crucial for nuanced clinical decision-making. Its robust performance underlines the importance of having a model that balances simplicity with accuracy, particularly in diverse healthcare settings. The Gradient Boosting Machine (GBM) model reached a sensitivity of 95.2%, a high mark indicating its substantial capability to identify patients likely to experience events, which is particularly crucial in clinical settings to reduce misdiagnoses. Additionally, GBM achieved the highest F1 score at 82.8%, reflecting a good balance between precision and recall. The XGBoost model also displayed strong performance in both AUC and F1 scores, indicating its high level of performance across multiple aspects. The Random Forest (RF) and AdaBoost models showed balanced performance across all metrics, but did not excel in any single metric. The Support Vector Machine (SVM), while performing well in terms of sensitivity, fell slightly short in overall performance assessment. These results provide solid data support for clinicians in selecting appropriate models for real-world clinical decisions, highlighting the potential value of using machine learning approaches from various angles to assist in assessing the risk of acute heart failure. Furthermore, the generalizability of the models assessed via five-fold cross-validation demonstrates their capacity to handle unseen data and reliably perform risk assessments in new clinical situations. This comprehensive approach not only enhances the predictive accuracy of the models but also increases their reliability in practical applications.

Through analysis, we found that acute myocardial infarction is a significant risk factor for acute heart failure prior to hip fracture surgery in older adults with hypertension. The long-term effects of hypertension on blood vessels and cardiac function, such as arteriosclerosis and increased left ventricular load, can decrease the myocardium’s tolerance to hypoxia and stress. In these patients, an acute myocardial infarction can trigger a sudden decline in cardiac pumping function, leading to acute heart failure [33]. Specifically, after an acute myocardial infarction, potential left ventricular dysfunction may cause a decrease in cardiac output, which is particularly severe in older hypertensive adults with already increased cardiac preload. In our study, multivariate dependency graph analysis showed that the impact of acute myocardial infarction on heart failure risk appears to be negative in the presence of chronic obstructive pulmonary disease. This suggests that although we would typically expect the coexistence of acute myocardial infarction and chronic obstructive pulmonary disease to increase the risk of heart failure, several explanations are possible for this negative relationship. Firstly, acute myocardial infarction is a very potent risk factor for heart failure, and when it occurs, it may have already caused extremely severe impacts on cardiac function. Therefore, even though chronic obstructive pulmonary disease adds an additional load to the heart, its contribution to the overall risk of heart failure may not be statistically significant due to the high-risk baseline state caused by acute myocardial infarction. Secondly, patients with chronic obstructive pulmonary disease may receive more aggressive treatment and monitoring after an acute myocardial infarction event. For example, following the acute onset of acute myocardial infarction, doctors might closely monitor the cardiac and respiratory status of chronic obstructive pulmonary disease patients, implementing more aggressive interventions to optimize oxygenation and reduce cardiac load. This enhanced care and treatment could help reduce the incidence of heart failure [34]. Lastly, this negative interaction could reflect underlying biological differences or variances in physiological responses after myocardial infarction in patients with chronic obstructive pulmonary disease, which warrants further research [35]. In clinical practice, these findings underscore the importance of personalized assessment and management in older hypertensive adults with both myocardial infarction and chronic obstructive pulmonary disease. Physicians managing these cases need to carefully balance the interactions of various risk factors and may need to take more proactive cardiac protective measures in patients with chronic obstructive pulmonary disease.

In our study, anemia was relatively common among older hypertensive adults with hip fractures, accounting for 32.3% of the total population and rising to 43.4% among those with acute heart failure. Anemia, characterized by reduced red blood cell and hemoglobin counts, diminishes the capacity for oxygen transport, exacerbating the heart’s oxygen deficiency. This additional strain on cardiac function is particularly pronounced in elderly individuals who already face increased cardiac workload due to long-term hypertension [36]. Research indicates a direct link between preoperative anemia following a fracture and heart failure [37]. After trauma-induced stress, the body’s demand for oxygen and nutrients significantly increases, and the drop in hemoglobin levels in anemic patients means the heart must work harder to maintain sufficient oxygen supply throughout the body. Under these circumstances, the cardiac burden intensifies, especially in older adults with long-standing hypertension, making it easy to trigger or exacerbate heart failure [38]. Our analysis through multivariate dependency graphs indicates a correlation between anemia and coronary artery disease. This finding underscores the importance of considering both factors in preoperative assessments, especially in older hypertensive adults with hip fractures. Anemia and coronary artery disease collectively impact cardiac oxygen supply and load. Anemia reduces the oxygen-carrying capacity of the blood, lowering tissue oxygen supply and forcing the heart to increase pumping efforts to meet the body’s oxygen demands. This is a significant burden on a heart already compromised by coronary artery disease. Coronary artery disease can restrict coronary blood flow, reducing oxygen supply to the heart, further stressing it under the already low oxygen conditions caused by anemia. This dual burden intensifies cardiac workload and can lead to further deterioration of heart function, increasing the risk of acute heart failure [39]. In older adults, this risk is particularly prominent due to their generally lower physiological reserves and poorer tolerance to various stressors. Therefore, comprehensive preoperative risk assessments must take into account both the patient’s anemia and coronary artery disease status. By improving the anemia condition and appropriately managing coronary artery disease, the risks of preoperative and postoperative complications can be significantly reduced, enhancing the safety and success rate of the surgery. This comprehensive assessment and management strategy is crucial for improving the overall prognosis of older hypertensive adults with hip fractures.

In older adults, mobility is often limited following a fracture, leading to a necessity for bed rest or reduced activity, which increases the risk of hypostatic pneumonia. Additionally, post-traumatic stress typically results in a decreased immune function, rendering patients more susceptible to pulmonary infections. Pulmonary infections can cause significant inflammatory responses, leading to the release of systemic inflammatory mediators such as tumor necrosis factor (TNF) and interleukins (ILs). These inflammatory factors can directly affect myocardial cells, weakening myocardial function and potentially reducing cardiac pumping capacity [40]. The respiratory dysfunction caused by pulmonary infections decreases pulmonary oxygenation capacity and increases the cardiac demand for oxygen. Particularly in the context of cardiac structural changes and functional disorders induced by hypertension, the imbalance between cardiac oxygen demand and supply becomes more pronounced. This hypoxic state further exacerbates cardiac burden, promoting the development of acute heart failure [41]. Our study, through multivariate dependency graph analysis, found a correlation between pulmonary infections and ventricular arrhythmias. This finding indicates that pulmonary infections not only have a direct impact on the respiratory system but can also significantly affect the cardiovascular system, especially in older hypertensive adults [42]. In older adults, pulmonary infections often lead to a systemic inflammatory response, which can affect cardiac function through several mechanisms. Firstly, inflammatory factors such as TNF and ILs can directly act on myocardial cells, potentially inducing myocardial damage. Secondly, the hypoxic state caused by pulmonary infections increases the metabolic demands of the heart, which may respond with an excessive stress reaction, including increased and irregular heart rates, all of which are triggers for ventricular arrhythmias [43]. Arrhythmias themselves are risk factors for heart failure, as they can lead to decreased cardiac output and reduced efficiency of blood circulation. This impact is particularly pronounced in older hypertensive adults whose hearts have been under long-term high load, potentially leading quickly to acute heart failure. Therefore, in preoperative assessments of older hypertensive adults with hip fractures, physicians should pay special attention to the diagnosis and management of pulmonary infections and their potential impact on cardiac function. Appropriate anti-infection treatment and maintaining electrolyte balance may help reduce the risk of ventricular arrhythmias in these patients and improve their overall prognosis.

Our study found that among older hypertensive adults with hip fractures, 79.8% of those who experienced preoperative acute heart failure were female. This disparity may be linked to biological, hormonal, and gender-specific differences in cardiac pathology. Firstly, hormonal changes in women, particularly regarding estrogen levels, directly impact cardiac structure and function. Estrogen has a cardioprotective effect, helping to maintain vascular elasticity and reduce arteriosclerosis. However, as women age, particularly post-menopause, this protective effect diminishes, increasing the risk of heart disease, especially heart failure [44]. Additionally, sociopsychological factors may also influence women’s cardiac health, such as lack of social support and increased psychological stress, which may indirectly impact heart health and the risk of heart failure. Therefore, comprehensive management strategies and personalized treatment are crucial when dealing with these patients, taking into account not only physiological factors but also psychological and social factors. Our multivariate dependency graph analysis also revealed a correlation between female gender and hypoproteinemia, showing that these factors together increase the risk of heart failure. Hypoproteinemia, often caused by malnutrition, chronic disease, or acute infection, can occur in older hypertensive adults after a hip fracture, particularly in women, possibly due to generally lower muscle mass and nutritional reserves in females. Combined with pain and reduced activity, elderly women may find it challenging to maintain adequate protein intake, while the demand for protein increases due to the need for tissue repair, further increasing the risk of hypoproteinemia [45]. Once hypoproteinemia occurs, it reduces plasma colloid osmotic pressure, causing fluid to leak from blood vessels into interstitial spaces, increasing cardiac preload, and thereby adding to cardiac burden. Additionally, protein deficiency can weaken the repair and maintenance of cardiac muscle, exacerbating the decline in cardiac function. These findings emphasize that in treating older hypertensive adults with hip fractures, attention must be paid not only to their cardiac and fracture status but also to gender factors and nutritional status, particularly monitoring and managing the risk of hypoproteinemia to reduce the incidence of heart failure. Through multivariate dependency graph analysis, a more comprehensive understanding and prediction of how these risk factors jointly affect the patient’s overall health can be achieved, providing more precise clinical treatment and management strategies.

Limitations

This study has several limitations. First, it is a retrospective study, and the data extracted from electronic medical records may be subject to selection bias. Second, the data for this study were sourced from a single-center study and underwent only internal validation, without external validation to assess the model’s generalizability to other hospitals or populations. Third, although the study conducted a comprehensive analysis of important related risk factors, the presence of residual, unmeasured confounding factors cannot be excluded. Fourth, this study primarily focused on the risk of acute heart failure before hip fracture surgery in older hypertensive adults and did not address postoperative acute heart failure. This indicates a need for recalibration to enhance specificity in future model iterations. Therefore, it is necessary to further explore the risk factors throughout the perioperative period for these patients to evaluate the various factors that may cause acute heart failure. Fifth, the Clinical Impact Curves demonstrate that at risk thresholds below 0.4, our model tends to overestimate the risk of acute heart failure, leading to a high number of false positives. Despite these clear limitations, this study still provides valuable insights for future research. Future studies could improve the accuracy and reliability of the model predictions by expanding the sample size, including more potential risk factors, and implementing long-term follow-up.

Conclusion

This study employed logistic regression (LR) and five different machine learning methods (including random forest, support vector machines, and gradient boosting machines) to construct predictive models for acute heart failure prior to surgery in older hypertensive adults with hip fractures. The LR model used important medical indicators such as gender, age, coronary artery disease, pulmonary infection, ventricular arrhythmias, acute myocardial infarction, and anemia as predictor variables and provided an intuitive risk assessment tool through the construction of a Nomogram. Additionally, LR was compared with five machine learning models to evaluate each model’s optimal performance across multiple assessment indicators for different clinical situations. Further analysis through SHAP values revealed the importance of each variable and their interactions, enhancing the model’s interpretability. This analysis not only demonstrated the contribution of each factor to the risk of acute heart failure but also clarified the complex interactions between variables, providing a scientific basis for clinicians to assess high-risk patients. The development and validation of these models offer early prediction tools that are significant for devising prevention strategies, optimizing patient management, and reducing the incidence of acute heart failure. Moreover, the results of this study underscore the potential of applying machine learning technology in real-world clinical decision-making, advancing the practice of precision medicine in high-risk surgical patient populations.