Predicting depression in patients with heart failure based on a stacking model

Abstract

BACKGROUND

There is a lack of literature discussing the utilization of the stacking ensemble algorithm for predicting depression in patients with heart failure (HF).

AIM

To create a stacking model for predicting depression in patients with HF.

METHODS

This study analyzed data on 1084 HF patients from the National Health and Nutrition Examination Survey database spanning from 2005 to 2018. Through univariate analysis and the use of an artificial neural network algorithm, predictors significantly linked to depression were identified. These predictors were utilized to create a stacking model employing tree-based learners. The performances of both the individual models and the stacking model were assessed by using the test dataset. Furthermore, the SHapley additive exPlanations (SHAP) model was applied to interpret the stacking model.

RESULTS

The models included five predictors. Among these models, the stacking model demonstrated the highest performance, achieving an area under the curve of 0.77 (95%CI: 0.71-0.84), a sensitivity of 0.71, and a specificity of 0.68. The calibration curve supported the reliability of the models, and decision curve analysis confirmed their clinical value. The SHAP plot demonstrated that age had the most significant impact on the stacking model's output.

CONCLUSION

The stacking model demonstrated strong predictive performance. Clinicians can utilize this model to identify high-risk depression patients with HF, thus enabling early provision of psychological interventions.

Key Words: National health and nutrition examination survey; Depression; Heart failure; Stacking ensemble model; Machine learning

Core Tip: In this study, we utilized easily accessible demographic data and laboratory indicators to construct a stacking ensemble model for predicting depression in heart failure patients. We then compared the performance of this stacking model with that of a single machine learning (ML) model and discovered that the stacking model outperformed the single ML model.

INTRODUCTION

Heart failure (HF) represents the final stage of various heart diseases, in which the cardiac output is unable to adequately meet the body's metabolic demands due to structural or functional issues. This leads to a cluster of symptoms resulting from inadequate blood perfusion to organs and tissues[1]. HF has emerged as being a significant health concern worldwide. HF is often accompanied by various complications, with comorbid psychiatric disorders garnering increased attention in recent years[2]. In particular, depression is highly prevalent among individuals with HF, with a fivefold greater incidence than in the general population[3].

Depressive disorder is a complex mood disorder influenced by various factors, and its primary symptom is anhedonia. It often shows a tendency to relapse, with many individuals experiencing relief during remission periods. Some patients may exhibit residual symptoms or develop a chronic condition. Individuals who experience multiple relapses typically have a poor prognosis[4]. Depressive disorder is a major contributor to the global burden of mental health issues and disability. In some European countries, the prevalence of depression in individuals with HF ranges from 10% to 79%, with 10% being asymptomatic and up to 40% experiencing severe functional impairment[5]. The prevalence of depression appears to be linked to the severity of HF[6].

The occurrence of depression in patients with HF is closely linked to physiological and metabolic disorders, mental state imbalances, and social behavioral abnormalities. Factors such as old age, female sex, primarily Caucasian race, low education level, low socioeconomic status, being unmarried, smoking, platelet aggregation, anemia and diabetes are all potential risk factors for depression in HF patients[7-10]. There are numerous variables that impact depression in patients with HF. Some factors are challenging to identify in clinical practice because most patients with HF are more likely to seek treatment in cardiology departments than in psychiatry departments. Therefore, to enhance the applicability of the model, we developed a prediction model using demographic information and common laboratory markers that are readily accessible in clinical practice. Research indicates that the rate of depression among patients with HF ranges from 14% to 40%[11]. Depression significantly impacts the prognosis of individuals with HF, as those with severe depressive symptoms tend to have higher readmission rates, longer hospital stays, and an increased risk of mortality compared to those without such symptoms[12]. Therefore, the identification of the risk factors for depression in HF patients, prompt recognition of depression in this population, and providing timely psychological interventions can positively influence the prognosis of HF patients.

Prediction models play a crucial role in the early screening of diseases and are highly valuable in clinical diagnosis, treatment, and prognosis guidance for patients. Although separate prediction models have been developed for HF and depression[13,14], there is a lack of clinical prediction models for depression in HF patients. Machine learning (ML), which is a branch of artificial intelligence, can analyze and explore large datasets using advanced statistical and probability techniques to create intelligent and effective prediction models, thus showing significant promise in medical research[15]. Ensemble learning, which combines multiple learners, often leads to better generalization performance than the use of a single learner[16]. Current ensemble learning methods can be broadly categorized into serial methods and parallel methods. However, when dealing with a large amount of training data, a more effective strategy is to utilize the learning method, which involves combining data with another learner. Stacking, which is a prominent example of a learning method, leverages the strengths of different algorithms and complements each other, thus resulting in an overall enhancement in model performance[17,18]. To date, there is a lack of literature discussing the utilization of the stacking ensemble algorithm for predicting depression in patients with HF.

This study aimed to develop a stacking model for predicting the probability of depression in HF patients by integrating easily accessible clinical and laboratory data, with the goal of identifying high-risk populations. Additionally, this study compared the performance of a single-layer ML model with that of a stacking model to assess the potential utility of the stacking model in predicting depression in HF patients.

MATERIALS AND METHODS

Data source

The National Health and Nutrition Examination Survey (NHANES) is a widely recognized survey conducted in the United States that aims to provide representative data on the health and nutritional status of the population[19]. The survey utilized a complex multistage random sampling design to ensure that the sample was representative of the entire population. The data for this study were collected from the NHANES database (https://www.cdc.gov/nchs/nhanes/). This study utilized data from seven consecutive survey cycles (2005-2018) of the NHANES database, involving a total of 70190 participants. After excluding 68894 participants with missing HF data or non-HF data and 212 participants with missing Patient Health Questionnaire-9 (PHQ-9) data, the final analysis included 1,084 eligible participants (Figure 1). The NHANES protocol was approved by the Ethics Review Committee of the National Centre for Health Statistics, and written informed consent was obtained from each participant. All of the methods were conducted in strict adherence to the relevant guidelines and regulations.

Figure 1 Flowchart of participant selection. NHANES: National Health and Nutrition Examination Survey; HF: Heart failure; PHQ-9: Patient Health Questionnaire-9.

Participants

It is challenging to definitively diagnose HF due to the limited availability of data on cardiac troponin, N-terminal pro-B-type natriuretic peptide, B-type natriuretic peptide, and cardiac ultrasound imaging in the NHANES database. As a result, HF diagnosis relied on information obtained from personal interviews in a health questionnaire. Individuals were categorized as having HF if they responded affirmatively to the question “Have you ever been told by a doctor or other health professional that you have HF?”. Several published articles have also endorsed the use of this questionnaire method for diagnosing HF in NHANES participants[20-23].

Outcomes

The PHQ-9 is a standardized tool used for screening depression that is based on the criteria outlined in the Diagnostic and Statistical Manual of Mental Disorders, 4^th edition. The PHQ-9 comprises 9 items encompassing areas such as sleep, mood, fatigue, appetite changes, negative emotions, concentration, interest in activities, behavior, and thoughts of self-harm. Each item is rated on a scale of 0 to 3, corresponding to “hardly ever”, “a few days”, “more than half the time”, and “almost every day”. The total score of the PHQ-9 can range from 0 to 27, with a score of 10 or higher indicating the presence of depression, as per the criteria that were used in this study.

Predictors

The study collected data on patient demographics, laboratory parameters, and comorbidities. Demographic information included age, sex, poverty income ratio (PIR), body mass index (BMI), waist circumference, race, marital status, education attainment, smoking status, and alcohol intake status. The PIR represented the household income proportion based on the survey year and state-mandated poverty guidelines. BMI was calculated by dividing weight by the square of height. Smoking status was determined by asking participants “Do you now smoke cigarettes?”. Alcohol intake was categorized as mild, moderate, or heavy. Heavy alcohol consumption was defined as consuming > 3 drinks per day for women or > 4 drinks per day for men. Moderate alcohol use was defined as consuming 2-3 drinks per day for women and 3-4 drinks per day for men. Mild alcohol use was classified as any other level of consumption[24]. The laboratory parameters included white blood cell count, red blood cell distribution width, triglyceride level, total cholesterol (TC) level, high-density lipoprotein cholesterol level, low-density lipoprotein cholesterol (LDL-C) level, lymphocyte count, neutrophil count, and platelet count. Information on comorbidities included hypertension, anemia, diabetes and coronary artery disease. Hypertension is diagnosed by a systolic blood pressure ≥ 140 mmHg and/or a diastolic blood pressure ≥ 90 mmHg[25]. Anemia was defined as a hemoglobin < 120 g/L for men and < 110 g/L for women. Coronary heart disease was identified by asking participants if a doctor or other health care professional had ever told them they had the condition. Diabetes was defined as a positive response to the question “Have you ever been told by a doctor or health professional that you have diabetes or sugar diabetes?”.

Feature screening

To enhance statistical efficiency, this study utilized multiple imputation methods to address missing data. Normally distributed data are expressed as mean ± SD, nonnormally distributed data are expressed as median (interquartile range), and count data are expressed as n (%). The t test was applied to continuous variables with a normal distribution, whereas the Mann-Whitney U test was employed for those without a normal distribution. Categorical variables were analyzed by using the χ² test. After univariate analysis, variables unrelated to depression in patients with HF were eliminated. A significance level of P < 0.05 indicated a potential association with depression in HF patients. Subsequently, the artificial neural network (ANN) algorithm ranked the selected features from univariate factor analysis based on importance. The top five features were used to construct a stacked integrated prediction model for depression in HF patients.

Model development and performance evaluation

To construct and validate the stacking model, the entire dataset was randomly split into a training dataset comprising 70% of the data and a test dataset comprising 30% of the data. Stacking learning can enhance robustness by training multiple learning algorithms simultaneously. In this study, the LightGBM, extreme gradient boosting (XGBoost), and decision tree (DT) models were utilized as base learners in the first layer of the stacking algorithm, whereas Lasso was employed to construct the meta-learner in the second layer. The parameters of the base learner were tuned via a random grid search to construct the optimal stacking model. Fivefold cross-validation of the training dataset was performed during the stacking process to ensure model fitting. The structural diagram of the stacking model framework is shown in Figure 2. The training set was initially divided into 5 parts. A basic model (LightGBM) was trained on 4 parts and used to predict the 5th part. Subsequently, the model was trained on the entire training dataset. Predictions were then made on the test set. This process was repeated for two additional models (XGBoost and DT), thus resulting in two more sets of predictions for both the training and test sets. Finally, the predictions from the training set were utilized as features to create a new integrated model, which was then used to make final predictions on the test set. Receiver operating characteristic (ROC) curves were generated for the prediction models in both the training and validation sets. The area under the curve (AUC), 95%CI, sensitivity, specificity and cutoff values were subsequently calculated to assess the differential diagnostic performance of these parameters in the test dataset. Calibration curves were used to evaluate the consistency between the actual results and the predicted probabilities. Additionally, decision curve analysis (DCA) was employed to determine the clinical utility of the models. Furthermore, to better understand the impact of each variable on the stacking model, the Shapley additive exPlanations (SHAP) plot was utilized to explain the feature importance.

Figure 2 Stacking model framework structure diagram. DT: Decision tree.

RESULTS

Baseline characteristics

This study analyzed a total of 1084 HF patients, with a mean age of 66.86 years. Among the population, 56.73% were males, and 43.27% were females. The results of the study are presented in Table 1. The study demonstrated a depression incidence of 20.66% among HF patients, with a greater occurrence observed in females and young individuals than in nondepressed participants. Depressed participants also exhibited elevated levels of TC and LDL-C compared to their nondepressed counterparts. Moreover, depression in HF patients was associated with an increased incidence of diabetes mellitus. Factors such as PIR, marital status, and education attainment were also found to be correlated with depression in HF patients.

Table 1 Characteristics of study participants with or without depression among heart failure patients, n (%)/mean ± SD/ median (25^th-75^th percentiles).

Variables	Total	Nondepression	Depression	P value
n	1084	860	224
Age (years)	69.00 (60.00-78.00)	70.00 (61.00-78.00)	64.00 (55.25-73.00)	< 0.001
Gender				< 0.001
Female	469 (43.27)	352 (41.93)	117 (52.23)
Male	615 (56.73)	508 (59.07)	107 (47.77)
PIR	1.55 (1.00-2.69)	1.63 (1.06-2.94)	1.23 (0.78-2.05)	< 0.001
BMI (kg/m2)	30.60 (26.45-36.40)	30.40 (26.41-35.80)	32.15 (26.67-37.94)	0.070
Waist circumference (cm)	108.93 ± 17.23	108.58 ± 17.13	110.31 ± 17.60	0.180
Race/ethnicity				0.104
Mexican American	94 (8.67)	69 (8.02)	25 (11.16)
Non-Hispanic Black	298 (27.49)	241 (28.02)	57 (25.45)
Non-Hispanic White	557 (51.38)	451 (52.44)	106 (47.32)
Other Hispanic	76 (7.01)	53 (6.16)	23 (10.27)
Other Race	59 (5.44)	46 (5.35)	13 (5.80)
Marital status				0.002
Married	501 (46.22)	416 (48.37)	85 (37.95)
Separated/widowed/divorced	449 (41.42)	350 (40.70)	99 (44.20)
Never married	90 (8.30)	59 (6.86)	31 (13.84)
Living with partner	44 (4.06)	35 (4.07)	9 (4.02)
Education attainment				< 0.001
Below high school	454 (41.88)	334 (38.84)	120 (53.57)
High school	257 (23.71)	212 (24.65)	45 (20.09)
Above high school	373 (34.41)	314 (36.51)	59 (26.34)
Current smoking status				0.198
Smoking	339 (31.27)	261 (30.35)	78 (34.82)
Nonsmoking	745 (68.73)	599 (69.65)	146 (65.18)
Alcohol intake				0.586
Mild	726 (66.97)	580 (67.44)	146 (65.18)
Moderate	238 (21.96)	185 (21.51)	53 (23.66)
Heavy	120 (11.07)	95 (13.95)	25 (11.16)
Laboratory features
WBC count (1000 cells/μL)	7.57 ± 2.66	7.50 ± 2.73	7.82 ± 2.34	0.119
RDW	13.90 (13.10-14.98)	13.90 (13.10-15.00)	13.80 (13.10-14.80)	0.528
TG (mmol/L)	1.32 (0.93-1.87)	1.30 (0.89-1.86)	1.40 (1.01-1.95)	0.052
TC (mmol/L)	4.40 (3.65-5.20)	4.29 (3.62-5.12)	4.60 (3.88-4.60)	0.002
LDL- C (mmol/L)	2.41 (1.81-3.13)	2.37 (1.78-3.10)	2.60 (1.91-3.30)	0.018
HDL- C (mmol/L)	1.19 (1.01-1.49)	1.19 (1.01-1.47)	1.16 (0.99-1.53)	0.740
L count (1000 cells/μL)	1.97 ± 1.67	1.78 ± 0.06	1.12 ± 0.07	0.497
N count (1000 cells/μL)	4.68 ± 1.82	1.78 ± 0.06	1.95 ± 0.13	0.154
PLT count (1000 cells/μL)	222.32 ± 68.90	220.68 ± 69.49	228.62 ± 66.36	0.125
Hypertension				0.854
No	717 (66.14)	570 (66.28)	147 (65.63)
Yes	367 (33.86)	290 (33.72)	77 (34.38)
Anemic				0.109
No	959 (88.47)	754 (87.67)	205 (91.52)
Yes	125 (11.53)	106 (12.33)	19 (8.48)
Diabetes				0.040
No	579 (53.41)	473 (55.00)	106 (47.32)
Yes	505 (46.59)	387 (45.00)	118 (52.68)
Coronary heart disease				0.788
No	628 (57.93)	500 (58.14)	128 (57.14)
Yes	456 (42.073)	360 (41.86)	96 (42.86)

PIR: Poverty income ratio; BMI: Body mass index; WBC: White blood cell; RDW: Red cell distribution width; TG: Triglyceride; TC: Total cholesterol; HDL-C: High-density lipoprotein cholesterol; LDL-C: Low-density lipoprotein cholesterol; L: Lymphocyte; N: Neutrophil; PLT: Platelet.

Predictor screening

This study identified 8 features with a significance level of P < 0.05 through univariate factor analysis: Age, sex, PIR, marital status, education attainment, TC, LDL-C, and diabetes. Subsequently, the ANN algorithm was employed to identify the most important characteristics of these variables. Features with lower importance were eliminated, and the top five features with the highest importance scores were retained as the optimal subset for further modeling. This process aimed to reduce feature dimensionality, simplify the model, and enhance its generalization ability. The final selected characteristics were age, PTR, TC, education attainment, and LDL-C (Figure 3A).

Figure 3 Feature screening and model performance evaluation. A: Rankings of feature importance; B-E: Receiver operating characteristic curves of the four models for predicting depression among heart failure patients in the training dataset and test dataset; F: Performance evaluation of the four models in the test dataset; G: Calibration curves of the four models in the test dataset; H: The clinical utility of the four models was evaluated via decision curves in the test dataset. PIR: Poverty income ratio; TC: Total cholesterol; LDL-C: Low-density lipoprotein cholesterol; DT: Decision tree.

Construction of the models

Participants were randomly assigned to a training dataset (n = 758) or a test dataset (n = 326) at a ratio of 7:3. The training dataset was utilized to create prediction models for depression in HF patients, whereas the test dataset was employed to assess the performance of the models. The XGBoost model includes parameters such as mtry, min-n, tree-depth, learn-rate, loss-reduction, and sample-prop, whereas the DT model includes tree-depth, min-n, and cost-complexity. Moreover, the LightGBM model parameters consist of tree-depth, trees, learn-rate, mtry, min-n and loss-reduction. During the construction process, basic models set various parameters within specific ranges. By utilizing the random grid search method and 5-fold cross-validation, the parameter settings were continuously adjusted to achieve the best configuration and to ultimately obtain the optimal models. After multiple iterations, the optimal XGBoost model was found to have mtry = 5, min-n = 19, tree-depth = 1, learn-rate = 0.0275, loss-reduction = 0.159, and sample-prop = 0.907. Similarly, the optimal DT model was characterized by tree-depth = 7, min-n = 7, and cost-complexity = 0.00628, whereas the optimal LightGBM model features tree-depth = 1, trees = 398, learn-rate = 0.0330, mtry = 2, min-n = 9 and loss-reduction = 0.514. Following the acquisition of cross-validation training dataset prediction outcomes and the average prediction results of all of the base models in the test dataset, the Lasso meta-learner was trained, and the final prediction was made based on the base model predictions.

Assessment of the models

This study analyzed the ROC curves of the four models in both the training and test sets and determined the AUC and 95%CIs for each (Figure 3B-E). In the training dataset, the stacking model had an AUC of 0.79 (95%CI: 0.75-0.82), the LightGBM model had an AUC of 0.73 (95%CI: 0.69-0.77), the DT model had an AUC of 0.73 (95%CI: 0.66-0.77), and the XGBoost model had an AUC of 0.70 (95%CI: 0.66-0.74). For the test dataset, the AUC of the stacking model was 0.77 (95%CI: 0.71-0.84), the AUC of the LightGBM model was 0.72 (95%CI: 0.65-0.79), the AUC of the DT model was 0.71 (95%CI: 0.64-0.79), and the AUC of the XGBoost model was 0.70 (95%CI: 0.63-0.78). The test dataset was utilized to compare the prediction performance indicators of the four models (Figure 3F). The calibration curves of the four models in the test set are illustrated in Figure 3G. The results indicated that all four models demonstrated strong agreement between the predicted probabilities and actual probabilities, thus enabling accurate prediction of depression risk in patients with HF. DCA demonstrated that all four models yielded substantial net benefits in detecting depression among HF patients in the test dataset, as illustrated in Figure 3H. Furthermore, a table (Table 2) was created to display the AUC, 95%CI, sensitivity, specificity, and cutoff values of the four models in the test dataset. The results indicated that the stacking model significantly outperformed the other base learning models. Furthermore, the stacking model demonstrated superior prediction results compared to the single-based learning model. Consequently, the stacking model was chosen as the final prediction model for this study.

Table 2 Area under the curve values of the models in the test dataset.

Model	AUC	95%CI	Sensitivity	Specificity	Cutoff value
Stacking	0.77	0.71-0.84	0.71	0.68	0.20
LightGBM	0.72	0.65-0.79	0.68	0.67	0.21
DT	0.71	0.64-0.79	0.52	0.85	0.28
XGBoost	0.70	0.63-0.78	0.62	0.66	0.21

DT: Decision tree; AUC: Area under the curve.

Visualization of feature importance

The SHAP algorithm was utilized to determine the importance of each predictor variable in the stacking model's prediction outcomes. The SHAP summary diagram, which is depicted in Figure 4A, provides an overview of the functionality of the model. The predictor variables were ranked based on their importance, with age being the most influential variable, followed by PTR, education attainment, TC, and LDL-C. SHAP values indicated that elevated levels of TC and LDL-C may contribute to depression in patients with HF. Conversely, increasing age, higher education level, and higher PIR were associated with a protective effect against the risk of depression in HF patients. In addition, this study also included the SHAP force plot of the model to aid decision-makers in gaining trust in the model and in understanding how each feature influences the model's decision-making process (Figure 4B). SHAP values indicated the predictive features of individual HF patients and the extent to which each feature contributes to mortality prediction. Yellow elements signify an elevated risk of depression, whereas red elements indicate a decreased risk of depression. The length of the arrow corresponds to the magnitude of the feature's impact on the model's output.

Figure 4 Visualization of feature importance. A: The SHapley additive exPlanations (SHAP) summary plot; B: SHAP force plot. PIR: Poverty income ratio; TC: Total cholesterol; LDL-C: Low-density lipoprotein cholesterol; SHAP: SHapley additive explanation.

DISCUSSION

To develop a predictive tool for identifying depression in HF patients, we implemented a stacking model using data from the NHANES database spanning from 2005 to 2018. The stacking model utilized three ML–algorithms (LightGBM, DT, and XGBoost) as base learners, with Lasso serving as the meta-learner to combine the predictions of the base models. The final stacking model consisted of 5 predictors: Age, PIR, TC, education attainment, and LDL-C. Performance evaluation on a test set demonstrated that the stacking model outperformed the individual ML models, achieving an AUC of 0.77 (95%CI: 0.71-0.84), a sensitivity of 0.71, a specificity of 0.68, and a cutoff value of 0.20. Calibration curve analysis indicated good consistency of the models, whereas DCA demonstrated the clinical applicability of the models. Additionally, SHAP plots were utilized to interpret the stacking model, which demonstrated that age had the most significant impact on the model output, followed by PIR, education attainment, TC, and LDL-C.

The study included a total of 1084 HF patients and reported a prevalence of depression of 20.66%. Univariate analysis demonstrated that variables such as age, sex, PIR, education attainment, marital status, TC, LDL-C, and comorbid diabetes were significantly different between patients with and without depression. The research noted a negative correlation between age and depression in HF patients, which is consistent with previous studies indicating a decrease in depression probability with advancing age[26]. This trend may be attributed to the high competitive pressure and future uncertainty experienced by younger individuals. Furthermore, this study confirmed that women are more susceptible to depression than men are, which aligns with the literature[27]. This susceptibility may be linked to the various physiological changes that women experience throughout their lives, such as the menstrual cycle, pregnancy, childbirth, lactation, and menopause. These changes can impact estrogen and testosterone levels, thus influencing mood and psychological well-being[28]. Furthermore, despite societal advancements, women still encounter unequal treatment and bias in certain spheres, thus potentially leading to feelings of isolation, helplessness, and despair and thereby elevating their risk of depression. Marital status also plays an important role in the onset and progression of depression in individuals. Married individuals exhibit lower levels of depression than individuals who are single, divorced, or widowed. This could be attributed to the emotional and social support that marriage offers, thus aiding in stress reduction and better management of life's adversities. Nonetheless, certain studies have suggested that marriage may not always have a positive impact on mental well-being. Issues within a marriage, such as communication challenges and frequent conflicts, can result in diminished marital satisfaction and potentially contribute to depressive symptoms in individuals[29]. Previous studies have shown a correlation between higher household income and higher education levels and lower rates of depression[30], which is a trend that aligns with our own findings. The reduced prevalence of depression in individuals with greater economic means and educational attainment could be attributed to their enhanced stress-coping mechanisms, heightened self-awareness and self-regulation skills, improved living standards, and elevated social standing. Despite potentially experiencing a fast-paced lifestyle, individuals in this demographic group seem to experience less detrimental effects on mental health than those facing existential stress in lower social strata[31]. Depressive symptoms have been found to be positively correlated with serum levels of TC and LDL-C, even after controlling for factors such as age, sex, and education level[32]. This association may be attributed to the proinflammatory effects of high TC and LDL-C levels in the body. Research indicates that elevated levels of serum inflammatory markers are linked to the development of depression[33,34]. Furthermore, animal studies have demonstrated that chronic HF can exacerbate susceptibility to depression through an associated increase in the inflammatory response[35]. This study also demonstrated a greater likelihood of depression among patients with both HF and diabetes, which is possibly attributed to shared pathophysiological mechanisms such as neurohormonal activation, rhythm disorders, inflammation, and a hypercoagulable state[36,37]. Moreover, individuals with diabetes often require strict dietary regimens and medications, which could contribute to feelings of moodiness and depression.

Depression is a serious condition that can potentially lead to suicide, thus underscoring the importance of early detection and psychological interventions. However, many patients may be hesitant to accurately report their depressive symptoms, which poses challenges in promptly diagnosing depression. It is crucial to recognize that different diseases exhibit unique characteristics of depression. For instance, symptoms of HF may overlap with those of depression, such as fatigue, sleep disturbances, and loss of appetite, thus increasing the likelihood of depression being overlooked[38]. Therefore, it is essential to develop specific predictive models for identifying depression in various patient groups. Compared to individuals without health issues, HF patients are at a greater risk of experiencing depression[11]. Furthermore, HF patients with comorbid depression often experience a reduced quality of life and an elevated risk of recurrent cardiovascular events and mortality[5,39]. The early detection of high-risk groups for depression in HF patients and timely intervention and treatment can significantly improve long-term outcomes for individuals with HF.

We aimed to develop a high-performance model for identifying HF patients at high risk of depression. Initially, we employed the ANN algorithm to determine the importance of features from eight variables that were identified through univariate analysis. The top five variables were selected as predictors, including age, PIR, education attainment, TC, and LDL-C. By recognizing the limitations of using a single algorithm to create a prediction model, such as reduced accuracy and universality, we sought to build a stacking model for improved performance. Stacking is an integration algorithm that merges multiple base models through meta-models, thus forming a multilayer learning system with a parallel structure[17,18]. This approach involves primary and secondary learners, wherein the secondary learner is produced by the primary learner during training. By implementing stacking ensemble learning, we can leverage the strengths of individual models to enhance the predictive capability of the overall model. Although stacking integration strategies have been utilized in the medical field[40-42], there is a gap in research regarding the application of the stacking algorithm for identifying HF patients at high risk of depression. In our study, we selected DT, XGBoost and LightGBM as the first-level base learners, with Lasso serving as the second-level meta-learner. Through a global search method, we traversed all of the possible combinations of base classifier algorithms to train and obtain the stacking ensemble model. We then compared the predictive performance of this model with that of a single model.

A DT is a classification and regression method that operates on a tree structure. Its fundamental concept involves identifying the optimal decision boundary through iterative data division. Beginning from the root node, the data are split based on feature values until specific stopping conditions are met, such as reaching the maximum depth or having a node with a sample count below a predefined threshold[43]. XGBoost, which is an ensemble learning algorithm, is a rooted gradient boosting DT, which is typically a classification and regression tree; afterwards, these weak classifiers are amalgamated to create a robust classifier[44]. Similarly, LightGBM is a comprehensive learning algorithm that also relies on gradient boosting DT. It employs a histogram-based approach to expedite training and implements a leaf node optimization strategy to minimize memory usage[45]. Each of the three individual models has its own set of advantages and disadvantages, with the stacking model combining the strengths of these models. This research evaluated the predictive performance of the four models by using data from the test dataset. The stacking model achieved an AUC of 0.77 (95%CI: 0.71-0.84), a sensitivity of 0.71, a specificity of 0.68, and a cutoff value of 0.20. The LightGBM model had an AUC of 0.72 (95%CI: 0.65-0.79), a sensitivity of 0.68, a specificity of 0.67, and a cutoff value of 0.21. The AUC of the XGBoost model was 0.70 (95%CI: 0.63-0.78), with a sensitivity of 0.62, a specificity of 0.66, and a cutoff value of 0.21. The DT model achieved an AUC of 0.71 (95%CI: 0.64-0.79), a sensitivity of 0.52, a specificity of 0.85, and a cutoff value of 0.28. Overall, the stacking model demonstrated the best performance.

Our study utilized seven cycles of NHANES data, which are reflective of the diverse racial and gender composition of the United States population, thus enhancing the generalizability of our findings. We developed three fundamental learning models along with a stacking ensemble model and subsequently compared their performances. Our results indicate that the stacking model exhibited superior performance and demonstrated a greater ability to predict high-risk groups for depression among HF patients. The implications of this study are significant because this model can assist clinicians in promptly identifying HF patients at high risk for depression, thus enabling them to initiate psychological interventions and treatments at an earlier stage and ultimately improving the prognosis of HF patients. However, this study was limited by the reliance on questionnaires for diagnosing HF and depression, which may have affected the accuracy of disease diagnosis. Additionally, the prediction model in this study considered only laboratory indicators and demographic information as predictors.

CONCLUSION

In this research, a stacking ensemble model was developed by utilizing three base –models (LightGBM, DT, and XGBoost) with data from the NHANES. The predictive performance of the four models was compared, thus demonstrating that the stacking model outperformed the individual base models. The stacking model proved to be effective in predicting the likelihood of depression in HF patients, thus demonstrating both consistency and high clinical predictability. This tool offers clinicians the ability to identify HF patients at high risk of depression for timely psychological interventions and treatments, thus ultimately improving the prognosis of HF patients.