1. Introduction

It has been revealed that heart disease has become the leading cause of death for the majority of people with one person dying every 33 s from cardiovascular disease (CVD) based on American statistics [1, 2]. Great attention has been aroused with ~252.2 billion dollars in year 2019 spent on healthcare service, medicine, and lost productivity due to heart diseases [3]. In order to focus on cardiovascular health (CVH) to maintain great health conditions, the American Heart Association (AHA) has introduced a scoring metric called Life’s Essential 7 (LE7) with consideration of health behavior elements of diet, nicotine exposure, and physical activities as well as health factor elements of blood pressure, blood glucose, blood lipids, and body mass index (BMI) [4, 5]. Subsequently, an improved method called Life’s Essential 8 (LE8) with the addition of sleep variable in 2018 for its effects in the determination of cardiovascular-related risky factors [6, 7] and these eight subindicators would comprehensively reflect such lifestyle-related and clinical aspects toward CVH [8, 9].

Numerous findings from the United States and international studies have demonstrated a consistent protective relationship between ideal CVH and several clinical and preclinical conditions [10], involving premature all-cause mortality, CVD mortality, heart failure (HF), cognitive decline, stroke, etc. For instance, the risk of incident HF was proven to be 61% lower for those with more than four ideal CVH metrics than those with only 0–2 ideal metrics for the black population [10]. Similarly, the negative association was illustrated for LE8 scoring metrics with the presence of nonalcoholic fatty liver disease (NAFLD) [11, 12], and specific LE7 elements of smoking, BMI, physical activities, blood pressure, and blood glucose were statistically associated with risk of chronic kidney disease (CKD) [13], where the impairment severity was measured by an eGFR < 60 mL/min per m², identified as GFR categories of G3a (mildly to moderately decreased) to G5 (kidney failure) based on the latest 2024 KDIGO [14]. Furthermore, the negative correlation between advanced LE8 metrics and both all-cause and CVD-specific mortality was proven with the application of different databases of the National Health and Nutrition Examination Survey (NHANES), UK Biobank, and specific Kailuan prospective cohort studies accordingly [15–17].

Moreover, in terms of mental health aspects, Boehm et al. [18] have illustrated the psychological well-being’s positive correlation with favorable CVH (FCVH) especially for the geriatrics, while the psychosocial discomfort including depressive symptoms was found to be negatively associated with CVH conditions [19, 20]. The scientific statement from the AHA has provided evidence for clear association between psychological health and CVD [21], and the potential role that psychological elements have played in terms of biological or behavioral processes affecting CVD was emphasized as well [21–23]. Additionally, positive psychosocial factors (PPFs) including happiness, satisfaction with life, and social support were found to be linked to CAD risks according to the UK Biobank database [24], while a large number of studies [25–28] based on NHANES have achieved similar conclusions demonstrating the inverse correlation between unfavorable CVH and decreased depressive severity.

For our research purposes, we have narrowed the psychological aspects to depression-specific factors, utilizing both elevated depressive severity and numerical measuring scores instead of previously widely applied binomial levels for the occurrence of disease [26, 27, 29]. Apart from that, we have applied LE8 metrics rather than LE7 to provide a more comprehensive landscape, and participants aged 18–20 were added to make comparison across different age brackets [9], where the majority of studies had their focus on adults with 20 years or older. In the third place, our paper would not only conduct traditional statistical association analysis but also attempt to make predictable estimates and visualized interpretations based on machine learning (ML) algorithms [30, 31], which seemed to be innovative study areas and helpful to clinical diagnosis and treatment processes in the long term.

2. Materials and Methods

3. Results

3.3. Sensitivity Analysis and RCS Plots

Sensitivity analysis has been conducted with stratification of all seven covariates for adults aged over 20 years old with low LE8 subgroup denoted as the reference group. According to Figure 1, the inverse relationship between LE8 and ranked depression conditions was consistent across all subgroups with all calculated OR smaller than one. More specifically, such negative association was apparently stronger in subgroups of female, people between 40 and 60, educational background of college or above, and married status. For age-specific items, changing from low to moderate and high LE8 groups has corresponded to OR changed from 0.464 (95% CI, 0.377–0.571) to 0.250 (95% CI, 0.185–0.337) for young adults aged 20–40, together with OR from 0.456 (95% CI, 0.375–0.555) to 0.219 (95% CI, 0.148–0.325) for people over 60. For middle-aged people aged 40–60, the odds of being more likely to fall into severe depression conditions were reduced by 66.3% for moderate LE8 and even dropped by ~87% for high LE8 with comparisons of low LE8 teams. Further interaction check has also been conducted, and almost all interactive terms were proven to be insignificant due to p-values larger than 0.05.

Moreover, separate RCS plots were presented to identify specific trends with four age brackets. The overall negative dose–response relationship between LE8 and depression levels was demonstrated due to all significant overall p-values. While for nonlinearity check, participants older than 20 years old had corresponding p-values smaller than cutoff value of 0.01 with group with 18–20 ages group failing the test accordingly. See details in Figure S2 for graphical explanations.

3.4. ML Predictive Models and Representative Interpretations

As could be vividly shown in Figure 2B, eight ML models were generated with AUC values all larger than or equal to 0.5. LR model has shown the greatest AUC of 0.745 with NB model following closely with figure of 0.737, and the XGBoost model ranked the third position with AUC equaling to 0.735. Moreover, different error bars were presented in Figure 2A, and in total, four algorithms models, namely, LR, NB, XGBoost, and RF, were found to present the best fitted predicted results with cutoff AUC number of 0.7. Furthermore, four indicative elements of accuracy, recall, precision, and F1 score were summarized in Figure 2C to present comprehensive selection criteria for optimal selection, and the whole ML package involving all eight models would be of suitability in terms of predictive purposes based on corresponding simulation and validation conducted.

Apart from that, representative interpretations using SHAP values based on XGBoost model were presented according to primary features related to CVH scores and subindicators. As was concluded from Figure S3A, characteristics such as gender, education, and marital would exert significance effects on predictive results with negative relation for males and married participants in contrast with positive impact for higher academic background. Based on Figure S3C,D, health behavior score, gender categories, and poverty status have become the top three significant variables with respect to impact on predictive models, and the seven elements identified in Figure S3B were proven to represent evident effects toward final prediction results.

4. Discussion

Based on our NHANES analysis toward cardiovascular risks represented by LE8 indicators as well as depression variables, overall statistical analysis has robustly illustrated the negative association with specified concentration on four different age brackets. Additionally, a package of eight ML models were selected to present predictive results with SHAP values applied for increasing model interpretability.

CVDs have continuously aroused people’s attention due to its high prevalence and mortality, accounting for ~18 million deaths in recent years based on the World Health Organization (WHO) statistics [37]. Several prospective cohort studies discovered the negative associations between ideal LE7 and depressive symptoms according to participants established from the United States, China, France, Brazil [28, 38–40], showing the declining trend of depression risks with greater number of CVH levels. With further adoption of new LE8 scoring metrics with incorporation of sleep factor, similar negative association between LE8 and depression would also be proven [6, 7, 27].

Depression is a complex condition with diverse symptom patterns, including feelings of hopelessness or pessimism and unusual behavioral changes, such as increased alcohol or drug usage [41]. Symptoms can vary widely across individuals across age, race, gender, and cultural background, differing in type, frequency, severity, and so on [41–43]. In addition to psychological signs, depression may also present with physical manifestations, such as a racing heartbeat, chest tightness, chronic headaches, or digestive problems [41, 43]. Therefore, a comprehensive and objective quantifying instrument is of great significance with consideration various dimensions. The PHQ-9, a useful nine-item instrument currently widely applied, could not only identified as the depression screener but also validated as a continuous measure of depressive severity [33, 36]. It has been proven that PHQ-9 is acceptable for major US sociodemographic groups and subgroup comparisons by Patel et al. with scientific assessment of the factor structure and measurement invariance of PHQ-9 [44]. Such PHQ-9 administration held by trained interviewers was further proven to hold diagnostic validity in comparison with the original clinician-administered PRIME-MD but even with more efficiency in practical applications as well [34].

The majority of previous studies commonly applied binomial-dependent variables with occurrence of depression disease with a cutoff PHQ-9 score of 10 points [33, 34, 45]. Our studies, on the contrary, has the revolution to apply elevated or ranked depressive severity levels and assess relationship between ranked ordinal-dependent variables with both continuous and categorical LE8 indicators. Our interpretative results have robustly showed the greater likelihood with worse depressive severity conditions with increasing LE8 scores, regardless of age groups selected and covariates about gender, education, marital, family income, and size added.

Apart from that, another advanced investigation could be the age-specific subgroup analysis. Adult participants have aroused great attention with more mature status and thorough information [4, 5, 9] in previous studies, while our researches have not only included observations from eligible teenagers aged 18–20 whose PHQ-9 responses toward depressive symptoms were available but also studied sensitivity of associated correlation across different age groups with age 20 as a cutoff threshold. In terms of CVH conditions, comparisons across four age groups in our studies revealed that unfavorable CVH outcomes were more apparent in older population, while the majority of those with high CVH belong to young adults in 20–40 age range. This observation was consistent with several studies [46–48] highlighting that older patients tend to experience a higher prevalence of CVD and risk factors [49, 50], often accompanied by age-related comorbidities [50, 51]. For instance, a study by Lettino et al. [46] found that the prevalence of hypertension was 28% among individuals aged 20–40 but rose significantly to 77% in those over 65 [46, 52]. Similarly, the incidence of hyperlipidemia was 76% for older individuals with a mean age of 69, compared to 41% in younger ones with a mean age of 29 [53], and cognitive or emotional factors of the geriatric could influence their capabilities to adhere to lifestyle changes or pharmacological treatment to some extent [46, 51, 54].

With regard to age-specific depression analysis, our studies revealed that adults aged 40–60 had worst depression conditions with 9.2% with a diagnosis of depressive illness, in contrast to 6.4%, 8.3%, and 8.4% for individuals with age of 18–20, 20−40, and over 60 accordingly. This finding is of importance in arousing greater public awareness of depression, especially for middle aged or the elderly. According to WHO statistics and Global Health Estimates (GHE) [55, 56], around 14% of adults over 60 live with a mental disorder, accounting for 10.6% of total disability measured by disability-adjusted life years (DALYs), and those elderly are suffering from a greater risk of depression and anxiety with mental health conditions underrecognized and undertreated. The term “geriatric giants” was introduced in 1965 and evolved to encompass four new syndromes where depression has played a unignorable role for the elderly [57]. Major depression has affected 2% of adults aged 55 and older, while 10%–15% experience clinically significant depressive symptoms as well, with the corresponding prevalence increasing with age [58]. Indeed, depressed older people were more likely to exhibit cognitive impairments, somatic complaints, and diminished interest than the younger individuals [59, 60]. Therefore, with purpose of investigating depression among the geriatric patient, several instruments were put forward. One tool called Geriatric Depression Scale (GDS), including three versions of GDS-30, GDS-15, and GDS-5, was regarded as the widely applied method for depression detection toward older population rather than general population [49, 50, 61]. Another method called PHQ-9, which was implemented in our NHANES analysis as well, was also appraised for its superior reliability and validity for detecting major depression [62]. Furthermore, when examining PHQ-9 and GDS-15 in older adults aged more than 60 years, both of them achieved superior outcomes with the inner consistency rate of 96.1%, and PHQ-9 seemed more attractive due to its ease of use and relative brevity [63]. Additionally, based on comparisons across various depression screening tools with respect to older veterans, the PHQ-9, PHQ-2/PHQ-3, GDS-15, and GDS-5 were all deemed effective. Notably, GDS-15 and GDS-5 were particularly recommended for older adults in outpatient psychiatric clinics or those who are medically fragile and hospitalized, as these populations tend to have higher base rates of depression. In contrast, the PHQ-9 was advised for medically stable older individuals or those in outpatient settings without the heavy burden imposed on fully functional populations [64].

When it comes to specific LE8 scoring components, health behavior elements involving diet, exercise, smoking, and sleeping had larger OR compared with health factor variables of BMI, blood glucose, blood lipids, and blood pressure. Such findings were plausible for the closer relationship as well as direct effects between depressive mental disease and behavior changes or patterns, which were supported by studies like Zhang et al. [28] and Zeng et al. [27] stating the robust linkage with respect to depression and health behaviors. Moreover, demographic characteristics have revealed portrait of subgroups belonging to high CVH with higher percentage of female, non-Hispanic White, college or above background, married status, family size of one to three members, and wealthier conditions (PIR above 3.5). Such phenomena would be explained by similar findings between socioeconomic status (SES) and CVD burden [65], with less advantageous background subgroups showing prevalence in disease incidence or mortality with biological, behavioral, and psychosocial factors taking effect [66, 67]. Furthermore, with age-specific considerations, sleep variable seemed to have larger scale of change in terms of depressive severity for adults over 20 compared with other seven variables. This finding was supported by analyzing relationship between short or extended sleep and depressive risky levels, and such association was proven apparent with adults and elderly particularly [68–70].

Instead of merely conducting association analysis based on statistical techniques, another improved point was generating ML models for cohort data predictions. Eight widely used ML algorithms were employed [71, 72], with four of them proven to be best fitting in our studies. Specifically, such diagnostic effects would play an important role in identifying and estimating possibility of depressive diseases based on biomarkers or features related to LE8 scoring and SES conditions at a very early stage, which seemed to be more direct and easy-to-implement with quantitative scoring metrics implying obscure and mysterious depressive challenges [30, 31]. Furthermore, SHAP tools were applied to increase the interpretability by prioritizing the features and visualizing the model’s decision-making capacities [73, 74], which would generate predictable explanations for decisions and estimates in a precise manner [72, 75].

Several strengths were apparent in our studies. Firstly, representative and nationwide survey was analyzed with large-scale observations, and possible weight adjustment enabled extension and estimation to population perspectives as well. Secondly, advanced LE8 metrics replaced previous LE7, and four age brackets especially the young adults aged 18–20 were included as a supplement. Thirdly, statistical analysis was improved by utilizing ordinal regression model to analyze elevated or ranked data instead of common binomial responses, and further predictable ML models together with SHAP values and feature diagrams were added which would contribute to clinical diagnosis procedures.

However, some limitations could not be ignored. Since subjective answers from questionnaires were collected, memory distortion and recall bias related to diet, physical activity, smoking, and sleeping might be inevitable. And such cross-sectional study instead of prospective cohorts failed to imply causal relationships, and other covariates or compounders might be omitted from our selected variables. Thirdly, while PHQ-9 is widely used due to its effectiveness and convenience, it is important to acknowledge that certain atypical symptoms of depression may go unnoticed, and some physical discomforts might be overlooked as well. And given that PHQ-9 in NHANES dataset could make assessment over a fixed 2-week period, episodic patterns or seasonal affective disorder (SAD) might not be captured. Therefore, more advanced instruments would be called for with inclusion of mental or physical, common, or unusual symptoms, along with interviews with wider research window such as the similar Canadian Community Health Survey conducted using PHQ-9 throughout the whole year [76]. Additionally in terms of ML predictive models, the lack of adaptive and automatic hyperparameter tuning process might become problematic, and manual hyperparameter adjustment and extensive experimentation was time-consuming and computationally expensive to some extent.

5. Conclusion

In conclusion, a negative relationship between CVH measured by LE8 and elevated depressive severity was demonstrated, and further subgroup stratification and sensitivity check have proven the model robustness. Four age groups with inclusion of teenagers below 20 were classified to make comparisons, and sleep variable was found to have more sensitive effects toward depression in adults aged 20 or above. Participants with profiles of less advantageous SES conditions had worse depressive symptoms, and a package of ML models were identified to make predictable results with SHAP tools providing assistance in clinical diagnosis and treatment.

Ethics Statement

The following protocol approval numbers were provided by the NCHS Ethics Review Board for the presented survey years: Continuation of Protocol #2005–06 (NHANES 2007–2010), Protocol #2011–17 (NHANES 2011–2012), Continuation of Protocol #2011–17 (NHANES 2013–2016), and Protocol #2018–01 (NHANES 2017–2018).

Consent

This is involved in the NHANES program for eligible participants.

Conflicts of Interest

The authors declare no conflicts of interest.

Author Contributions

Zhixing Wu and Lin Zhang conceived and designed the initial idea of the study. Zhixing Wu and Pengyuan Xu collected, cleaned the data, and performed the statistical analysis. Zhixing Wu and Pengyuan Xu applied the machine learning algorithms and conducted the model predictions and visualization. Zhixing Wu drafted the manuscript, and all authors revised the manuscript with comments. All authors read and approved the final manuscript. Zhixing Wu is the guarantor of this work, having full access to all the study data and could take responsibility for the integrity of the data and accuracy of analysis. Zhixing Wu and Pengyuan Xu contributed equally to this work and are first authors.

Funding

This study is supported by the Leading International Talent Funding at Dushu Lake, Suzhou (MSRI8001021).

Supporting Information

Additional supporting information can be found online in the Supporting Information section.