Biopsychosocial phenogroups in an urban population with coronary artery disease and their associations with adverse cardiovascular events

Abstract

Aims

Patients with stable coronary artery disease (CAD) are clinically heterogeneous and exhibit differential treatment responses. We aimed to identify distinct CAD phenogroups based on psychosocial, autonomic, and cardiovascular characteristics and to examine their risk of adverse cardiovascular events.

Methods and results

We pooled data from 949 participants with stable CAD enrolled in two related studies. Phenogroups were derived using Gaussian mixture model clustering applied to 12 pre-specified markers readily obtainable in ambulatory clinical settings: autonomic dysfunction, psychosocial stress, myocardial injury, obstructive CAD burden, left ventricular ejection fraction, and blood pressure. Hazard ratios (HRs) were estimated to examine the associations between the derived phenogroups and composite major adverse cardiovascular events (MACE), cardiovascular disease (CVD)-specific, and all-cause mortality. The mean age was 60 ± 10 years; 34% were women and 41% were Black. Four phenogroups were identified: phenogroup 1 (n = 257), the ‘low-risk cluster’; phenogroup 2 (n = 387, most prevalent), the ‘psychosocial stress-angina cluster’; phenogroup 3 (n = 142), the ‘autonomic dysfunction cluster’; and phenogroup 4 (n = 163), the ‘ischaemic cardiomyopathy cluster’. Compared with the low-risk cluster, the ischaemic cardiomyopathy cluster had the highest risk of composite MACE (HR, 2.93; 95% CI, 1.80–4.78), followed by the autonomic dysfunction cluster (MACE HR, 2.07; 95% CI, 1.21–3.53) and the psychosocial stress-angina cluster (HR, 1.65; 95% CI, 1.04–2.61). Similar associations were observed for CVD-specific and all-cause mortality.

Conclusion

Psychosocial and autonomic factors, when combined with cardiovascular risk characteristics, result in clinically meaningful CAD phenogroups with varying risk of adverse cardiovascular events that may help guide personalized interventions.

Lay summary

Patients with stable coronary artery disease exhibit distinct clusters of psychosocial, autonomic, and cardiovascular characteristics that influence their risk of adverse cardiovascular events

• Among this high-risk population, the most prevalent group was characterized by elevated psychosocial stress and frequent angina and was associated with a higher risk of adverse cardiovascular events.

• More research is needed to test whether integrating psychosocial stress assessment and targeted behavioral or psychosocial interventions into secondary prevention strategies, alongside standard medical therapy, may help reduce the risk of adverse cardiovascular events in these high-risk patients.

Graphical Abstract

Graphical Abstract.

Introduction

Individuals with stable coronary artery disease (CAD) represent a clinically heterogeneous group with varying psychosocial, autonomic, and cardiovascular profiles that may differentially respond to various treatments, but unfortunately, limited studies have examined personalized models of secondary CAD prevention.^1,2 Psychosocial stressors are prevalent and independently associated with adverse cardiovascular outcomes in patients with CAD, while also shaping health behaviours and quality of life; however, they are not routinely incorporated into secondary prevention strategies in clinical cardiology practice.^3,4 More research is needed to understand how psychosocial factors may combine with biomedical factors to influence the risk of adverse cardiovascular events. For example, while some individuals with CAD have severe mental illness and relatively low amounts of CAD burden, others may have high CAD burden and fewer psychosocial comorbidities. By phenomapping the different patterns of psychosocial and cardiovascular pathologies amongst individuals with CAD, we can develop more personalized secondary prevention approaches that include varying amounts of biobehavioral, neuromodulator, and psychotropic interventions.^5,6

Unfortunately, relatively few studies have incorporated psychosocial factors into their phenomapping algorithms, given the challenges of collecting psychosocial data in large cohorts.^7,8 Nonetheless, certain measures like patient-reported outcomes (PROs) and resting heart rate variability (HRV) are increasingly available using free tools and consumer wearable devices.⁹ Autonomic nervous system dysregulation, as reflected by reduced HRV, has been strongly linked to adverse cardiovascular outcomes, including mental stress-induced myocardial ischaemia and cardiovascular death.^5,10,11 In addition, psychosocial stress and autonomic dysfunction appear to have a bidirectional relationship, whereby chronic psychosocial stress may impair autonomic regulation, while autonomic dysregulation may also contribute to heightened stress responsiveness and poorer cardiovascular adaptation.^12,13 Thus, the inclusion of both HRV and PROs of psychological stress may provide complementary information on the physiological (HRV) and perceived measured stress (PROs), which may differentially characterize patient groups and influence outcomes. Phenomapping may help uncover the clustering of psychosocial and cardiovascular comorbidities in a way that can guide treatment decisions by clustering group. One example is cardiac rehabilitation, an intervention that improves psychological health and reduces the risk of CVD death, but can also be customized to the specific phenogroup of individuals.^14,15 Integrated and intensive versions of cardiac rehabilitation may be more appropriate for some individuals with particularly high risk, but more research is needed to identify the best candidates for these programs, given the investment required from both the patient and healthcare system.¹⁶

In this study, we sought to derive novel phenogroups in patients with CAD with detailed psychosocial and autonomic function characterization. To facilitate reproducibility, we restricted the candidate features to a limited number of traditional CVD predictors, cardiac autonomic measures, and psychosocial self-reported stress that are feasible to obtain in many clinical settings. We hypothesized that psychosocial and autonomic factors would play a key role in the formation of these groups, that they would vary by age and sex, and that they would be differentially associated with major adverse cardiovascular events (MACE), CVD-specific and all-cause mortality.

Methods

Study population and design

Our study sample was drawn from 2 related studies: MIPS (Mental Stress Ischaemia Mechanisms and Prognosis Study; n = 636) and MIMS2 (Myocardial Infarction and Mental Stress Study 2; n = 313).¹⁷ Both studies recruited participants with CAD and adhered to similar protocols as previously described.¹⁸ While MIMS2 recruited post-MI patients within 8 months of their event, MIPS recruited from a larger pool of participants with stable CAD. The Emory University Institutional Review Board approved the research protocol for both studies, and all participants provided written informed consent during enrolment.¹⁷

To enhance reproducibility and clinical applicability, we limited our analysis to 12 variables based on clinical expertise and prior literature, as well as previous literature showing that limiting the number of features improves model stability and generalizability.^19,20 We included markers of autonomic dysfunction- low frequency (LF), high frequency heart rate variability (HF HRV), and heart rate- myocardial injury and obstructive burden (high sensitivity troponin I, Gensini score, angina frequency), left ventricular ejection dysfunction (LVEF), blood pressure, and psychosocial stress burden for the clustering algorithm. These variables were chosen based on our findings from previous research and feasibility to obtain them in clinical settings.^5,21,22

Data collection

Sociodemographic factors (age, sex, race, income, employment, and current smoking status), medical history [hypertension, hyperlipidaemia, diabetes mellitus, obesity, heart failure, previous percutaneous transluminal coronary angioplasty, and previous coronary artery bypass graft (CABG)], and medications taken (antidepressants, beta blockers, aspirin, and statins) were assessed using standardized questionnaires assisted by trained clinical staff. More detailed information on the methodologies used to collect this information has been reported in previous studies.¹⁷

Assessment of psychosocial factors

We included data from three related but distinct dimensions of psychosocial stress. This includes:

1. Depressive symptoms, assessed via the Beck Depression Inventory II (BDI-II), a 21-item self-report assessment of depressive symptoms.²³

2. Current PTSD symptoms, assessed using the PTSD symptom checklist (PCL), a 17-item self-report measure of the 17 Diagnostic Statistical Manual IV symptoms of PTSD (range, 17–85).²⁴

3. Perceived discrimination, assessed via the 10-item version of the Everyday Discrimination Scale (EDS).²⁵ Sample items include being treated with less courtesy than other people, receiving poorer service than other people at restaurants or stores, etc., without reference to race, gender, or any other characteristic. The EDS has been widely used with strong reliability and validity across racial/ethnic groups.²⁶

Collectively, these three scales provide a multidimensional assessment of psychosocial burden that may have varying effects on mechanisms of CVD risk, including internalizing psychopathology from depressive symptoms leading to hypothalamic-pituitary-adrenal axis dysregulation (BDI-II), trauma-related distress leading to sympathetic overdrive (PCL), and structural/interpersonal stress exposure causing a generalized increase in allostatic load (EDS).

We also included angina symptoms because of their relevance to quality of life and well-known relationship with psychological factors. Previous research has shown that angina is more strongly correlated with psychological stress than myocardial ischaemia or atherosclerotic burden.²⁷ Angina symptoms were assessed using the Seattle Angina Questionnaire’s (SAQ) angina frequency subscale which measures frequency of angina and use of nitroglycerin for chest pain over the previous four weeks. The SAQ has been validated in different populations and correlates with electronic daily diary entries for angina and nitroglycerin use.²⁸

Cardiovascular measures

CAD severity was quantified using the Gensini semiquantitative angiographic scoring system, which takes into account the potential functional significance of the coronary lesion with a multiplier for specific coronary tree locations.²⁹ Native plaque burden was measured before revascularization. CAD severity data were not available in all participants; in most cases, this was due to the fact that their detailed angiogram data were obtained outside of the Emory Hospital network and were not accessible.

LVEF was derived from myocardial perfusion imaging for participants in MIPS, which has a strong correlation with echocardiography.³⁰ In MIMS2, it was derived from echocardiography, coronary angiography, or myocardial perfusion imaging (based on availability) at the time of hospitalization for MI. Blood pressure and heart rate were also recorded at the baseline visit.

Baseline high-sensitivity troponin I (hs-cTnI) is another biomarker of risk that is commonly used for risk stratification in chest pain, but also recently found to have long-term prognostic value as well.³¹ We included this as a model feature because of its strong prognostic value, relationship with CVD outcomes, and clinical availability. Blood was drawn while at rest for hs-cTnI measurement. Samples were processed and stored at −80°C. Plasma hs-cTnI was measured using the ARCHITECT STAT Hs-cTnI assay (Abbott Laboratories, Abbott Park, Illinois), which has a limit of detection of 1.2 pg/mL and an interassay coefficient of variation of <10% at 4.7 pg/mL. Of note, hs-cTnI was obtained as part of a post-study ancillary proposal and not available in all participants due to insufficient sample.

Autonomic measures

The relationship of psychological stress and autonomic function is well-known; we have previously found that both resting 5-min HRV and stress-induced changes were associated with CVD mortality in this cohort.⁵ Each participant wore a Holter monitor (GE Marquette SEER digital system; GE Medical Systems, Waukesha, WI) while in a seated position. HRV measures were obtained in a 5-min window period while at rest, at approximately the same time of day across participants using GE MARS 8.0.2 (2015) software, to reduce potential variability due to circadian influences on autonomic function. Participants were instructed to abstain from certain substances and medications, including beta-blockers, on the morning of their research visit, as these factors can acutely influence HRV measurements. Fast Fourier transformation was used to classify HRV into 2 frequency bands: low-frequency (LF) (0.05 to <0.15 Hz) and high-frequency (HF) (0.15 to <0.40 Hz). Frequency-domain HRV measures were selected based on their established physiological relevance and prognostic value in patients with CAD, including in this cohort.^5,10 Time-domain measures were not included because of their redundancy with the selected frequency-domain measures, particularly the close correlation between RMSSD and HF HRV, and because SDNN reflects total variability arising from multiple physiological sources and is therefore a less specific index of autonomic function.³² Additional adjustment for respiration was not performed during HF HRV ascertainment because prior studies suggest that such adjustment may overcorrect the signal, given that respiration itself reflects ANS activity and HF HRV captures both parasympathetic modulation and respiration-related influences.³³ We obtained autonomic measures from Holter monitoring as part of an ancillary investigation that started after recruitment started, which resulted in a significant amount of missing data.

Evaluation of outcomes

Follow-up data were collected through participant contacts, medical record review, and the Social Security Death Index. Medical records were obtained and reviewed for hospitalizations. In MIPS, participants were contacted every 6 months for the first 3 years and then at the 5-year mark. In MIMS2, participants were contacted at the 3- and 5-year anniversary of their initial visit (approximately). All events were adjudicated by a team of cardiologists who were blinded to baseline study data. We evaluated MACE and CVD-specific mortality as the primary endpoints and all-cause mortality as a secondary endpoint. MACE was defined as a composite of non-fatal myocardial infarction, non-fatal stroke, and CVD-specific mortality.

Model-based clustering approach

We used model-based clustering with parameterized finite Gaussian mixture models to identify distinct clinical phenogroups based on pre-specified risk factors.³⁴ Gaussian mixture modelling is a probabilistic clustering approach that assumes the observed data arise from a heterogeneous population composed of several underlying subpopulations (clusters), with each cluster represented by a mixture of Gaussian distributions characterized by its own mean and covariance structure. Model parameters are estimated using the expectation-maximization algorithm, which iteratively estimates cluster membership probabilities and updates the parameters of each Gaussian distribution until convergence.^35,36 The model assigns each data point a probability of belonging to different clusters, and the final cluster membership is determined based on the highest posterior probability. The optimal model and number of clusters were determined using the Bayesian Information Criterion (BIC). The Mclust package (version 6.1.1) in R (version 4.5.1; The R Foundation for Statistical Computing, Vienna, Austria) was used to perform the model-based clustering algorithm.³⁷ All variables were standardized to have a mean of 0 and a standard deviation (SD) of 1 before clustering. Variables with missing data—heart rate (0.8%), blood pressure (0.8%), LVEF (1.0%), angina frequency (2.7%), BDI-II (3.4%), PCL (4.1%), hs-cTnI (11.2%), Gensini score (11.9%), LF/HF HRV (20.7%), and EDS (24.0%)—were imputed using predictive mean matching under a missing-at-random assumption.³⁸

Statistical analysis

Baseline sociodemographic and clinical characteristics were compared across clusters using one-way analysis of variance (ANOVA) for continuous variables and chi-square tests for categorical variables. The assumptions of normality and homogeneity of variance were assessed prior to applying parametric tests. When the assumption of equal variances was not met, Welch’s ANOVA was used. Fisher’s exact test was used for categorical variables with expected cell counts below five. These comparisons are valid since cluster assignment was based on posterior probabilities from the model-based clustering algorithm. Time-to-event analyses for composite MACE, CVD-specific, and all-cause mortality were performed using cumulative incidence curves, competing risk models, and Cox proportional hazards regression, adjusting for age, sex, and race. Because the sample was pooled from two separate cohorts, we included cohort membership as a fixed effect variable in the adjusted model to account for baseline differences between the two cohorts. Hazard ratios (HR) with 95% confidence intervals (CI) were calculated using the lowest-risk cluster as reference. For all outcomes, we used cause-specific hazard ratios as our primary estimates because our goal was to evaluate the mechanistic effect of the novel clusters on adverse cardiovascular events. Prior studies have shown that cause-specific hazard ratios are more appropriate for addressing etiologic questions than sub-distribution hazard ratios derived from competing-risk models.³⁹ However, given the presence of non-CVD death as a potentially important competing event, we additionally performed sensitivity analyses using the Fine and Gray competing risk for MACE and CVD-specific mortality to assess the robustness of the observed associations to alternative handling of competing risks. We evaluated the proportional hazards assumption using Schoenfeld residuals, and the results indicated no violations.

As an additional sensitivity analysis, we further adjusted for diabetes, hypertension, dyslipidemia, smoking, prior CABG, prior PCI, heart failure, and the use of beta-blockers, statins, antidepressants, and aspirin to evaluate the independent prognostic contribution of phenogroup membership to MACE beyond established risk factors. Finally, as a form of internal validation, we repeated the clustering and survival analyses within the MIPS cohort, which comprised ∼67% of the overall study population. This analysis assessed the consistency of the cluster structure, phenogroup characteristics, and prognostic associations within a single parent cohort.

Results

The overall cohort had a mean age of 60 (10) years; 34% were women and 41% were Black. Over a median follow-up of 5.9 years, 167 (18%) total MACEs, 51 (5%) CVD-specific, and 90 (9%) all-cause mortality events were recorded.

Model selection and cluster formation

First, we examined the correlations among the variables and found no strong correlations (−0.6 < r < 0.6; Supplementary material online, Figure S1). The results of the model-based cluster analysis are shown in Supplementary material online, Figure S2. The optimal model involved a four-cluster solution with a BIC of −23737. This model consisted of four clusters with the VVE covariance parameter (variable volume, variable shape, and equal orientation). The average posterior probabilities for the most likely cluster classification were 0.95, 0.94, 0.99, and 0.93 for clusters 1, 2, 3, and 4, respectively, indicating a high degree of certainty in the classification of the four clusters.

Phenogroup characteristics

The four phenogroups (clusters) demonstrated distinct sociodemographic and clinical characteristics (Table 1). Phenogroup 1 included the highest proportion of White and employed participants, the highest income levels, and the lowest prevalence of prior CABG and heart failure. In contrast, phenogroup 2 had higher female representation, the highest proportion of Black participants and low-income individuals (family income <$20 000/year), and the highest prevalence of antidepressant users. Phenogroup 3 comprised the oldest participants and had the highest prevalence of heart failure. Phenogroup 4 was the youngest group, with the greatest proportion of males and the highest prevalence of prior CABG.

Table 1.

Participant characteristics

	Phenogroup 1	Phenogroup 2	Phenogroup 3	Phenogroup 4	P-value
	(N = 257)	(N = 387)	(N = 142)	(N = 163)
Demographic factors
Age, mean (SD), years	61 (10)	58 (10)	63 (9)	57 (11)	<0.001
Sex, n (%)					0.002
Male	178 (69)	228 (59)	100 (70)	120 (74)
Female	79 (31)	159 (41)	42 (30)	43 (26)
Race, n (%)					<0.001
White	179 (70)	201 (52)	96 (68)	88 (54)
Black or African American	78 (30)	186 (48)	46 (32)	75 (46)
Income <$20,000, n (%)	24 (9.3)	106 (27.4)	19 (13.4)	39 (23.9)	<0.001
Employed, n (%)	129 (50.2)	133 (34.4)	38 (26.8)	58 (35.6)	<0.001
CVD risk factors
BMI, mean (SD), kg/m2	30 (5)	30 (7)	30 (5)	31 (7)	0.12
Current smoking, n (%)	106 (41)	149 (39)	72 (51)	61 (37)	0.06
Diabetes, n (%)	75 (29)	127 (33)	56 (39)	49 (30)	0.18
Hypertension, n (%)	198 (77)	311 (80)	100 (70)	124 (76)	0.11
Dyslipidemia, n (%)	221 (86)	311 (80)	111 (78)	135 (83)	0.18
CVD history
History of PTCA, n (%)	145 (56)	236 (61)	72 (51)	92 (56)	0.19
History of CABG, n (%)	60 (23)	96 (25)	50 (35)	77 (47)	<0.001
History of Heart failure, n (%)	17 (7)	43 (11)	33 (23)	31 (19)	<0.001
Selected Medications
Antidepressant use, n (%)	42 (16)	109 (28)	37 (26)	25 (15)	<0.001
Beta blocker use, n (%)	187 (73)	307 (79)	102 (72)	136 (83)	0.02
Aspirin use, n (%)	224 (87)	326 (84)	121 (85)	132 (81)	0.39
Statins use, n (%)	227 (88)	330 (85)	114 (80)	134 (82)	0.13

P-values were obtained from one-way or Welch’s analysis of variance for continuous variables and chi-square tests for categorical variables. Statistical testing for these comparisons is interpretable since cluster assignment was based on posterior probabilities determined by the model-based clustering algorithm.

Abbreviations: BMI, Body mass index; CABG, Coronary artery bypass graft; CVD, Cardiovascular disease; PCTA, Percutaneous transluminal coronary angioplasty.

As illustrated in Figures 1 and S3 and summarized in Table 2, the phenogroups also differed across the clinical domains. Phenogroup 1, the ‘low risk’ group, consistently had the lowest levels of all five domains. Phenogroup 2, the ‘psychosocial stress-angina’ group, exhibited the highest levels of psychosocial stress, alongside low to intermediate levels of the other domains. Phenogroup 3, the ‘autonomic dysfunction’ group, was characterized by elevated levels of autonomic dysfunction as indexed by lower HRV and by elevated levels of myocardial injury. Finally, phenogroup 4, the ‘ischaemic cardiomyopathy’ group, demonstrated low LVEF and elevated levels of obstructive CAD.

Figure 1.

A radar chart showing the pattern of variables in each phenogroup. Radar plot showing the mean values of variables for each cluster. Points closer to the centre indicate lower values, while points farther from the centre indicate higher values. Phenogroup 1 represents the ‘low-risk group’; Phenogroup 2 represents the ‘psychosocial stress-angina group’; Phenogroup 3 represents the ‘autonomic dysfunction group’; Phenogroup 4 represents the ‘ischaemic cardiomyopathy group’. BDI, Beck’s depression inventory score; DBP, Diastolic blood pressure; HF-HRV, High frequency heart rate variability; HR, Heart rate; LF-HRV, Low frequency heart rate variability; PTSD, Post-traumatic stress disorder; SBP, Systolic blood pressure.

Table 2.

Summary of phenotypic domains for each phenogroup

Phenogroup	Autonomic dysfunction	Myocardial Injury & Obstructive burden	LV systolic dysfunction	Blood pressure	Psychosocial stress-angina
1	Low	Low	Low	Low	Low
2	Intermediate	Low	Low	Intermediate	High
3	High	High	Intermediate	Intermediate	Intermediate
4	Low	High	High	Intermediate	Intermediate

Phenogroup 1 represents the ‘low-risk group’; Phenogroup 2 represents the ‘psychosocial stress-angina group’; Phenogroup 3 represents the ‘autonomic dysfunction group’; Phenogroup 4 represents the ‘ischaemic cardiomyopathy group’. High, intermediate, and low assignments were based on the relative levels of individual variables within each domain as depicted in the radar plot. Domains and their representative variables included: Autonomic dysfunction (low- and high-frequency heart rate variability, heart rate); Myocardial injury/obstructive burden (troponin I, Gensini score); LV systolic dysfunction (LV ejection fraction); Blood pressure (systolic and diastolic values); and Psychosocial stress-angina(PTSD score, BDI score, discrimination score, and angina frequency).

Abbreviations: BDI, Beck’s depression inventory scale; LV, Left ventricular; PTSD, Post-traumatic stress disorder.

Clinical outcomes and prognostic significance

Overall, the cumulative incidences of composite MACE, CVD-specific, and all-cause mortality were highest among participants in the ischaemic cardiomyopathy phenogroup and lowest among those in the low-risk phenogroup (Figure 2). Specifically, the ischaemic cardiomyopathy phenogroup recorded 46 (28%) MACEs, 23 (14%) CVD-specific, and 28 (17%) all-cause deaths. The autonomic dysfunction phenogroup had 29 (20%) MACEs, 10 (7%) CVD-specific, and 22 (16%) all-cause deaths; the psychosocial stress-angina phenogroup had 66 (17%) MACEs, 14 (3.6%) CVD-specific, and 27 (7%) all-cause deaths; and the low-risk phenogroup had 26 (10.1) MACEs, 4 (2%) CVD-specific, and 13 (5%) all-cause deaths.

Figure 2.

Cumulative incidence curves for CVD-specific and all-cause mortality by phenogroups. (A) Composite MACE. (B) CVD-specific Mortality. (C) All-cause Mortality. Crude cumulative incidence curves for composite MACE (A), CVD-specific (B), and all-cause (B) mortality, stratified by the novel phenogroups. Phenogroup 1 represents the ‘low-risk group’; Phenogroup 2 represents the ‘psychosocial stress-angina group’; Phenogroup 3 represents the ‘autonomic dysfunction group’; Phenogroup 4 represents the ‘ischaemic cardiomyopathy group’.

Using the low-risk group as the reference, significantly higher risks of composite MACE, CVD-specific, and all-cause mortality were observed across the other three groups in both crude and adjusted models (Table 3). Participants in the ischaemic cardiomyopathy phenogroup carried the greatest risk overall, with an approximately threefold higher risk of composite MACE (adjusted HR, 2.93; 95% CI, 1.80–4.78), eightfold higher risk of CVD-specific mortality (adjusted HR, 8.45; 95% CI, 2.89–24.71) and a 3.7-fold higher risk of all-cause mortality (adjusted HR, 3.72; 95% CI, 1.91–7.26) after adjusting for demographic factors and cohort membership, compared with the low-risk phenogroup (Table 3). Participants in the autonomic dysfunction group also had a twofold higher risk of composite MACE (adjusted HR, 2.07; 95% CI, 1.21–3.53), a 3.7-fold higher risk of CVD-specific mortality (adjusted HR, 3.67; 95% CI, 1.13–11.88) and a more than twofold higher risk of all-cause mortality (adjusted HR 2.39; 95% CI, 1.19–4.81) compared with the low-risk phenogroup.

Table 3.

Association of novel phenogroups with adverse cardiovascular events

	Low risk phenogroup (n = 257)	Psychosocial stress-angina phenogroup (n = 387)	Autonomic dysfunction phenogroup (n = 142)	Ischaemic cardiomyopathy phenogroup (n = 163)
Composite MACE
Total No. of events, n (%)	26 (10.1)	66 (17.1)	29 (20.4)	46 (28.2)
Rate per 100 patient-years	1.9	3.4	3.6	6.1
Rate difference per 100 patient-years (95% CI)	Ref.	1.54 (0.43, 2.64)	1.73 (0.22, 3.24)	4.14 (2.25, 6.04)
Crude HR (95% CI)	Ref.	1.83 (1.16, 2.88)	1.98 (1.16, 3.36)	3.20 (1.98, 5.17)
Adjusted HR (95% CI)	Ref.	1.65 (1.04, 2.61)	2.07 (1.21, 3.53)	2.93 (1.80, 4.78)
CVD-specific mortality
Total no. of events, n (%)	4 (1.6)	14 (3.6)	10 (7.0)	23 (14.1)
Rate per 100 patient-years	0.3	0.7	1.2	2.8
Rate difference per 100 patient-years (95% CI)	Ref.	0.39 (−0.06, 0.83)	0.87 (0.10, 1.63)	2.47 (1.31, 3.64)
Crude HR (95% CI)	Ref.	2.31 (0.76, 7.01)	3.73 (1.16, 12.01)	9.72 (3.36, 28.12)
Adjusted HR (95% CI)	Ref.	2.13 (0.70, 6.54)	3.67 (1.13, 11.88)	8.45 (2.89, 24.71)
All-cause mortality death
Total no. of events, n (%)	13 (5.1)	27 (7.0)	22 (15.5)	28 (17.2)
Rate per 100 patient-years	0.9	1.3	2.5	3.4
Rate difference per 100 patient-years (95% CI)	Ref.	0.37 (−0.33, 1.07)	1.61 (0.44, 2.78)	2.44 (1.10, 3.78)
Crude HR (95% CI)	Ref.	1.37 (0.71, 2.66)	2.52 (1.26, 5.05)	3.65 (1.89, 7.04)
Adjusted HR (95% CI)	Ref.	1.36 (0.70, 2.66)	2.39 (1.19, 4.81)	3.72 (1.91, 7.26)

Low-risk represents phenogroup 1; Psychosocial stress-angina represents phenogroup 2; Autonomic dysfunction represents phenogroup 3; ischaemic cardiomyopathy represents phenogroup 4.

The low-risk group was used as a reference to calculate the rate difference and cause-specific hazard ratio for the other 3 groups.

95% Confidence intervals are wide due to the limited number of events overall and within phenogroups. Adjusted models included age, sex, race, and cohort membership.

Abbreviation: CVD, cardiovascular disease; HR, Hazard ratio; MACE, major adverse cardiovascular events

Lastly, participants in the psychosocial stress-angina phenogroup had a 65% higher risk of composite MACE compared with the low-risk phenogroup (adjusted HR, 1.65; 95% CI, 1.04–2.61). Although the association of this phenogroup with CVD-specific and all-cause mortality did not reach statistical significance (Table 3), the observed effect estimates were directionally consistent with those for MACE.

Sensitivity analyses

Using the Fine and Gray competing-risk model, the sub-distribution hazard ratios for composite MACE and CVD-specific mortality were broadly similar to the corresponding cause-specific hazard ratios across the phenogroups (see Supplementary material online, Table S1). Additionally, after accounting for established prognostic factors in a stable CAD population, participants in the ischaemic cardiomyopathy phenogroup continued to have more than a 2.5-fold higher risk of composite MACE compared with the low-risk phenogroup (adjusted HR, 2.58; 95% CI, 1.55–4.29). Similarly, participants in the autonomic dysfunction phenogroup had an ∼80% higher risk of composite MACE compared with the low-risk phenogroup (adjusted HR, 1.81; 95% CI, 1.06–3.11).

As a form of internal validation, we repeated the clustering and survival analyses among participants in the MIPS cohort only. In this subcohort, a four-cluster solution was also identified. Despite some differences in magnitude and distribution, the phenogroups demonstrated broadly similar patterns to those observed in the full cohort (see Supplementary material online, Figure S4). Phenogroup 4 in this subcohort showed characteristics similar to the low-risk phenogroup in the full cohort and was therefore used as the reference group in the survival analyses. Despite the smaller number of events in this subcohort, the identified high-risk phenogroup remained associated with an increased risk of adverse cardiovascular outcomes, including an 89% higher risk of composite MACE (adjusted HR, 1.89; 95% CI, 1.08–3.29), a more than fourfold higher risk of CVD-specific mortality (adjusted HR, 4.33; 95% CI, 1.82–10.26), and a four-fold higher risk of all-cause mortality (adjusted HR, 4.10; 95% CI, 2.18–7.72), compared with the low-risk phenogroup. Supplementary material online, Table S2 presents the associations between the phenogroups in the MIPS cohort and study outcomes.

Discussion

In this study of over 900 participants with clinical CAD, we identified 4 primary phenogroups using model-based clustering, which encompassed distinct pathophysiological patterns with different thematic elements. The identified phenogroups, in order of increasing risk of adverse cardiovascular events, included: low-risk (phenogroup 1), psychological stress-angina (phenogroup 2), autonomic dysfunction and myocardial injury (phenogroup 3), and ischaemic cardiomyopathy (phenogroup 4). The highest risk for composite MACE (3-fold), CVD-specific (over 8-fold), and all-cause mortality (over 3.5-fold) was observed in the ischaemic cardiomyopathy group. Nonetheless, the psychosocial stress-angina phenogroup was the most prevalent (41%) and exhibited a 65% higher risk of composite MACE. While the association between this phenogroup and mortality was not statistically significant, potentially due to limited statistical power resulting from the relatively small number of mortality events, the magnitude of the effect suggests potential clinical relevance and underscores the need for more research. Notably, the low-risk phenogroup, which was characterized by a lower burden of psychosocial stress, had the lowest event rates, further supporting the potential importance of psychosocial well-being in maintaining cardiovascular stability and highlighting the value of incorporating stress-related PROs into cardiovascular risk assessment. To our knowledge, this proof-of-concept study is one of the first to phenogroup patients with CAD using psychosocial and autonomic data. It also underscores the need for more research to develop and test more personalized and integrated therapies.⁴⁰

This study builds upon previous phenomapping studies of CAD patients by integrating psychosocial and autonomic factors as critical factors that help define groups. Previous studies have found these factors are highly predictive of outcomes.^3,5,41 Although our results require validation with more research in larger cohorts, the need to more fully evaluate psychosocial factors and offer biobehavioral therapies is a longstanding need that warrants more research with translational and implementation research.⁴² Studies are also needed to investigate the feasibility of short-term autonomic testing in clinical settings, which complements the collection of psychosocial stress factors through patient-reported outcomes and also may signal higher risk due to diabetic autonomic neuropathy.⁴³

The ‘ischaemic cardiomyopathy’ phenogroup demonstrated the highest risk of adverse cardiovascular events, as expected, although many individuals may not receive the necessary imaging tests to establish this diagnosis.⁴⁴ This may occur especially in situations of significant barriers to care, such as cost or distance from clinical settings. Other considerations include challenges of implementation of medical and lifestyle therapies that are already indicated by the guidelines. For example, a disproportionately low proportion of heart failure patients receive cardiac rehabilitation.⁴⁵ A particularly striking finding was that this phenogroup represented the youngest cohort, with the highest proportion of males and a high prevalence of prior CABG. One possible explanation relates to selection processes inherent to cohort enrolment, including survival bias. Individuals with a combination of severe coronary disease, heart failure, and high psychosocial burden may be at increased risk of premature mortality or may be less likely to participate in intensive research protocols, resulting in their underrepresentation among older participants. This depletion of higher-risk individuals at older ages may lead to an apparent paradox whereby the highest-risk phenogroup appears younger. At the same time, this pattern may also reflect a more aggressive, early-onset form of CAD progressing to ischaemic cardiomyopathy. While this group exhibited a lower psychosocial burden on average, psychosocial and behavioural factors may still play an important role in accelerating disease progression in susceptible individuals.⁴⁶ Future studies could explore potential biobehavioral mechanisms, such as early-life adversity, chronic occupational stress, and other cumulative stress exposures, that may contribute to accelerated cardiovascular risk in these younger patients despite lower observed stress scores.

Although all patients with CAD should receive guideline-directed medical therapy, their management from a health behavioural perspective calls for further discussion. Our findings highlight the importance of considering the heterogeneity in psychosocial and autonomic risk factors when informing secondary prevention strategies in this patient population. Notably, we identified a highly prevalent psychosocial stress-angina phenogroup which was associated with a higher risk of major adverse cardiovascular events. This phenogroup also had the highest proportion of Black and low-income participants, suggesting that social and structural factors may contribute to the observed psychosocial burden. Low socioeconomic status and race-related structural stressors may influence cardiovascular risk through chronic stress exposures and differential access to healthcare and preventive resources. Collectively, these findings show how a substantial proportion of patients may carry elevated risk profiles driven primarily by psychosocial factors rather than traditional or structural cardiovascular disease severity alone. Current guideline-based therapies and coverage decisions are largely anchored to traditional cardiovascular risk markers, functional capacity of patients, and structural disease burden, which may inadvertently exclude patients with high psychosocial burden whose disease appears clinically stable. For example, although cardiac rehabilitation is a guideline-based recommended therapy for certain patients with CAD, many individuals with stable ischaemic forms do not qualify based on Medicare criteria.⁴⁷ Assessment of psychosocial burden using multidimensional scales such as the BDI, PCL, and EDS is not universally recommended in clinical settings for CAD patients, and certain pharmacological treatments are controversial in this group.⁴⁸

Our proof-of-concept findings support the need for more research on the role of autonomic and psychosocial factors in helping to identify clinically significant phenogroups, and also developing personalized treatment strategies that address their risk factor profiles. Future research involving biopsychosocial phenogroup identification and management findings may be particularly useful in tailoring comprehensive lifestyle programs such as cardiac rehabilitation to patients’ predominant risk profiles. For example, patients in the psychosocial stress-angina phenogroup may benefit from greater emphasis on stress-reduction strategies and cognitive behavioural therapy within cardiac rehabilitation, while those with autonomic dysfunction may benefit from targeted biobehavioral or neuromodulation interventions such as vagal nerve stimulation.⁴⁹ The low-risk phenogroup appears to be the most optimally managed and would need the fewest adjustments. The ischaemic cardiomyopathy phenogroup may warrant frequent follow-up with goal-directed medical therapy and aggressive lipid-lowering and, depending on the atherosclerotic burden, antiplatelet therapies.

Our study has several strengths, including the inclusion of a large cohort of patients with CAD and the use of a data-driven clustering approach that allows probabilistic phenogroup assignment and improves interpretability and reproducibility compared with traditional clustering methods. Moreover, this is the first study to demonstrate the feasibility and utility of incorporating psychosocial and cardiac autonomic factors into phenomapping patients with stable CAD. Finally, the prospective design enabled us to assess the differential risk of adverse cardiovascular events across the novel phenogroups, with complementary findings observed in competing-risk analyses using the Fine and Gray model. The study also has limitations. The sample was drawn from a single urban metropolitan area, potentially limiting generalizability across certain ethnic and socioeconomic populations, such as rural populations. Although using a pooled cohort increased sample size and may have introduced between-study differences that potentially influenced cluster formation, the identification of similar cluster patterns during internal validation within the more stable CAD population of the MIPS cohort supports the robustness and interpretability of our findings in a broader CAD population. However, external validation in independent cohorts is needed to confirm phenotype stability and clinical utility. We also acknowledge that heterogeneity in LVEF measurement technique is a limitation; despite this, LVEF remains a robust and highly significant predictor of mortality and MACE, suggesting that it has significant clinical validity despite this limitation. Missing data were handled using predictive mean matching. We cannot prove that missingness was not at random, although the reasons are based on non-biological factors such as availability of data from non-Emory hospitals or delayed initiation of certain ancillary-funded study procedures (e.g. HRV, discrimination survey). The uncertainty caused by imputation also limits the immediate use of this model in clinical settings; nonetheless, these proof-of-concept findings lay the groundwork for future studies in larger cohorts that would allow for more robust, definitive models that are ready for clinical translation. Translation of phenotyping approaches into clinical practice requires development of automated tools that can rapidly compute phenotype probabilities from routine clinical data; thus far, not all metrics considered in this study (e.g. HRV and detailed psychosocial measures) are routinely available in clinical practice, which limit their immediate translation. Nonetheless, such data are increasingly available, as seen in modern wristband wearable devices, for example.⁵⁰ Although the psychosocial scales in our study are feasible to collect in clinical settings, they are not typically part of standard care for this patient group. Future studies should focus on designing a simple survey that integrates the key elements from these scales, along with other relevant psychosocial measures, to facilitate routine clinical implementation. Also, because model-based clustering uses probabilistic rather than discrete assignment, some individuals may have similar posterior probabilities of belonging to more than one phenotype, resulting in a potential ‘grey zone’ of group assignment ambiguity. Additionally, incorporating sleep-related measures, such as sleep duration and quality, may further improve the characterization of biopsychosocial risk profiles in patients with CAD.⁴⁶

Conclusions

We were able to identify four distinct phenogroups amongst a group of stable CAD patients that integrated cardiac autonomic function, myocardial injury marker haemodynamic parameters, and psychosocial factors into clinically meaningful groups. These findings may help facilitate individualized care strategies, especially for behavioural interventions. These findings support a more holistic model that considers psychosocial and behavioural underpinnings to CVD pathogenesis. This model may have significant impact on risk if the right therapies are prescribed, warranting more research to both validate these findings and apply them in clinical trials.

Data availability

The data underlying this article will be shared upon reasonable request to the corresponding author, subject to institutional, ethical, and data protection requirements.

Supplementary material

Supplementary material is available at European Heart Journal Open online.

Funding

We would like to acknowledge the following funding sources: National Institutes of Health/National Heart, Lung and Blood Institute K23 HL127251, R01 HL155711 for A.J.S., National Institutes of Health/National Heart, Lung and Blood Institute R01 HL163998 and National Institutes of Health/National Heart, Lung and Blood Institute R01 HL109413 for V.V., and P01 HL101398 for A.A.Q. L.Y.L. is supported by the National Centre for Advancing Translational Sciences of the National Institutes of Health under Award Number UL1TR002378 and TL1TR002382. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.