Bringing doctors together

or

A free, private network for verified physicians

Leading voices

Dr. Spencer Dorn

Functional & Motility Disorders · UNC

View

Dr. Saurabh Sethi

General GI · Oja Health

View

Dr. Sara Horst

IBD · Vanderbilt

View

Dr. Nneka Ufere

Liver Transplant · MGH

View

Dr. Wendi LeBrett

General GI · ModernGut

View

Dr. Treta Purohit

General GI · Oshi Health

View

Dr. Laura Sahyoun

IBD · WCM

View

Dr. Uma Mahadevan

IBD · UCSF

View

Dr. Rushabh Modi

General GI · Kaiser Permanente

View

Dr. Brennan Spiegel

General GI · Cedars-Sinai

View
  • Discover your community

  • Stay current

  • A profile that represents you

  • Find your people

About RoonPrivacyTerms

    Overview

    Unsupervised Learning Analysis of Triglycerides, Inflammation, Cholesterol, and the Risks of Incident Cardiovascular Disease and Type 2 Diabetes in the Women's Health Study.

    Sep 5, 2025DOI: 10.1161/jaha.123.039381Journal ArticleAHA

    Written by: Edward K DuranNancy R CookAaron W AdayJulie E BuringPaul M RidkerAruna D Pradhan

    METHODS

    Data will not be made publicly available but will be provided to researchers upon request.Study participants were derived from the WHS (Women's Health Study), a previously completed, randomized double‐blind, placebo‐controlled trial of low‐dose aspirin and vitamin E in the primary prevention of cardiovascular disease in 39 876 US women aged ≥45 years.11 Enrollment began in September 1992 and the trial completed in March 2004, at which time subjects were invited to participate in observational follow‐up, which has continued through 2023. Women eligible for the current study were those who provided adequate baseline blood specimens (n=28 345) with hsCRP concentrations <10.0 mg/L (n=26 443) to account for the skewness of this variable. All participants were continuously followed for the occurrence of incident myocardial infarction (MI), ischemic stroke, coronary revascularization, peripheral artery disease (PAD), and T2D via annual questionnaires. Adjudication of coronary, cerebrovascular, peripheral artery, and T2D endpoints in the WHS have been previously described.12, 13, 14 Events accrued during the initial trial period and observational follow‐up period through January 2017. All participants provided written informed consent, and the study was approved by the Institutional Review Board at Brigham and Women's Hospital.Baseline blood samples were stored in vapor phase liquid nitrogen (−150 °C to −180 °C) and thawed at time of analysis. Measurement methods of triglycerides, hsCRP, LDL‐C (directly measured), and other lipid and lipoprotein biomarkers used in this study have been described in detail.13, 15 Total cholesterol (TC) to high‐density lipoprotein cholesterol (HDL‐C) ratio was equal to TC divided by HDL‐C, whereas non–HDL‐C was equal to the difference between TC and HDL‐C.16We performed an unsupervised learning analysis using k‐means clustering to generate data‐defined (rather than investigator‐defined) exposure group identities for each subject according to their individual aggregate biomarker profiles. In k‐means clustering analyses, groupings (clusters) are mutually exclusive and defined by an algorithm solely on the basis of maximizing within‐group similarity and between‐group differences in continuous variables. Two‐step iterative fitting was performed using a modified expectation–maximization algorithm which refined cluster definitions until stable cluster centers (means) were obtained. We allowed this process to test up to 6 clusters at a time, with optimal clustering number (k) determined by cubic cluster criteria based on minimizing within‐cluster variation.17 Clustering was performed using the “K Means Cluster” platform of JMP Pro 14.0 (SAS Institute, Cary, NC).Baseline concentrations of LDL‐C, hsCRP, and triglycerides were selected as input variables for clustering on the basis that they each have widely available clinical assays and represent 3 divergent atherosclerotic pathways for which pharmacologic therapies have shown clinical benefit in cardiovascular outcome trials. We repeated the unsupervised analysis using varying combinations of other lipid and lipoprotein biomarkers (TC, HDL‐C, apolipoproteins A and B, natural logarithm of hsCRP) as input variables in supplemental analyses. The alternative cluster solutions were also used to evaluate cardiovascular and diabetes risk to determine if risk associations from our main analysis were considerably alerted by inclusion of alternate lipid and lipoprotein biomarkers. Clustering solutions were compared using Cramer's V statistic for strength of association between the main and supplemental clusters. As our focus was on proof of concept for a novel methodology, we did not include other candidate biomarkers that were either unavailable in the entire cohort or had less established risk associations for our outcomes of interest. A schematic diagram depicting our unsupervised learning analysis is shown in Figure 1.Study population consisted of apparently healthy women free of baseline cardiovascular disease enrolled in the Women's Health Study. Data inputs for learning analysis were limited baseline LDL‐C, hsCRP, and triglyceride concentrations. Application of these methods (specifically, iterative k‐means clustering) to minimize within‐group variation and maximize between‐group variation yielded 4 novel exposure groups (cluster identities). Group assignments were mutually exclusive. Resultant profiles value are median (interquartile range). Each person icon represents ≈100 study participants. This methodology was repeated for supplemental analyses using varying combinations of lipids, lipoproteins, and hsCRP. hsCRP indicates high‐sensitivity C‐reactive protein; and LDL‐C, low‐density lipoprotein cholesterol.Baseline characteristics were summarized within each cluster. Continuous variables were reported as median (interquartile range [IQR]) and categorical variables were reported as proportions, and differences across clusters were tested using the Kruskal–Wallis and χ2 tests, respectively.We evaluated the risk of incident CVD according to cluster identity using the Kaplan–Meier method and multivariable adjusted Cox proportional hazards models. Specifically, outcomes evaluated were total CVD (nonfatal MI, nonfatal ischemic stroke, PAD, death from CVD causes, or coronary revascularization), MI (fatal and nonfatal events), and ischemic stroke (fatal and nonfatal events). PAD outcomes were defined as intermittent claudication and peripheral artery revascularization. We also examined risk associations with incident T2D to illustrate that cluster formation was not trained specifically for cardiovascular events. Age‐adjusted models accounted for age and original WHS trial treatment assignment (aspirin, vitamin E), with additional models adjusting for smoking status (current, former, never), continuous body mass index (BMI), baseline diabetes (for CVD models), family history (of MI for CVD models; of diabetes for T2D models), systolic blood pressure, hormonal replacement therapy use (current, former, never/non–hormonal replacement therapy), and weekly exercise (model 2). In all analyses, cluster 1 was used as the reference group for reasons detailed below. These analyses were repeated for supplemental cluster analyses A through D, with the cluster containing lowest (or highest, for HDL‐C and Apo A) values of each biomarker used as reference. Log‐rank test statistic was performed for Kaplan–Meier analyses using an α level of 0.05.We then compared the utility of our novel categorization method to traditional methods of cardiovascular risk stratification, which use established joint thresholds of corresponding biomarkers. Specifically, cluster‐based risk association estimates were compared with those obtained using clinically defined hypercholesterolemia without low‐grade inflammation (LDL‐C ≥160 mg/dL and hsCRP <2.0 mg/L), low‐grade inflammation without hypercholesterolemia (LDL‐C <160 mg/dL and hsCRP ≥2.0 mg/L), and mixed dyslipidemia (HDL‐C <50 mg/dL and triglyceride ≥150 mg/dL). Thresholds applicable to the WHS were selected on the basis of American Heart Association/American College of Cardiology primary prevention population recommended thresholds for abnormal levels for each individual biomarker: LDL‐C ≥160 mg/dL, hsCRP ≥2.0 mg/dL, triglyceride ≥150 mg/dL, and HDL‐C ≤50 mg/dL.18 To highlight that the cluster‐based solutions were mutually exclusive, while traditional classifications were not, we evaluated within‐cluster frequency of participants meeting criteria for the aforementioned biomarker thresholds. We then computed model reclassification and improvement statistics for cluster‐identities versus jointly dichotomized classifications (triglyceride ≥150 mg/dL and HDL‐C <50 mg/dL; LDL‐C ≥160 mg/dL and hsCRP ≥2.0). All analyses were adjusted for model 2 covariates. All statistical analyses were performed in SAS 9.4 for Unix (SAS Institute).

    RESULTS

    Four mutually exclusive clusters emerged and were subsequently named on the basis of their resulting biomarker profiles (Table 1, Figure 2). Cluster 1 (n=12 101, hereafter referred to as the healthy cluster) was defined by lowest biomarker concentrations among the 4 clusters (triglyceride, 88 [IQR, 67–118] mg/dL; hsCRP, 1.1 [IQR, 0.5–2.0] mg/L; LDL‐C, 105.0 [IQR, 90.0–118.1] mg/dL). Cluster 2 (n=7424) was found to have the highest LDL‐C concentrations with a median of 153.7 (IQR, 141.6–170.6) mg/dL and was named the hypercholesterolemic cluster. Median hsCRP concentrations of 6.1 mg/L (5.0–7.6 mg/L) defined Cluster 3 (n=5056, inflammatory cluster), while baseline concentrations of triglycerides defined Cluster 4 (triglyceride‐rich cluster; 340.0 mg/dL [296–412 mg/dL]).Results of Unsupervised Learning Analysis and Baseline Concentrations of Cluster Input Variables in the Women's Health StudyValues shown are median and interquartile range. LDL‐C indicates low‐density lipoprotein cholesterol.Values shown are median and interquartile range. Unsupervised learning analysis (k‐means clustering) using only baseline concentrations of triglycerides, low‐density lipoprotein cholesterol, and high sensitivity C‐reactive protein identified 4 mutually exclusive cluster identities for ostensibly healthy women: cluster 1 (healthy), cluster 2 (hypercholesterolemic), cluster 3 (inflammatory), and cluster 4 (triglyceride‐rich).As shown in Table 2, all baseline characteristics were significantly different across clusters (P<0.001). The healthy cluster was marginally younger relative to the other clusters, which were otherwise identical in age. The healthy cluster also had the lowest BMI and prevalence of cardiovascular risk factors (diabetes, smoking, family history of MI). Current smoking status and family history of MI were comparable in the hypercholesterolemic and inflammatory clusters, while BMI and prevalent T2D appeared marginally higher in the inflammatory cluster. TC to HDL‐C ratio, non–HDL‐C, and apolipoprotein B were lowest in the healthy cluster, while hemoglobin A1c levels did not vary greatly across clusters. In the triglyceride‐rich cluster, risk enhancer prevalence overall was modestly higher than in the healthy cluster, although diabetes was >7 times as prevalent (7.8% versus 1.1%, respectively). Other lipid markers of atherogenic dyslipidemia were highly abnormal in the triglyceride‐rich cluster, with a median TC to HDL‐C ratio of 6.0 (IQR, 5.1–7.0), HDL‐C of 38.6 mg/dL (IQR, 33.1–46.0), and non–HDL‐C of 192.2 mg/dL (IQR, 169.1–220.1).Baseline Characteristics of Initially Healthy Women According to Unsupervised Learning (K‐Means Cluster) AnalysisValues are n (%) for categorical variables and median (interquartile range) for continuous variables. Frequency missing: family history of MI (n=441), smoking status (n=23), hormonal therapy use (n=52), systolic blood pressure (n=276), body mass index (n=21), apolipoprotein B (n=179), apolipoprotein A1 (n=175), HDL‐C (n=1), TC to HDL‐C ratio (n=1), hemoglobin A1c (n=130). Differences in baseline characteristics were tested using a Kruskal–Wallis test (continuous variables) and χ2 test (categorical variables). All baseline characteristics were significantly different across clusters (P<0.001 for all rows). HDL‐C indicates high‐density lipoprotein cholesterol; LDL‐C, low‐density lipoprotein cholesterol; and TC, total cholesterol.In comparing the nonhealthy clusters, cardiovascular risk enhancers did not vary greatly overall, while several differences in lipid profiles were noted. As anticipated, total cholesterol concentrations appeared highest in the cluster defined predominantly by elevated LDL‐C concentrations (hypercholesterolemic cluster), followed by the triglyceride‐rich and inflammatory clusters. In contrast with the triglyceride‐rich cluster, TC to HDL‐C ratios were clinically normal (<5.0) in the hypercholesterolemic and inflammatory clusters. Apolipoprotein B and non–HDL‐C concentrations in the triglyceride‐rich cluster were slightly higher than those in the hypercholesterolemic cluster despite considerably lower LDL‐C concentrations, indicating marked elevations in triglyceride‐rich lipoproteins (TRLs) and their associated cholesterol content by the Friedewald relationship (TRL‐C=non–HDL‐C minus LDL‐C).19, 20, 21The proportion of subjects with biomarker concentrations considered clinically abnormal by the American Heart Association/American College of Cardiology definitions within each cluster is shown in Table 3. Clinical subgroups formed by a singular biomarker threshold were not limited to any particular cluster and could generally have any of the 4 cluster identities. Similarly, subjects with clinically defined isolated hypercholesterolemia, isolated inflammation, or dyslipidemia, were not restricted to having a hypercholesterolemic, inflammatory, or triglyceride‐rich cluster identify, respectively. These distributions varied considerably across clusters, and presence or absence of these clinical subgroups was not specific for any single cluster identity. In contrast, cluster assignments were inherently mutually exclusive.Distribution of Subjects with Clinically Abnormal Lipids and hsCRP Levels According to Cluster IdentityValues are n (% within each cluster). Dyslipidemia is defined as triglycerides ≥150 mg/dL and HDL‐C<50 mg/dL; isolated hypercholesterolemia is defined as LDL‐C ≥160 mg/dL and hsCRP <2.0 mg/L; Isolated inflammation is defined as LDL‐C<160 mg/dL and hsCRP ≥2.0 mg/L. HDL‐C indicates high‐density lipoprotein cholesterol; hsCRP, high‐sensitivity C‐reactive protein; and LDL‐C, low‐density lipoprotein cholesterol.The results of our survival analyses are shown in Figures 3 and 4A through 4C. The incidence of total CVD (Figure 3) was lowest in the healthy cluster, approximately equal in the hypercholesterolemic and inflammatory clusters, and highest in the triglyceride‐rich cluster (Kaplan–Meier cumulative incidence rates: 9.3%, 16.7%, 18.2%, 27.1%; P<0.001). When cardiovascular event subtypes were analyzed individually, this pattern was magnified for incident MI (Figure 4A), with cumulative incidence rates of 2.2%, 4.1%, 5.2%, and 10.2% according to cluster during the 25 years of follow up (P<0.001). The risk of incident PAD (Figure 4B) was also significantly different according to cluster (P<0.001), although the overall cumulative incidence of this outcome was low in this cohort as previously reported.22 The incidence of stroke (Figure 4C) was also lowest in the healthy cluster but did not vary as greatly among the remaining clusters (3.9%, 6.6%, 7.1%, 8.2%; P<0.001).Kaplan–Meier plots depicting the risk of incident total cardiovascular disease. Risk between clusters was significantly different, and subjects with a triglyceride‐rich cluster identity had the greatest risk of total cardiovascular events, followed by those in the inflammatory, hypercholesterolemic, and healthy clusters. P value is for log‐rank test.Kaplan–Meier plots depicting the risk of incident (A) myocardial infarction, (B) peripheral artery disease, and (C) stroke according to cluster identity are shown. For each outcome, subjects a triglyceride‐rich cluster identity had the greatest risk of each outcome, followed by those in the inflammatory, hypercholesterolemic, and healthy clusters. The risk differences for stroke were less pronounced, but all event rates were significantly different. P values are for log‐rank test.Hazard ratios (HRs) and 95% CIs for associations with composite and individual CVD outcomes are shown in Table 4. Age‐adjusted risk of CVD was 1.59 (95% CI, 1.42–1.77) in the hypercholesterolemic cluster, 1.86 (95% CI, 1.65–2.10) in the inflammatory cluster, and 3.02 (95% CI, 2.62–3.47) in the triglyceride‐rich cluster. This association pattern persisted in multivariable adjusted models accounting for model 2 covariates. Associations according to cluster somewhat varied for individual cardiovascular outcomes, with stronger associations for MI and PAD than observed for stroke. The multivariable adjusted HRs for incident MI and PAD in the triglyceride‐rich cluster relative to the health cluster were 3.22 (95% CI, 2.44–4.25) and 2.39 (95% CI, 1.30–4.38), respectively. Consistent with cumulative incidence estimates obtained using the Kaplan–Meier method, the multivariable adjusted risk of stroke was similar between the hypercholesterolemic, inflammatory, and triglyceride‐rich clusters.Cardiovascular Risk Associations According to Cluster Identity in the Women's Health StudyModel 1: age, Women's Health Study treatment assignment. Model 2: model 1+smoking status category, body mass index (continuous), baseline diabetes, family history of myocardial infarction, systolic blood pressure, hormone replacement therapy use category, weekly exercise. Cumulative incidence rates obtained using the Kaplan–Meier estimator. HR indicates hazard ratio.N per cluster in analyses for peripheral artery disease were 12 095 (healthy), 7412 (hypercholesterolemic), 5051 (inflammatory), and 1859 (triglyceride‐rich).Kaplan–Meier curves for incident T2D according to cluster assignment are shown in Figure 5. While the risk of T2D was again lowest in the healthy cluster, an increased risk according to hypercholesterolemic, inflammatory, and triglyceride‐rich cluster can be appreciated: the Kaplan–Meier incidence rate estimates according to cluster were 8.11%, 15.5%, 26.7%, and 37.3%, respectively (log‐rank P<0.001). In age‐adjusted and multivariable Cox models, risk associations persisted (Table 5): model 2 HRs were 1.65 (95% CI, 1.46–1.86) in the hypercholesterolemic cluster, 2.21 (95% CI, 1.95–2.50) in the inflammatory cluster, and 3.78 (95% CI, 3.29–4.35) in the triglyceride‐rich cluster. As with CVD, triglyceride‐rich clusters remained the greatest predictor of incident T2D in our supplemental analysis (Figures SA3 through SD3).Kaplan–Meier plots showing the risk of incident type 2 diabetes is shown. Subjects with triglyceride‐rich cluster identities had the greatest risk of developing type 2 diabetes, followed by those with inflammatory, hypercholesterolemic, and healthy cluster identities. Subjects with baseline diabetes were excluded from this analysis.Risk of Incident Type 2 Diabetes According to Cluster Identity in the Women's Health StudyModel 1: Age, Women's Health Study treatment assignment. Model 2: model 1+smoking status category, body mass index (continuous), family history of diabetes, systolic blood pressure, hormone replacement therapy use category, weekly exercise. Cumulative incidence rates obtained using the Kaplan–Meier estimator.N per cluster shown are for those free of diabetes at baseline.Comparisons with clinical criteria for isolated hypercholesterolemia (LDL‐C ≥160.0 mg/dL and hsCRP <2.0 mg/L), and isolated low‐grade inflammation (LDL‐C <160.0 mg/dL and hsCRP ≥2.0 mg/L), and dyslipidemia (HDL‐C <50.0 mg/dL and triglyceride ≥150.0 mg/dL) are shown in Figure 6. Risk associations according to mutually exclusive clusters were comparable or modestly stronger than those using categorization with traditional thresholds and most pronounced for the triglyceride‐rich cluster. For total CVD, assignment to the triglyceride‐rich cluster conferred higher risk than traditionally defined mixed dyslipidemia, while the inflammatory cluster conferred borderline stronger risk compared to the traditional counterpart. In contrast, the hypercholesterolemic cluster conferred similar risk to traditional the traditional definition. Similar findings were generally observed for MI and PAD, with broader estimated HR due to fewer events. Risk of MI according to the traditional classification for isolated hypercholesterolemia but was comparable to slightly stronger than that using the new cluster‐based definition. For stroke, only the triglyceride‐rich cluster conferred higher risk than established biomarker thresholds. Model improvement and reclassification statistics for total cardiovascular events are included in Table 6. There was a small and nonsignificant improvement in the C‐statistic for clustering‐based classification, which compared jointly dichotomized triglycerides (≥150 mg/dL) and HDL‐C (<50 mg/dL) as well as jointly dichotomized hsCRP (≥2.0 mg/L) and LDL‐C (≥160 mg/dL). Net‐reclassification improvement (categorical and continuous) and integrated discrimination improvement were statistically significant for models including cluster identities compared with models including dichotomized triglycerides and HDL‐C. Findings for cluster identities compared with dichotomized hsCRP and LDL‐C were less robust, as only continuous net reclassification improvement and integrated discrimination improvement were statistically significant. Model improvement statistics individual outcomes can be found in Table S1.Multivariable adjusted hazard ratios and 95% CIs for cardiovascular outcomes according to cluster identity and clinical categories representing dyslipidemia, inflammatory risk, and hypercholesterolemic risk are shown. Triglyceride‐rich and inflammatory cluster identities yielded stronger overall cardiovascular risk associations compared with traditional clinical counterparts. Triglyceride‐rich cluster identity additionally was the strongest overall predictor of future cardiovascular events.Total Cardiovascular Disease Model Improvement Statistics for Cluster Identities Versus Jointly Dichotomized Biomarkers in the Women's Health StudyAll analyses accounted for traditional cardiovascular risk factors included in model 2 (age, Women's Health Study treatment assignment, smoking status category, body mass index [continuous], baseline diabetes, family history of myocardial infarction, systolic blood pressure, hormone replacement therapy use category, weekly exercise). HDL‐C indicates high‐density lipoprotein cholesterol; hsCRP, high‐sensitivity C‐reactive protein; IDI, integrated discrimination improvement; LDL‐C, low‐density lipoprotein cholesterol; and NRI, net‐reclassification improvement.A summary of our supplemental clustering analyses is provided in Table 7 and in the online supplement. Supplemental clusters A through D evaluated various combinations of lipids, lipoproteins, and hsCRP, and clusters were numerically named in the order in which they emerged. Cramer's V test of association showed moderate‐to strong associations with our original clusters, with stronger associations for those including hsCRP. Triglyceride‐rich clusters (median triglyceride >300 mg/dL) with greatest CVD risk also emerged in each of our supplemental cluster analyses using the Kaplan–Meier method (supplemental analyses A through D) and in Cox proportional hazards models adjusting for model 2 covariates (Table 7). The direction and magnitude of association for the triglyceride‐rich supplemental clusters (supplemental clusters A1, B4, C5, D3) were similar to those in our main analysis.Summary of Unsupervised Learning Analyses with Alternate Input Biomarkers and Associated Risk of Total Cardiovascular Disease Events in the Women's Healthy StudyColors are used to help visually summarize results from multiple analyses and correlate to lowest (green) or highest (red) values within each column. Within each of the supplemental cluster analyses (A through D), the A1, B4, C3, and D4 clusters independently emerged as a “triglyceride‐rich cluster” with the highest risk of total cardiovascular disease events, which persisted in Cox regression models adjusting for model 2 covariates. Model 2 covariates: age, Women's Health Study; treatment assignment, smoking category, body mass index (continuous, family history of myocardial infarction, systolic blood pressure, hormone replacement therapy use category, weekly exercise. Apo A indicates apolipoprotein A; Apo B, apolipoprotein B; HDL‐C, high‐density lipoprotein cholesterol; HR, hazard ratio; hsCRP, high‐sensitivity C‐reactive protein; LDL‐C, low‐density lipoprotein cholesterol; NA, not applicable; and TC, total cholesterol.

    DISCUSSION

    In this unsupervised learning analysis clustering subjects on the basis of baseline LDL‐C hsCRP and triglyceride concentrations, we identified 4 clusters defined primarily by the absence (cluster 1, healthy cluster) or predominance of hypercholesterolemia (cluster 2, hypercholesterolemic cluster), subclinical inflammation (cluster 3, inflammatory cluster), or hypertriglyceridemia (cluster 4, triglyceride‐rich cluster). The principal strength of our study is the formation of data‐driven, mutually exclusive cluster identities to examine relative contributions of established biologic pathways to atherosclerotic risk, which has historically proven difficult due to the high degree of biomarker correlation. Subjects with triglyceride‐rich cluster identities had the greatest risk of coronary and limb events and incident T2D in unadjusted Kaplan–Meier and Cox analyses with only modest attenuation when accounting for traditional risk factors. Risk associations in the hypercholesterolemic and inflammatory clusters were comparable with each other for cardiovascular outcomes, while the risk of T2D was greater in the inflammatory cluster than in the hypercholesterolemic cluster. We further showed the potential utility of unsupervised learning‐based categorization vis‐à‐vis stratification using traditional biomarker thresholds and found stronger risk associations for triglyceride‐rich and inflammatory cluster identities. These findings illustrate how data‐driven methods using established, ubiquitous clinical biomarkers can be applied to improve cardiovascular risk stratification and inform pathophysiologic mechanisms of disease.Traditional biomarker‐based epidemiologic studies often use established (eg, hsCRP ≥2.0 mg/L, triglyceride ≥150 mg/dL) or distribution‐based (medians, quartiles, quintiles, etc) thresholds to generate exposure groups, and rely on multivariable analyses to account for correlation or confounding by other risk factors. The purpose of this study was to illustrate the use of an underused methodology, unsupervised learning, to generate novel, unique exposure group (cluster) identities derived from naturally occurring codistributions of routine clinical biomarkers. Our method is similar to established techniques in that it does not rely on training and knowledge of outcome data (ie, blinded to outcomes),17, 23 a feature distinct from machine learning and other supervised learning analyses.5, 24 We sought to efficiently disentangle the relative effects of 3 mechanistically distinct atherosclerotic pathways and used incident T2D to highlight both the strength of an unsupervised (ie, not trained on a specific outcome) analysis and explore underlying pathophysiologic connections given their many shared risk factors.25 Refinement and validation of our approach in contemporary cohorts may yield a powerful tool for the clinician's growing arsenal of cardiometabolic risk evaluation.A small proportion of the WHS study population (n=1862) formed the triglyceride‐rich cluster in whom atherosclerotic risk was greatest and disproportionately elevated compared with baseline concentrations of LDL‐C and hsCRP (Table 4; Figures 3, 4, and 6). Subclinical inflammation is known to associate with hypertriglyceridemia and promote plaque destabilization,26, 27 and was present but to a much lesser degree in the triglyceride‐rich than in the inflammatory cluster, suggesting that inflammation alone cannot account for atherosclerotic risk in these subjects. More likely, elevations in highly atherogenic particles containing apolipoprotein B (TRLs; small, dense low‐density lipoprotein) and their non–HDL‐C content (TRL‐C, small, dense LDL‐C) characteristic of the hypertriglyceridemic state greatly contributed to these findings.28, 29, 30Emerging data continue to indicate that TRL‐C31, 32 (ie, remnant cholesterol) and small, dense LDL‐C33, 34 are strong independent predictors of CVD, and we previously showed risk also associates independent of hsCRP and LDL‐C.16 Evidence for a causal association between TRL‐C and incident MI has also been reported in several genetic Mendelian randomization studies.35, 36 In the current study, it is likely that both biomarkers were highest among those with a triglyceride‐rich cluster identity, with LDL‐C existing primarily in the form of small, dense LDL‐C as is typical of hypertriglyceridemia.28 Low‐normal LDL‐C levels coupled with relative elevations in atherogenic lipoprotein particles and cholesterol (apolipoprotein B, non–HDL‐C) in the triglyceride‐rich cluster also indicate substantially elevated TRL particles and cholesterol content,37 as evidence by invoking Friedewald approximations of TRL‐C (TRL‐C=non‐HDL‐C minus LDL‐C).19, 20, 21 Drivers of the unique signature of atherogenic lipids and lipoproteins portending heightened atherosclerotic risk in the triglyceride‐rich cluster remain unclear and may represent an interplay of lifestyle and genetic factors that warrant further investigation.A key finding in our study was that cardiovascular risk in the inflammatory cluster (median hsCRP, 6.1 mg/L; LDL‐C, 121 mg/dL; triglyceride, 150 mg/dL) was at least comparable, if not slightly greater, than that according to the traditionally defined low‐grade inflammation group (hsCRP ≥2.0 mg/L, LDL‐C <160 mg/dL) and both methods for defining hypercholesterolemia (Table 2, Figures 3 and 4). These data suggest that mechanisms involving the NRLP3 inflammasome, immune cell activation, and release of inflammatory cytokines may play as important a role as those involving LDL‐C.38, 39, 40 Analyses of baseline and on‐treatment hsCRP and LDL‐C (hsCRPOT, LDL‐COT) from recent proprotein convertase subtilisin/kexin type 9 inhibitor cardiovascular outcome trials support this notion. In the SPIRE (Studies of PCSK9 Inhibition and the Reduction in Vascular Events) trials participants with hsCRPOT ≥3.0 mg/L had a 62% increased risk of cardiovascular events compared with those with hsCRPOT <1.0 mg/L despite aggressively reduced LDL‐COT (41.7 versus 29.3 mg/dL, respectively). In the FOURIER (Further Cardiovascular Outcomes Research with PCSK9 Inhibition in Subjects With Elevated Risk) trial, the adjusted cardiovascular event rate for those with baseline hsCRP ≥3.0 mg/L and LDL‐COT <20 mg/dL was comparable with those with baseline hsCRP <1.0 mg/L and LDL‐COT ≥100 mg/dL (13.1% versus 12.3%, respectively). The 81% increased risk of MI, 47% increased risk of stroke, and 67% increased risk of PAD conferred by an inflammatory cluster identity in the present study, together with the above data, underscore the importance of continued efforts to characterize and reduce inflammatory cardiovascular risk.Importantly, application to a noncardiovascular outcome also served an illustrative purpose to highlight the untrained nature of our analysis, which is not optimized for a single outcome. Our study may therefore be relevant to the emerging field of diabetocardiology,41 as risk associations between measures of low‐grade inflammation14, 15, 42 and dyslipidemia of insulin resistance43, 44, 45 with incident T2D are well established in population health studies. Using novel methodology, we corroborate these findings and submit that a predominantly dyslipidemic phenotype (triglyceride‐rich cluster identity) may be a greater marker of risk for T2D than low‐grade inflammation (inflammatory cluster identity). The current cluster identities could be readily optimized for detecting incident T2D through addition or exchanging of relevant exposures, similar to the approach shown in our supplemental analyses. Further, risk associations with future thrombotic, microvascular, or nonvascular outcomes (eg, chronic kidney disease, nonalcoholic fatty liver disease) could be readily explored to inform any common pathophysiology using altered or unaltered cluster identities.There are several limitations and strengths of the current study that merit discussion. First, as our study population was composed primarily of White women from the United States, these results, although internally valid, may not extend to groups with greater sex and racial diversity. Second, low prevalence of statin use at the time of WHS enrollment suggests that cluster recalibration to account for lower biomarker levels may be indicated when scoring contemporary cohorts. Third, hypercholesterolemic cluster identity was not a stronger risk predictor than the traditional categorization method (LDL‐C >160 mg/dL, hsCRP <2.0 mg/L) possibly due to lower LDL‐C concentrations in the former (median, 153.7 [IQR, 141.6–170.6] mg/dL). Fourth, we did not include repeated biomarker measures or advanced lipid and lipoprotein predictors of CVD as input variables in our unsupervised learning analysis, which may have altered our cluster characteristics and risk associations. However, we leveraged a more parsimonious analysis using only widely available routine clinical biomarkers, which may improve generalizability and reproducibility. Further, our supplemental analyses show risk according to triglyceride‐rich cluster was consistently greatest in cluster iterations using other routine lipids and lipoproteins. We additionally illustrate how an unsupervised learning analysis, untrained for a specific outcome, can be used to study and inform risk of a multiple outcomes in a manner not readily accomplished with supervised learning analyses.

    CONCLUSIONS

    This unsupervised learning analysis (clustering) on the basis of aggregate LDL‐C, hsCRP, and triglyceride profiles yielded novel methodologic and pathophysiologic insights into cardiovascular risk stratification by partitioning distinct contributions of hypercholesterolemia, subclinical inflammation, and dyslipidemia to cardiovascular risk. Subclinical inflammation may be as important a risk factor as hypercholesterolemia attributable to LDL‐C, while dyslipidemia defined by triglyceride‐rich cluster identity is a stronger indicator of future cardiovascular events than other cluster identities and traditionally defined dyslipidemia (triglyceride ≥150 mg/dL, HDL‐C <50 mg/dL). Triglyceride‐rich cluster identity was also the strongest predictor of T2D, illustrating how unsupervised learning analyses (ie, not trained on a specific outcome) may be leveraged to study multiple disease outcomes and elucidate shared pathophysiology. Replication in large, completed cardiovascular outcomes trials to examine baseline and on‐treatment cluster identities may better inform cardiovascular risk and improve upon risk stratification methods currently used in clinical practice.

    Sources of Funding

    This work was supported by the National Institutes of Health under Award Number 5T32HL007575‐33 (Dr Duran) and K23 HL151871 (Dr Aday). The Women's Health Study was funded by grants CA047988, HL043851, HL099355, HL080467, and UM1 CA182913. The Women's Health Study was funcing by several grants from the NIH/NHLBI/NCI. Those are the grant numbers listed and are confirmed to be correct by the PI's (Cook Rider, Buring) of the WHS.

    Disclosures

    Dr Ridker reports having received research funding support from multiple not‐for‐profit entities including the National Heart, Lung, and Blood Institute; the National Cancer Institute; the American Heart Association; the Doris Duke Charitable Foundation; the Leducq Foundation; and the Donald W Reynolds Foundation. Dr Ridker also reports having received investigator‐initiated research support from AstraZeneca, Novartis, Pfizer, and Kowa. Dr Ridker is listed as a coinventor on patents held by the Brigham and Women's Hospital that relate to the use of inflammatory biomarkers in cardiovascular disease that have been licensed to Siemens and AstraZeneca, and has served during the past year as a research consultant to Iqvia, Novartis, Corvidia, Inflazome, Amgen, Merck, Sanofi, Civibio, and Janssen, and serves as co‐principal investigator of the PROMINENT (Pemafibrate to Reduce Cardiovascular Outcomes by Reducing Triglycerides in Patients With Diabetes) trial NCT03071692). Dr Aday reports serving as a consultant in the past year to CRC Oncology and Aeglea. Dr Pradhan receives investigator‐initiated research support from Kowa Research Institute and serves as co‐principal investigator of the PROMINENT trial (NCT03071692). The remaining authors report no conflicts.

    Supporting information

    Table S1
    AADr. Ariel Aday

    Dr. Ariel Aday

    Gastroenterology · Functional & Motility Disorders
    The University of Texas Southwestern Medical Center
    arieladay