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

Background: Traditional cardiovascular risk assessment entails investigator-defined exposure levels and individual risk markers in multivariable analysis. We sought to determine whether an alternative unbiased learning analysis might provide further insights into vascular risk.

Methods: We conducted an unsupervised learning (k-means cluster) analysis in the Women's Health Study (N=26 443) using baseline levels of triglycerides, high-sensitivity C-reactive protein, and low-density lipoprotein cholesterol to form novel exposures. We then evaluated cluster-based risk of incident coronary, cerebrovascular, and limb events using the Kaplan-Meier method and multivariable Cox models, followed by comparison with established clinical biomarker thresholds. Finally, we illustrated clinical applicability to a nonvascular outcome (type 2 diabetes).

Results: Four clusters emerged and were named according to aggregate biomarker profiles: Cluster 1 ("healthy," n=12 101), cluster 2 ("hypercholesterolemic," n=7424), cluster 3 ("inflammatory," n=5056), and cluster 4 ("triglyceride-rich," n=1862). Triglyceride-rich cluster identity conferred the highest risk of future cardiovascular events (adjusted hazard ratio [HR], 2.24 [95% CI, 1.93-2.60]) compared with those in the healthy cluster (reference group). Risk was intermediate in the hypercholesterolemic (predominantly elevated low-density lipoprotein cholesterol) and inflammatory (predominantly elevated high-sensitivity C-reactive protein) clusters (HR, 1.44 [95% CI, 1.28-1.61]; and 1.54 [95% CI, 1.35-1.75], respectively). Clustering yielded stronger total cardiovascular disease risk associations than traditionally defined mixed dyslipidemia with modest improvement in reclassification statistics. Cluster identities also predicted incident type 2 diabetes, with the greatest risk among the triglyceride-rich cluster (HR, 3.78 [95% CI, 3.29-4.35]).

Conclusions: Unsupervised learning analyses demonstrated associations that may be useful when refining cardiovascular risk and may inform atherosclerosis development in healthy individuals better than traditional classification methods.

Registration: URL: https://clinicaltrials.gov; Unique identifier: NCT00000479.