Latent Mixture Modeling to Identify Homogeneous Subgroups Predictive of Cerebrovascular Reactivity

Abstract

IntroductionPrecision medicine is a growing topic of research within the medical community. However, the complex factors that influence disease progression and treatment efficacy often make the identification of small subgroups within large clinical populations challenging. Without strategies to identify subgroups that differentially respond to intervention, researchers and clinicians could overlook treatments that may benefit specific aspects within sizable clinical populations. Cerebrovascular disease results from the interaction of multiple risk factors, leading to varying pathologies including stroke and dementia. Identifying subgroups within the scope of cerebrovascular dysfunction risk will allow researchers to better identify individuals at risk for cerebrovascular disease and ultimately inform optimal treatment strategies. Latent mixture modeling (LMM) presents a novel, unbiased method to identify small homogeneous subgroups that have similar traits within the context of large heterogeneous cohorts. Using LMM, it is possible to detect the presence of latent categorical variables which correspond to unobserved subgroups. To assess the feasibility of LMM as a technique for forming subgroups that predict cerebrovascular health, we retrospectively analyzed data from 78 individuals including 12 measures of physical, metabolic, and cognitive health. Our hypothesis was that latent mixture modeling would identify subgroups formed around similar characteristics of physical, metabolic, cognitive, and psychological health. Further, we hypothesized that these subgroups would display differences in cerebrovascular reactivity (CVR).MethodsData from previous experiments conducted in our laboratory between 2018‐2020 were organized and prepared for analysis. Health measures included: body mass index, systolic blood pressure, heart rate, non‐HDL cholesterol, white blood cell count, albumin/globulin ratio, blood glucose, serum creatinine, total recall memory score, and patient‐reported perception of fear and anxiety. CVR was measured by transcranial Doppler ultrasound and calculated as the percent change (from baseline) in middle cerebral artery blood velocity (MCAv) per mmHg following a forced +9mmHg increase in end‐tidal CO2. LMM was used to identify latent classes based on the selected health measures, and then to assess whether CVR differed among the identified classes. The number of classes was picked based after examining a number of criteria including: Akaike, Bayesian, Sample Size Adjusted Bayesian Information Criteria, Entropy, and the VLMR and adjusted VLMR likelihood ratio tests. The BCH procedure was used for post‐hoc comparisons among classes.ResultsWe identified 4 homogeneous subgroup classes within our data. CVR was found to be significantly different when comparing class 1 and class 2 (p= 0.012) and there was a trend towards significance when comparing class 2 and class 4 (p= 0.063). These data suggest that the interaction of healthspan indicators influence differences in cerebrovascular endothelial function.