Gaussian mixture modelling of invasive haemodynamics identifies reproducible phenotypes and risk gradients in a broad spectrum of cardiovascular disease

Background Invasive haemodynamic assessment provides a multidimensional characterisation of cardiovascular physiology in a broad spectrum of cardiovascular disease. Some hemodynamic patterns can be associated with potential success of different treatment modalities and outcomes in specific populations. However, these variables are inherently covariant, limiting conventional interpretation based on isolated parameters or threshold-based definitions and potentially obscuring underlying physiological patterns. Purpose To evaluate whether gaussian mixture modelling (GMM) can leverage physiologically covariant haemodynamic data to uncover reproducible cardiovascular phenotypes with clinical relevance. Methods In 2,836 adults undergoing invasive haemodynamic evaluation in a broad variance of cardiovascular disease, six parameters reflecting preload, right ventricular load, pulmonary circulation, systemic pressure, left ventricular filling and cardiac output were analysed using GMM with tied covariance in a complete case analysis. Model selection integrated information criteria with bootstrap stability (adjusted rand index, ARI). Robustness was assessed through train–test validation, sex-stratified analysis and simulated missingness. Clinical relevance was explored using cumulative incidence analysis and age- and sex-adjusted cause-specific Cox regression. Results A four-phenotype solution demonstrated optimal stability (bootstrap ARI 0.71) with consistent cluster prevalence across validation sets and high assignment confidence (mean posterior probability 0.89). Reproducibility was maintained across sex (ARI 0.71–0.75), simulated 30%-missingness of a single variable with median imputation (ARI 0.68–0.92), and LOVO analysis (ARI 0.34–0.85). Phenotypes comprised a large normodynamic group (n=2164) and three pathological haemodynamic states: congestive preserved-output (n=296), congestive low-output (n=284) and high-output (n=92). Phenotypes aligned with clinically recognisable haemodynamic syndromes. In-hospital mortality increased stepwise across phenotypes (1.2% in reference, 2.7% in congestive preserved-output, 5.6% in congestive low-output, 7.6% in high-output). Compared with the reference phenotype, adjusted hazard ratios for in-hospital mortality were 2.32 (p=0.040) in congestive preserved-output, 5.04 (p<0.001) in congestive low-output, and 5.32 (p=0.001) in high-output. Conclusion Model-based clustering of invasive haemodynamics reveals robust physiological phenotypes that translate into clinically meaningful risk gradients. These findings suggest that gaussian mixture modelling represents a suitable approach for analysing highly covariant and multidimensional clinical data, such as hemodynamic measurements, and may augment conventional interpretation to support data-driven peri-procedural risk stratification in broad variance of cardiovascular disease. External validation is required.