**Background:** Hypertension is a leading cause of death worldwide, and RAAS-targeting antihypertensive drugs (ACEi and ARB) are central to treatment. Population-based studies offer opportunities to assess real-world treatment effectiveness but often lack high-quality drug documentation, especially when electronic health record linkage is unavailable. Self-reported treatment is imprecise and subject to classification bias. Recent advances in LC-MS/MS allow simultaneous measurement of RAAS biomarkers (angiotensin I, angiotensin II, aldosterone) in biobanked samples. This study investigated whether unsupervised cluster analysis of these biomarkers could identify undertaken antihypertensive treatments in the general population.
**Methods:** The study included 800 participants from the CHRIS population-based study (baseline 2011-2018). Participants were sampled across 8 age- and sex-matched groups: normotensive, untreated hypertensive, non-AHD drug users, ACEi monotherapy, ACEi+diuretics, ARB monotherapy, ARB+diuretics, and beta blocker monotherapy. RAAS biomarkers were measured via LC-MS/MS (RAAS Triple A testing, Attoquant Diagnostics). Lower limits of quantification were 5 pg/ml for each biomarker. K-means unsupervised clustering was applied to log-transformed biomarker values. Optimal cluster number was determined by Silhouette score. Agreement between clusters and drug classification was assessed using weighted kappa (κ_w), sensitivity, and specificity. Lasso penalized regression identified clinical characteristics associated with each biomarker after accounting for cluster and treatment effects.
**Key Results:** Three well-separated clusters were identified: Cluster 1 (n=444, 55%) preferentially included individuals not receiving RAAS-targeting drugs (normotensive, hypertensive, non-AHD, beta blockers). Cluster 2 (n=235, 30%) identified ARB users (κ_w=74%; sensitivity=73%; specificity=83% for ARB monotherapy; for ARB+diuretics: κ_w=51%, sensitivity=69%, specificity=76%). Cluster 3 (n=121, 15%) identified ACEi users (κ_w=82%; sensitivity=55%; specificity=90% for ACEi monotherapy; for ACEi+diuretics: κ_w=78%, sensitivity=67%, specificity=92%). When combining ACEi groups, cluster 3 showed κ_w=82% (95%CI: 79-86%), sensitivity=61% (95%CI: 54-68%), specificity=100% (95%CI: 99-100%). When combining ARB groups, cluster 2 showed sensitivity=71% (95%CI: 64-77%), specificity=84% (95%CI: 81-87%). Cluster 1 had 85% sensitivity and 74% specificity for identifying non-RAAS users. Individuals in clusters 2 and 3 had higher BMI (29.0 and 29.3 vs. 27.4 kg/m², p=0.0021), higher fasting glucose (104.38 and 102.48 vs. 98.76 mg/dl, p=4.42×10⁻⁵), higher HbA1c (6.04% and 6.04% vs. 5.91%, p=1.05×10⁻⁶), and higher diabetes prevalence (22% and 21% vs. 9%, p=1.45×10⁻⁵) compared to cluster 1. Cluster 3 had lower eGFR (73.55 vs. 78.34 ml/min/1.73m², p=0.0070). Lasso regression revealed that after removing cluster and treatment effects, age, sex, eGFR, DBP, and diabetes were still associated with angiotensin I; age, sex, eGFR, and SBP with angiotensin II; and sex, BMI, eGFR, and cortisol with aldosterone.
**Clinical Implications:** Unsupervised clustering of angiotensin-based biomarkers can reliably identify individuals on ACEi (near-perfect specificity of 100% for combined ACEi groups) and, to a lesser extent, ARB users in population-based settings. This technique could help reconstruct likely AHD treatment a posteriori in studies lacking electronic health record linkage, reducing classification bias. The imperfect sensitivity (61% for ACEi, 71% for ARB) suggests that factors such as treatment non-adherence, drug dosage, and individual biological variability affect biomarker levels. The association of age, sex, and kidney function with biomarkers independent of treatment highlights the need to account for these factors when using biomarkers for treatment classification. The method is limited by cross-sectional measurement, inability to assess treatment adherence, and lack of urinary sodium data. Independent replication and calibration of the clustering algorithm in other settings is warranted before clinical application.