Blood Pressure and Heart Rate Patterns Identified by Unsupervised Machine Learning and Their Associations with...

We applied unsupervised machine learning to analyze blood pressure (BP) and resting heart rate (HR) patterns measured during a 1-year period to assess their cross-sectional relationships with subclinical cerebral and renal target damage. Dimension reduction via uniform manifold approximation and projection, followed by K-means++ clustering, was used to categorize 362 community-dwelling participants (mean age, 56.2 years; 54.9% women) into three groups: Low BP and Low HR (Lo-BP/Lo-HR), High BP and High HR (Hi-BP/Hi-HR), and Low BP and High HR (Lo-BP/Hi-HR). Cerebral vessel lesions were defined as the presence of at least one of the following magnetic resonance imaging findings: lacunar infarcts, white matter hyperintensities, cerebral microbleeds, or intracranial artery stenosis. A high urinary albumin-to-creatinine ratio (UACR) was defined as the top 10% (≥ 12 mg/g) of the mean value from ≥2 measurements. Poisson regression with robust error variance, adjusted for demographics, lifestyle, and medical history, showed that the Hi-BP/Hi-HR group had relative risks of 3.62 (95% confidence interval, 1.75-7.46) for cerebral vessel lesions and 3.58 (1.33-9.67) for high UACR, and the Lo-BP/Hi-HR group had a relative risk of 3.09 (1.12-8.57) for high UACR, compared with the Lo-BP/Lo-HR group. These findings demonstrate the utility of an unsupervised, data-driven approach for identifying physiological patterns associated with subclinical target organ damage.