MACHINE LEARNING CLUSTERING FOR BLOOD PRESSURE VARIABILITY: VALIDATION FROM THE SPRINT TO THE HONG KONG COMMUNITY COHORT. (April 2021)

• Record Type: Journal Article
• Authors: Tsoi, Kelvin; Chan, Nicholas; Yiu, Karen; Poon, Simon; Ho, Kendall; Lin, Bryant
• Is Part Of: Journal of hypertension. Volume 39(2021)e-Supplement 1
• Publication Date: 2021-04
• Subjects: Hypertension -- Periodicals; Hypertension -- Periodicals; 616.132005
• DOI: 10.1097/01.hjh.0000745088.97388.2a
• Languages: English
• ISSNs: 1473-5598

Abstract

Abstract : Objective: Visit-to-visit BPV is associated with risks of cardiovascular diseases. Our aim is to investigate the classification methods of patients with varying levels of BPV using different machine learning algorithms. Design and method: Two sets of visit-to-visit blood pressure (BP) readings were extracted from (i) SPRINT in the United States and (ii) BP cohort in Hong Kong (HK). BPV were defined as the mean absolute residuals of regression trends on BP over time. Patients were clustered into low, medium and high levels of BPV with the traditional quantile clustering and five machine learning algorithms, including K-means clustering, Partitioning Around Medoids (PAM), Spectral clustering, Ward's method and Expectation Maximization. Clustering methods were assessed by the Stability Index, and similarities were assessed by the Davies-Bouldin Index (DBI) and the Silhouette Index. Regression models were fitted to compare the risk of stroke. Results: Results: A total of 8, 133 participants were included from SPRINT with the mean BP measurement 14.7 times in 3.28 years of follow-up, and1, 094 participants were recruited from HK cohort with the mean BP measurement 165.4 times in 1.37 years. Quantile clustering assigned one-third participants as the high level of BPV, but machine learning methods only assigned 10 to 27%. Quantile clustering is the most stable method (Stability Index: 0.982 in the SPRINT and 0.948 in the HK cohort) but shown to have certain levels of cluster similarities (DBI: 0.752 and 0.764, respectively) (Table 1). Across the machine learning algorithms, K-means clustering is the most stable method (Stability Index: 0.975 and 0.911, respectively) with lowest similarities on cluster classification (DBI: 0.653 and 0.680, respectively). Patients with high level of BPV under the machine learning models showed stronger association with stroke. Figure. No caption available. Conclusions: Among the machine learning algorithms, K-means clustering shows the most stable and reliable results. One-seventh of the population with high level of BPV had elevated risk of stroke compared to conventional BPV classification. Machine learning can be potentially used in the electronic health system for better patient management.