Machine learning-driven integration of 24-hour ambulatory blood pressure and its variability.

Twenty-four hour ambulatory blood pressure (BP) monitoring (24-hour ABPM) is considered the best out-of-office BP measurement to assess hypertension. Yet, the complexity of the obtained data, hinders their interpretation. Therefore, in this study, we investigated whether a clustering algorithm could facilitate the integrative interpretation of 24-hour ambulatory BP and its variability to construct distinctive clusters associated with cardiovascular (CV) outcomes. In 1344 community-dwelling individuals (mean age, 46.9 years; 50.6% women), we acquired 24-hour ABPM recordings, clinical data at baseline and we collected adverse outcomes (median follow-up time 19.4 years). We used dynamic time warping combined with the k-medoids clustering algorithm on the heart rate, systolic and diastolic BP recordings. Clinical characteristics and CV outcome were used to evaluate the validity of the derived clusters using the original cohort, while an external general population cohort (n = 1219) was used for further validation. We separated the dataset into 4 clusters. Cluster 1 showed the most favourable CV risk profile comprising younger individuals with the lowest 24-hour ABPM patterns. On the contrary, cluster 4 revealed the worst CV profile with older participants, higher office and 24-hour ABPM, higher medication intake and the highest BP variability. After adjusting for traditional risk factors, including office systolic BP cluster 4 had significantly higher risk of adverse CV events compared to cluster 1 (HR: 1.63; 95%CI:1.15-2.32; P = 0.006). Similar findings were made in the external validation cohort. Time series clustering of ABPM recordings may facilitate the integrative interpretation of 24-hour ambulatory BP and BP variability.