2025 Volume 20 Issue 2 Pages 24-00325
Heart failure (HF) is a significant global public health issue, and accurate diagnosis and effective management are crucial for improving patient outcomes. This study aims to identify potential HF patient subgroups based on the pulse wave time series via unsupervised clustering algorithms and to quantify the importance ratio (IR) of pulse wave morphological features within these subgroups. We collected and normalized pulse wave time series and clinical characteristics from 380 HF patients, which were clustered by introducing the K-means++ algorithm and the clustering performance was assessed along with the clinical characteristic differences between clusters. We then extracted time-frequency features from the pulse waves, analyzed the differences in these features between clusters, and quantified their IRs using a Random Forest classifier. Our results show the optimal clustering performance when the number of clusters is 2, with Silhouette coefficient, Calinski-Harabasz index, and Davies-Bouldin index values of 0.74, 585, and 0.75, respectively. Noticeable differences were observed between clusters in terms of age, heart rate (HR), and ejection time (ET), resulting in two HF patient subgroups: Cluster 0 (47% elderly patients; 34% with tachycardia; 54% with low ET) and Cluster 1 (57% elderly patients; 23% with tachycardia; 46% with low ET). Additionally, through the Random Forest classifier, it was found that upstroke time, mean, downstroke time, and skewness were the significant important pulse wave morphological features, with IRs of 20.8%, 18.7%, 17.1%, and 10.1%, respectively. This study is the first to apply the K-means++ algorithm for unsupervised clustering of pulse wave time series in HF patients, successfully identifying two patient subgroups and revealing significant differences in age, HR, and ET between the clusters. The findings provide preliminary evidence for stratified management of HF patients using non-invasive and easily accessible pulse wave signals.
