Cluster analysis in patients with heart diseases: the cornerstone for personalized and tailored care
Year: 2025
Type: article
Abstract: This invited commentary refers to ‘Identifying and preliminary validating patient clusters in coronary artery bypass grafting: integrating autonomic function with clinical and demographic data for personalized care’ by P. Singh et al. https://doi.org/10.1093/eurjcn/zvaf059 Postoperative care plays a vital role in facilitating recovery and enhances cardiovascular function in patients undergoing heart surgery.1 Although coronary artery bypass grafting has proven benefits for a significant proportion of patients with coronary artery disease,2 heterogeneity in patients’ demographical, clinical, physical, and mental characteristics complicates precision decision-making and personalized postoperative care, resulting in various disease outcomes.3 However, in the face of such a large disease population, thoroughly understanding each individual’s features and tailoring a customized care plan for each person is time-consuming and labour-intensive. Cluster analysis grants us insights to solve this problem. Cluster-informed tailored care may be an effective way to improve patients’ postoperative outcomes and support the transition toward personalized cardiovascular care. Singh and coworkers4 published an article in the European Journal of Cardiovascular Nursing on identifying and preliminary validating patient clusters in coronary artery bypass grafting based on autonomic function, demographic, and clinical data. In that paper, dimensionality reduction and data clustering structure visualization were performed by t-distributed stochastic neighbour embedding (t-SNE), while hierarchical clustering was used to determine the optimal number of clusters. Ultimately, two distinct clusters and predictors of complications were identified, enhancing patient monitoring and cluster-based tailored intervention to improve patient outcomes. While some traditional and more straightforward methods such as the Society of Thoracic Surgeons risk score and the European System for Cardiac Operative Risk Evaluation are used for risk stratification in cardiac surgery patients, these methods still exist certain limitations, including less robust calibration of the model.5 Singh’s study utilized advanced clustering techniques to recognize patient portraits based on a set of variables.4 Cluster analysis is a preferred approach for establishing homogenous subclassifications and has been used widely in many research to help understand patient heterogeneity. In Singh paper, t-SNE was used to simplify the variables information to optimize the clustering process. T-SNE is an unsupervised machine learning method for visualizing high-dimensional data structures. Compared with other dimensionality reduction techniques, such as principal component analysis, t-SNE is superior in capturing the non-linear relationships between variables.6 In t-SNE analysis, two-dimensional components were visualized in a scatterplot to promote an intuitive understanding of the data structure and relationships. These components were subsequently utilized in hierarchical cluster analysis to identify distinct groups.6 In hierarchical clustering, the number of clusters was selected based on the silhouette analysis and dendrogram inspection. The silhouette plot measures how similar each patient is to their cluster compared to others, with higher silhouette widths indicating better-defined clusters.7 Meanwhile, the dendrogram shows the hierarchical arrangement of patients with distinct branches representing each cluster. The number of clusters is determined by how tall this tree is. However, hierarchical clustering does not indicate the best trimming height for the resulting dendrogram.8 To address this, silhouette plots and other metrics like cubic clustering criterion, C-index, and SD-index should be employed to confirm cluster selection.8 In addition to hierarchical cluster analysis, there are also some other unsupervised machine learning methods used for clustering such as K-means clustering, Self-Organizing Maps (SOM), density-based clustering, etc. Each method has its advantages and disadvantages. Hierarchical clustering is easy to implement, but it is limited in large datasets. K-means clustering is computationally faster than any unsupervised machine learning algorithm, but it cannot be used in non-linear data and requires a pre-determined number of clusters.9 Density-based clustering is effective in dealing with noise and outliers, but it has limited use in data with variable densities.9 SOM and t-SNE are useful for visualizing high-dimensional data and dimensionality reduction, but SOM calculations are more complex and require longer training time.10 Above all, the choice of clustering method should be depended on the type and size of the dataset, the study objective, and computational expertise available. Patient-specific interventions should be conducted to improve patients’ postoperative outcomes, for that, healthcare professionals should identify patients’ characteristics for an in-depth understanding of their needs. However, without prior knowledge of this disease group, this task remains challenging. Cluster analysis offers a solution by detecting unknown patterns in this population, and patients belonging to the high-risk group which is associated with worse outcomes should receive more attention and intervention.11 Furthermore, different targeted interventions can be given according to different patient patterns. For instance, as shown in the Singh paper, patients in the high-risk response group who exhibited worse autonomic nervous systems should receive more comprehensive and tailored interventions to manage autonomic function, so as to promote postoperative cardiac rehabilitation.4 Clustering is an unsupervised machine learning method that analyzes unlabelled data in multidimensional feature space to identify patient clusters with similar features. However, a key challenge arises when new patient data become available—how should these patients be classified into existing clusters? It would be impractical to repeat the entire clustering process each time a new patient is added. Machine learning modelling methods provide new ideas to solve this problem. Once patient clusters have been established, they can serve as pre-defined class labels for supervised machine learning models to predict subgroup for new patients.12 By combining unsupervised machine learning methods with supervised machine learning methods, researchers can pave the way for real-time patient cluster recognition as well as fully realizing the potential of personalized care. Bei Dou [(Writing—original draft: Lead); RN MSc] and Philip Moons (Writing—review and editing: Lead) B.D. is funded by a grant of the China Scholarship Council. No data were used in this paper.
Authors: Bei Dou, Philip Moons
Institutions: KU Leuven, University of Cape Town, University of Gothenburg
Language: English
Subfield: Surgery
Field: Medicine
Domain: Health Sciences
Open Access status: bronze
Funder: China Scholarship Council
OpenAlex ID: W4410079664