Identifying novel subgroups in heart failure patients with unsupervised machine learning: A scoping review

Jin Sun, Hua Guo, Wenjun Wang, Xiao Wang, Junyu Ding, Kunlun He, Xizhou GuanFrontiers in Cardiovascular Medicine 2022 · DOI

TL;DR

Heart failure (HF) is a complex condition with diverse underlying causes, making current broad classifications insufficient for personalized treatment. This scoping review synthesizes recent research using unsupervised machine learning (ML) to identify novel HF subgroups, revealing a growing trend in this area and highlighting key methodological choices and knowledge gaps for future clinical application.

What Problem Does This Paper Address?

Heart failure (HF) is a major global health challenge, characterized by high mortality and readmission rates. Current classifications, primarily based on ejection fraction (EF), fail to capture the significant heterogeneity in HF's clinical manifestations, progression, and treatment responses. This lack of granular understanding hinders the development of effective, personalized treatment strategies. Researchers are increasingly turning to unsupervised machine learning, particularly clustering analysis, to uncover hidden patterns and define more precise patient subgroups that could lead to better diagnosis and management. This paper addresses the need to systematically review and synthesize the rapidly growing body of literature employing these ML techniques for HF subtyping.

Methods

This study is a scoping review, following the PRISMA-ScR guideline, to map the existing research landscape. The authors searched the PubMed database for studies that defined new HF subgroups using clustering analysis methods. Exclusion criteria included reviews, commentaries, editorials, studies not focused on defining new subtypes, or those not employing unsupervised algorithms. Data extraction covered study characteristics, population, clustering methods, variables, and outcomes. Data synthesis involved descriptive statistics and visualization.

Datasets

The paper does not introduce new datasets. Instead, it analyzes data from 47 previously published studies. These studies utilized various data sources, including Electronic Health Records (EHR) (51.1%), Randomized Controlled Trial (RCT) data (19.1%), disease registries (17.0%), observational data (6.4%), and claims data (2.1%). Some studies combined multiple data types. The sample sizes varied significantly, with a median of 480 patients, but a wide range from 63 to over 318,000.

Key Insights and Ideas

  • The ML wave in HF is here to stay: The number of publications identifying novel HF subgroups using ML has surged, with a significant portion published in 2020 and later, indicating a strong and growing research interest.
  • Comorbidities are king (for now): Comorbidity data is the most frequently used variable type for clustering (61.7%), likely due to its accessibility and known impact on HF. This highlights a practical approach but also points to a potential underutilization of other rich data sources.
  • Hierarchical clustering leads the pack: Hierarchical clustering was the most common ML method (46.8%), offering flexibility by not requiring a predefined number of clusters, which is advantageous for exploratory subgroup discovery.
  • Geographic bias exists: The majority of studies originate from the United States and Europe, raising concerns about the generalizability of identified subgroups to populations in other regions with different demographic and health profiles.
  • The "black box" problem persists: While ML can identify subgroups, the interpretability of why these subgroups are distinct and their precise clinical implications are not always fully explored, limiting immediate clinical translation.

Key Results

  • 47 studies included: Out of 498 initially identified records, 47 studies met the inclusion criteria.
  • Publication trend: 61.7% of the studies were published in 2020 or later.
  • Geographic distribution: The United States contributed the largest number of studies (46.8%), followed by Europe. Most studies used data from a single country.
  • Most common ML method: Hierarchical cluster analysis (46.8%).
  • Most common variable type: Comorbidity (61.7%).
  • Least common variable type: Genomics (12.8%).
  • Sample size distribution: 48.9% of studies used datasets with less than 500 patients.
  • Outcome assessment: 85.1% of studies assessed the association between identified clusters and at least one outcome, with death (76.6%) and hospitalization (57.4%) being the most frequent.
  • External validation is rare: Only 12.8% of studies reported external validation of their identified subgroups.

Limitations and Future Work

The authors acknowledge several limitations:

  • Limited database search: Only PubMed was searched, potentially missing relevant studies from other databases.
  • Language bias: Only English-language publications were included, potentially excluding diverse perspectives.
  • Lack of external validation and evaluation criteria assessment: Many studies did not rigorously validate their subgroups or clearly define evaluation criteria, hindering clinical adoption.
  • Descriptive nature: This scoping review provides a comprehensive overview but does not perform a formal methodological quality assessment or meta-analysis.

Future research directions suggested include:

  • Multinational collaborations: Combining datasets from different countries to improve the generalizability of identified subgroups.
  • Focus on genomics and proteomics: While currently underutilized, these data types hold great potential for revealing deeper insights into HF prognosis.
  • Rigorous validation: Future studies should prioritize external validation and clear evaluation of subgroup criteria to facilitate clinical translation.
  • Prognostic implication exploration: More studies should clearly define clinical endpoints and evaluate the prognostic implications of identified subgroups, potentially including quality of life or daily behavior as outcomes.
  • Systematic review and meta-analysis: The authors plan to conduct a more in-depth systematic review and meta-analysis on studies with similar subgroup definitions.

Summary updated

Abstract

This scoping review provides an overview of recent studies proposing novel HF subgroups based on clustering analysis. Differences were found in study design, study population, clustering methods and variables, and outcomes of interests, and we provided insights into how these studies were conducted and identify the knowledge gaps to guide future research.
BibTeX
@inproceedings{sun2022identifying,
  author = {Jin Sun and Hua Guo and Wenjun Wang and Xiao Wang and Junyu Ding and Kunlun He and Xizhou Guan},
  title = {{Identifying novel subgroups in heart failure patients with unsupervised machine learning: A scoping review}},
  booktitle = {Frontiers in Cardiovascular Medicine},
  year = {2022},
  doi = {10.3389/fcvm.2022.895836},
}
Figure