Multiscale classification of heart failure phenotypes by unsupervised clustering of unstructured electronic medical...

Area of Science:

• Cardiology
• Medical Informatics
• Computational Biology

Background:

• Heart failure (HF) is a major cause of death and healthcare costs globally.
• Current HF subtyping methods lack precision due to diverse patient factors, limiting clinical decisions and therapy development.
• A data-driven approach is needed to better characterize real-world HF patient heterogeneity.

Purpose of the Study:

• To develop and validate a novel, data-driven methodology for characterizing heart failure patient subpopulations.
• To leverage natural language processing (NLP) and clustering on electronic health record (EHR) data to identify distinct HF patient groups.
• To uncover clinically interpretable subgroups based on patient complaints and manifestations.

Main Methods:

• Utilized NLP to process unstructured clinical notes from EHRs, extracting disease and symptom concepts (complaints).
• Constructed vectorized representations of patient complaints.
• Applied clustering algorithms to group HF patients based on complaint vector similarity and identified enriched complaints within each cluster via statistical testing.

Main Results:

• Successfully clustered heart failure patients into clinically interpretable subgroups based on complaint patterns.
• Identified known HF etiologies, risk factors, and comorbidities (e.g., ischemic heart disease, hypertension, diabetes) within the discovered subgroups.
• Provided deeper insights into the specific clinical manifestations of HF within each identified subgroup.

Conclusions:

• The novel NLP and clustering approach effectively characterizes real-world heart failure patient heterogeneity.
• This hypothesis-free methodology can reveal clinically meaningful HF subgroups and inform personalized treatment strategies.
• The approach is adaptable for discovering novel insights in other diseases and patient populations.