Deep Clustering of Electronic Health Records Tabular Data for Clinical Interpretation

Area of Science:

• Medical informatics
• Machine learning in healthcare
• Unsupervised learning for patient stratification

Background:

• Supervised machine learning relies on labeled data, which is often scarce and costly to obtain in medicine.
• Traditional unsupervised clustering methods like k-means offer limited insights for clinical applications requiring deeper patient data analysis.
• Existing clustering evaluations focus on accuracy, not clinical interpretability or data homogeneity.

Purpose of the Study:

• To propose a novel patient stratification strategy using clinical variables, bypassing the need for diagnostic labels.
• To establish robust evaluation metrics for clustering performance, focusing on within-cluster homogeneity and between-cluster statistical differences.
• To compare traditional clustering algorithms with a deep learning approach for tabular clinical data.

Main Methods:

• A deep learning-based clustering solution was developed for tabular patient data.
• Clustering performance was evaluated using metrics measuring within-cluster homogeneity and between-cluster statistical separation.
• The deep clustering method was compared against traditional algorithms like k-means.

Main Results:

• The deep clustering method demonstrated superior within-cluster homogeneity and between-cluster separation compared to k-means.
• Three statistically distinct and clinically interpretable patient clusters were identified for high blood pressure.
• The proposed strategy successfully stratified patients without relying on explicit diagnostic labels.

Conclusions:

• The developed deep clustering strategy and evaluation metrics enable effective patient stratification in large cohorts.
• This approach facilitates health science research by leveraging unlabeled clinical data for deeper patient insights.
• The findings highlight the potential of deep learning for advancing unsupervised patient data analysis in clinical settings.