Explainable hierarchical clustering for patient subtyping and risk prediction

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Data Analysis

Background:

• Accurate patient stratification is crucial for effective hospital management and predicting clinical deterioration.
• Existing scoring systems like the National Early Warning Score 2 (NEWS2) provide a general patient overview but may lack granularity.
• The integration of machine learning offers potential for more nuanced patient subtyping.

Purpose of the Study:

• To develop and evaluate a machine learning pipeline for automated identification and clinical interpretation of hospital patient subtypes.
• To compare the predictive performance of identified patient subtypes against the established NEWS2 scoring system.
• To explore the synergy between machine learning-driven subtyping and clinical expertise.

Main Methods:

• Utilized routinely collected hospital data (2017-2021) from a UK teaching hospital.
• Employed iterative, hierarchical clustering to identify key features for patient stratification.
• Applied explainability techniques for clinical interpretation of subtypes and trained outcome prediction models for each cluster.

Main Results:

• Identified distinct patient subtypes with clinically meaningful interpretations, validated by clinicians.
• Outcome prediction models for patient subtypes demonstrated superior forecasting of patient deterioration compared to NEWS2.
• Showcased the robustness of identified subtypes through explainability techniques and clinician assessment.

Conclusions:

• Machine learning-based patient subtyping can significantly enhance the prediction of patient deterioration.
• Combining computational analysis with clinical expertise offers a powerful approach to patient stratification.
• This methodology holds promise for improving personalized patient management and clinical decision-making.