Self-tuned healthy homogeneous core: Addressing heterogeneities in biomedical datasets

Area of Science:

• Biomedical Machine Learning
• Computational Biology
• Data Science

Background:

• Healthy population heterogeneity can negatively impact disease classification accuracy.
• Decision boundaries between healthy and diseased groups are often distorted by this heterogeneity.
• Existing methods may not adequately address the variability within healthy cohorts.

Purpose of the Study:

• To develop a method for creating a homogeneous healthy reference subset (H²C) from heterogeneous healthy data.
• To demonstrate that training on H²C improves classification performance compared to using the full healthy cohort.
• To provide a classifier-agnostic strategy for healthy cohort curation in biomedical modeling.

Main Methods:

• Decomposition of the healthy cohort into borderline and homogeneous core (H²C) subtypes.
• Utilizing an information-theoretic framework to analyze classification error bounds.
• Developing a healthy-only kernel density estimation with bootstrap-stability-based self-tuning.
• Evaluating the H²C approach on three biomedical datasets: myocardial infarction, arrhythmia, and migraine.

Main Results:

• Training on H²C consistently improved classification performance (F1, precision, recall, accuracy) across all tested datasets compared to the full healthy baseline.
• Significant improvements in mean F1 scores were observed: 0.899 to 0.952 on PTB-XL+, 0.693 to 0.854 on Arrhythmia, and 0.643 to 0.741 on Migraine.
• The H²C method showed the clearest advantage under a balanced training setup and subset-gated evaluation.

Conclusions:

• Curating a stable, homogeneous healthy reference subset (H²C) enhances downstream classification performance in biomedical machine learning.
• The H²C strategy offers a practical approach for constructing smaller, higher-quality healthy reference cohorts.
• This self-tuned, classifier-agnostic method addresses the critical issue of healthy cohort curation in biomedical modeling.