NIARE: Noise-Induced Analysis for Model Robustness Evaluation in Asthma Phenotype Classification

Area of Science:

• Computational Biology
• Biomedical Informatics
• Data Science

Background:

• Unsupervised modeling is vital for medical data but vulnerable to noise and outliers.
• Existing validation indices lack insight into noise resistance.
• Supervised adversarial robustness methods are not directly applicable to unsupervised biomedical phenotyping.

Purpose of the Study:

• Introduce Noise-Induced Analysis for Robustness Evaluation (NIARE) to quantify robustness in unsupervised biomedical clustering.
• Address the methodological gap in assessing noise resistance for clinical phenotyping.
• Provide a scalable and interpretable measure of robustness for high-dimensional biomedical data.

Main Methods:

• NIARE augments datasets with synthetic noise-only features from controlled distributions.
• Feature-reduction scoring identifies noise-dominated components for dimensionality selection.
• Clustering scoring assesses cluster stability by measuring the importance of injected noise features.

Main Results:

• NIARE applied to an asthma cohort identified a five-component PCA and five-cluster solution.
• These clusters represent clinically distinct asthma phenotypes with variations in lung function, inflammation, and treatment.
• External validation on the SARP cohort confirmed generalizability and recovery of known asthma phenotypes.

Conclusions:

• NIARE offers a robust and interpretable method for evaluating high-dimensional biomedical clustering.
• It serves as a practical alternative to supervised adversarial methods in unsupervised clinical phenotyping.
• NIARE enhances the reliability of clustering results for clinical interpretation and discovery.