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

Abstract

Robust unsupervised modeling is critical in medical applications where high-dimensional noise and outliers can destabilize clustering and mislead clinical interpretation. However, existing internal validation indices- such as the Silhouette score, Davies-Bouldin index, and Calinski-Harabasz index- mainly quantify cohesion and separation, providing limited insight into noise resistance. In addition, adversarial robustness techniques developed for supervised learning cannot be directly applied to unsupervised biomedical phenotyping, leaving a critical methodological gap. We introduce Noise-Induced Analysis for Robustness Evaluation (NIARE), a framework that quantifies robustness by augmenting datasets with synthetic noise-only features drawn from controlled distributions. NIARE operates in two modes: (i) feature-reduction scoring, which identifies the onset of noise-dominated components to guide dimensionality selection, and (ii) clustering scoring, which measures the importance assigned to injected noise features relative to real features to assess cluster stability. Applied to a high-dimensional asthma cohort of 1,112 patients across 3,471 clinical visits, NIARE selected a five-component PCA representation and a five-cluster operating point, yielding clinically distinct asthma phenotypes characterized by differences in lung function, inflammatory markers, oral corticosteroid burden, and visit frequency. External validation on the Severe Asthma Research Program (SARP) cohort further supported generalizability by recovering the established five-phenotype asthma structure without requiring ground-truth labels. Compared with conventional validation indices, NIARE provides a scalable and interpretable measure of robustness for highdimensional biomedical clustering, offering a practical alternative to supervised adversarial methods in unsupervised clinical phenotyping.