Abstract
This paper evaluates non-negative matrix factorisation and Gaussian mixture modelling for soft classification of kidney stones from FTIR spectral data. The clinical motivation is that accurate stone composition analysis supports treatment planning and recurrence prevention.
Authors
Aistis Raudys, Aušra Šubonienė, Arūnas Želvys
What the study evaluates
The study focuses on rare or mixed stone compositions, where data scarcity can limit neural-network approaches. It examines whether NMF can provide a useful low-dimensional representation and whether GMM can express uncertainty through soft class assignments.
Main findings
The combination of NMF and GMM provides a practical route for composition analysis under limited-sample conditions. The method reaches meaningful concordance with reference labels while preserving a probabilistic view of ambiguous cases.
This is especially relevant for medical datasets where rare classes are clinically important but underrepresented in training data.
Why it matters
Medical AI methods often need to work with small, imbalanced, and noisy datasets. This study shows how interpretable statistical methods can remain useful in such settings, particularly when uncertainty is part of the clinical decision process.