Toward Robust Machine Learning Models for MALDI-TOF MS: Novel Approaches for Mycobacterium abscessus Subspecies Identification

Distinguishing Mycobacterium abscessus subspecies presents significant diagnostic challenges due to their
genetic homogeneity and variability in analytical platforms. Our research
combines matrix-assisted laser desorption/ionization time-of-flight
(MALDI-TOF) mass spectrometry with machine learning (ML) approaches
to enhance discrimination accuracy, utilizing 325 spectra profiles
from diverse European hospitals. The analytical pipeline incorporates
specialized techniques for geographical data harmonization, feature
selection, and balancing class representation. The best model employs
support vector machines (SVMs) with ComBat correction, Boruta feature
selection, and centroid clustering for class imbalance, achieving
a discrimination performance of 97% F1 score and 97.17% AUC-ROC on
test samples. Noteworthily, most tested models improved their discrimination
performance with the approach and demonstrated consistent performance
metrics with high geometric mean (GEO) and index balanced accuracy
(IBA) metrics (>0.90), ensuring consistent sensitivity and specificity
across all subspecies. SHAP (SHapley Additive exPlanations) validated
the biological relevance of selected spectral features, particularly
improving discrimination of the diagnostically challenging M. abscessus subsp. bolletii. This work advances the state-of-the-art in M. abscessus classification, providing a scalable
analytical framework for enhanced microbial diagnostics and targeted
antimicrobial therapy selection.