Robust machine learning for Clostridioides difficile typing from mass spectrometry via data augmentation

Machine learning (ML) approaches applied to Matrix-Assisted Laser Desorption Ionization-Time of Flight Mass Spectrometry (MALDI-TOF MS) spectra have shown promise for the typing of Clostridioides difficile, yet their deployment in routine clinical settings remains challenging due to strong sensitivity to acquisition variability. Differences in culture media, incubation time, protein extraction protocols, and instrumentation across hospitals often lead to substantial performance degradation when models are evaluated under heterogeneous or previously unseen conditions. In this work, we systematically analyze the impact of methodological and technical variability on ML-based C. difficile typing and investigate whether data augmentation (DA) strategies can mitigate these effects. Using a dedicated dataset of 60 isolates acquired under diverse conditions, we show that DA substantially improves robustness to variability when training on spectra from selective C. difficile agar media. Importantly, models trained with DA achieve performance levels approaching those obtained using enriched Schaedler agar media, while relying exclusively on standard 24-hour incubation. Evaluation on an independent cohort of 28 newly acquired isolates confirms that DA significantly reduces performance degradation under real-world domain shift. To facilitate adoption and reproducibility, we release MAL-DIDA, an open-source Python library for DA of MALDI-TOF MS spectra.