The data science team led by Dr. Kosmas Kepesidis has published a new study addressing one of the most persistent challenges in the field: ensuring that machine learning models maintain their accuracy when applied across different infrared spectroscopy devices.
The paper, released on May 7, presents an approach to cross-device model generalization in blood-based infrared spectroscopy — an increasingly valuable technique for non-invasive blood analysis. A major obstacle to deploying such models in real-world clinical or research settings is the variability between devices, which can dramatically hinder predictive performance when a model trained on one instrument is applied to another.
To address this, the team employs a domain adaptation strategy grounded in data augmentation. Their method involves simulating device-specific variations within the training dataset, effectively teaching the model to accommodate a broader range of signal discrepancies. This simple yet effective approach significantly improves the model’s flexibility and robustness across different Fourier-transform infrared (FTIR) spectroscopy devices.
The researchers validated their approach through experimental testing on blood plasma spectra obtained from two FTIR instruments located in separate research laboratories. The results demonstrated notable improvements in prediction accuracy and reliability, underscoring the method’s practical applicability in multi-device environments.
“This work provides a crucial step toward making machine-learning models trained on spectroscopic data more transferable and robust,” said Dr. Kepesidis. “As we aim to scale up the use of these models, particularly in large population studies and clinical settings, device independence becomes a foundational requirement.”
Importantly, this publication also marks a milestone: it is the first scientific paper based on infrared spectra of blood plasma samples collected as part of the Health for Hungary (H4H) study — an ambitious initiative focused on advancing data-driven healthcare.
Picture: Thorsten Naeser
Original publication:
Flora B. Nemeth, Niklas Leopold-Kerschbaumer, Diana Debreceni, Frank Fleischmann, Krisztian Borbely, David Mazurencu-Marinescu-Pele, Thomas, Bocklitz, Mihaela Žigman and Kosmas V. Kepesidis
Bridging Spectral Gaps: Cross-Device Model Generalization in Blood-Based Infrared Spectroscopy
Analytical Chemistry Article ASAP, May 7, 2025
