Uncertainty Evaluation in classification-based ML measurements: A first step towards a GUM-based methodology

The adoption of Machine Learning (ML) techniques is becoming increasingly relevant across several application domains, including Measurement Science, where data-driven approaches are progressively complementing traditional model-based ones. Recent studies have demonstrated that ML regression models can support the conduction of measurements that are otherwise difficult, costly, or impractical to perform using conventional techniques, enabling new possibilities for indirect measurements. Nevertheless, according to the International Vocabulary of Metrology (VIM), measurement processes are not limited to regression-based approaches, as classification tasks involving ordinal quantities are also formally included within the definition of measurement. Despite this conceptual inclusion, the application of ML classification models in measurement contexts remains insufficiently explored, particularly with respect to the rigorous evaluation and expression of measurement uncertainty. As a consequence, a significant methodological gap persists in ensuring traceability and comparability of classification-based measurement results. Based on these considerations, this work presents the initial steps toward the definition of a methodology compliant with the Guide to the Expression of Uncertainty in Measurement (GUM) for uncertainty evaluation in measurement processes employing ML classification models. Experimental results obtained from a case study focused on battery state-of-health assessment show that the proposed procedure enables a comprehensive uncertainty evaluation, accounting for all relevant sources of variability along the measurement chain.