Overview
The present pCO2 dataset contains data from a study conducted in April 2019 to evaluate a new type of sensor for the Roche cobas b123 BGE. It includes a total of 1196 measurements of blood samples and quality control solutions using 10 different sensors.
In this VM1 project, we will investigate the use of machine learning to construct correction terms for physical formulas to make the measurement more controlled and explainable. We aim to achieve this goal through the use of neural networks that integrate the Nernst equation.
This approach should lead to faster and significantly more accurate pCO2 measurements compared to (standard) pCO2 measurements that rely solely on the Nernst equation, including the post-calibration phase. Furthermore, faulty measurements (anomalies) should be detected using supervised and semi-supervised physics-informed models that have a higher F1 score than models that ignore physical laws.
