Machine Learning Techniques for Analysis of Biochemical Data

This thesis explores the quantification of blood lactate using non-invasive sensors within diagnostic machinery for extracorporeal flows, considering technologies like optical spectral analysis, Near-Infrared, and Mid-Infrared as promising. The primary objective is the long-term monitoring of lactate concentration. In the initial tests, predictive multivariate analysis without neural networks is employed, emphasizing the accurate measurement of lactate amid disturbances. The study delves into the behavior of blood components, dedicating a chapter to understanding CO2, hemoglobin, and oxygen in the areas of interest. Machine Learning techniques, such as Multiple Linear Regression, Principal Component Analysis, Principal Component Regression, and especially Partial Least Squares, are thoroughly explored. The experimental phase is bifurcated, comprising quantitative dye analysis and exploring lactate concentrations in cell cultures. In the first part, the model effectively handles Red Alizarin and PrestoBlue concentrations, demonstrating robustness in complex scenarios. The second part extends the Machine Learning technique to biological samples. Initial tests and a larger experiment with 54 samples highlight adaptability to real-world complexities. Both parts underscore the positive versatility of the model. The comprehensive set forms a robust foundation for future development. This thesis marks the beginning of a broader journey, aiming to develop an advanced application for directly measuring lactate concentration in blood spectra, contributing significantly to innovative approaches in biochemical analysis based on machine learning techniques and spectroscopy.