The prediction of Creatinine and Bilirubin using Machine Learning Methods

Abstract

This thesis critically assesses the effectiveness of these models in handling the inherent imbalances and complexities within EHR data, particularly focusing on the predictive accuracy for Creatinine and Bilirubin levels. Oversampling techniques are meticulously applied to rectify class imbalances, enhancing the models' sensitivity towards less prevalent, yet clinically significant outcomes. The comparative analysis highlights the nuanced interplay between model choice, data preprocessing techniques, and the specific characteristics of the biomarkers in question, providing insightful implications for clinical applications. On evaluation with the 825 patients' data, the model achieved sensitivities of 95 (23/24) in the data labelled change of Creatinine, 79 (635/801) in not change of Creatinine,70 (87/124) in the data labelled change of Bilirubin, and 72 (509/701) in not change of Bilirubin.

The findings reveal a varied performance landscape across models and biomarkers, underscoring the importance of tailored approaches in predictive healthcare modeling. Supervised models demonstrated commendable accuracy in majority scenarios, while oversampling techniques offered nuanced benefits, particularly in bolstering the models' ability to detect significant changes in biomarker levels. The study further illuminates the challenges associated with EHR data, including variability, dimensionality, and quality issues, proposing avenues for future research focused on advanced preprocessing techniques, feature selection, and the exploration of deep learning models to surmount these obstacles.

In essence, this research contributes to the burgeoning field of medical informatics by showcasing the potential of ML models to advance predictive diagnostics and personalized medicine, ultimately aiming to enhance patient care through early detection and monitoring of health indicators.