Publication: Automatic analysis of high dimensional categorical variables in medical databases for the prediction of hospital bacteremia
Full text at PDC
This project aims to continue and consolidate the study for the bacteriemia detection process and its diagnosis carried out by some faculty companions last year. A first glance through the analysis of numerical variables allowed a deeper understanding and the trace of an approach for a quick detection model. Now, categorical variables take relevance too in order to successfully achieve higher results in the classifier models. The addition of categorical variables in classifier models has been around for at least five years due to the increase in computational capacity, and the benefits in the classifiers as direct consequence is clear. Yet, it is proven that, as complex and abstract as language is, classifiers do struggle when data with slang or abbreviations comes up for prediction, even if its linguistic register is heavily bounded, i.e. when strictly related to medical issues data is treated. Throughout the study we will apply text cleaning and text processing methods to prepare the variables for use, since their format is heterogeneous and unsuitable to be processed by Machine Learning tools. We will also apply the string similarity method to identify all those classes that can help in the algorithm classification process and we will assess the most suitable types of encoding for working with these variables. Finally, we will apply the Random Forest Machine Learning algorithm on the set with techniques that allow us to avoid data learning bias and we will assess the results in terms of the success rates and the relevance of the variables in the decision-making process of the algorithm.
Trabajo de Fin de Grado en Ingeniería Informática, Facultad de Informática UCM, Departamento de Arquitectura de Computadores y Automática, Curso 2020/2021.