Resumen de Machine Learning and Knowledge Management for Decision Support. Applications in Promotional Efficiency and Healthcare
The development achieved in Information and Communications Technologies in recent decades has brought an enormous growth in the collection and storage of data in such diverse fields as marketing, health, or safety. The availability of large amounts of data makes necessary the search for new machine learning paradigms, capable of addressing their automated analysis and the subsequent information extraction. Specifically, given a number of training examples (also called samples or observations) associated with desired outcomes, the machine learning techniques learn the relationship between them. In recent years, these techniques have experienced spectacular advances in both theoretical foundations and their application to a wide range of different knowledge domains. The general objective of this Thesis consists on the theoretical development and implementation of machine learning methods, with emphasis on the feature selection and predictive model design stages, allowing to tackle with the analysis of data of diverse nature, and creating specific procedures for each stage but at the same time applicable in various fields. This Thesis has addressed three specific areas of increasing economic and social interest: (a) interaction modeling between everyday products and promotional efficiency; (b) clinical decision support for early detection of complications after colorectal cancer surgery; (c) risk stratification of sudden cardiac death from predictive indices obtained from the electrical signals of the heart, using a clinical knowledge model and a standardized terminology. The data analysis in these applications shares the use of machine learning techniques according to the general goal. However, the diverse nature of these applications represents by itself a specific goal of this Dissertation. The first specific objective consists on further evaluation and analysis of promotional sales, traditionally based on classical statistical techniques. A substantial support decision making must necessarily come from the systematic analysis of massive data on the control and monitoring of promotions and their complex interactions. Therefore, a statistical analysis and comparison of various machine learning techniques is proposed. Another area of very different nature respect to the previous one, but with strong social interest, is healthcare. The analysis of clinical data, both structured (vital signs or blood tests) and unstructured (text-based documents), systematically and longitudinally collected from the electronic health record (EHR) of a large group of patients, can substantially increase the clinical knowledge and support decision-making. However, machine learning techniques and massive data analysis have provided, nowadays, a limited impact in the healthcare area. This situation is mainly due to the difficulty of extracting useful information from clinical data recorded in heterogeneous sources. In addition, there are few precedents of systems enabling the automatic analysis of information at the aggregated level among different hospital entities. There is a great need for suitable and relevant data as a basis for scientific advance, with greater impact on the clinical practice. In this Thesis, two healthcare domains highly relevant in most developing countries are analyzed, namely, colorectal cancer and cardiovascular diseases. The second specific objective is the adaptation and application of machine learning methods for early detection of complications after colorectal cancer surgery, analyzing both individually and jointly data from heterogeneous sources recorded in the EHR. The third specific objective is to build clinical knowledge models to enable data exchange and semantical understanding of clinical information from different EHR. In recent years, numerous predictors of cardiac risk indices have been proposed. Specifically, in this Thesis, the heart rate turbulence is analyzed to be a predictor of sudden cardiac death with clear and concise guidelines. Nowadays, the analysis of large amounts of data as well as the theoretical development of new machine learning algorithms undoubtedly represent a very active area of research in different domains. This Thesis contributes to improve knowledge and decision making in real-world applications of diverse nature which still share remarkable common denominators.
