Catálogo de publicaciones - libros

Current predictive models in the intensive care rely on summaries of data collected at patient admission. It has been shown recently that temporal patterns of the daily Sequential Organ Failure Assessment (SOFA) scores can improve predictions. However, the derangement of the six individual organ systems underlying the calculation of a SOFA score were not taken into account, thus impeding the understanding of their prognostic merits. In this paper we propose a method for model induction that integrates in a novel way the individual organ failure scores with SOFA scores. The integration of these two correlated components is achieved by summarizing the historic SOFA information and at the same time by capturing the evolution of individual organ system failure status. The method also explicitly avoids the collinearity problem among organ failure episodes. We report on the application of our method to a large dataset and demonstrate its added value. The ubiquity of severity scores and sub-scores in medicine renders our approach relevant to a wide range of medical domains.

The task addressed and the method proposed in this paper aim at improved understanding of differences between similar diseases. In particular we address the problem of distinguishing between thrombolic brain stroke and embolic brain stroke as an application of our approach of contrast set mining through subgroup discovery. We describe methodological lessons learned in the analysis of brain ischaemia data and a practical implementation of the approach within an open source data mining toolbox.

Coronary artery disease is one of the world’s most important causes of early mortality, so any improvements of diagnostic procedures are highly appreciated. In the clinical setting, coronary artery disease diagnostics is typically performed in a sequential manner. The four diagnostic levels consist of evaluation of (1) signs and symptoms of the disease and ECG (electrocardiogram) at rest, (2) ECG testing during a controlled exercise, (3) myocardial perfusion scintigraphy, and (4) finally coronary angiography (which is considered as the “gold standard” reference method). In our study we focus on improving diagnostic performance of the third diagnostic level (myocardial perfusion scintigraphy). This diagnostic level consists of series of medical images that are easily obtained and the imaging procedure represents only a minor threat to patients’ health. In clinical practice, these images are manually described (parameterized) and subsequently evaluated by expert physicians. In our paper we present an innovative alternative to manual image evaluation – an automatic image parametrization on multiple resolutions, based on texture description with specialized association rules, and image evaluation with machine learning methods. Our results show that multi-resolution image parameterizations equals the physicians in terms of quality of image parameters. However, by using both manual and automatic image description parameters at the same time, diagnostic performance can be significantly improved with respect to the results of clinical practice.

Monitoring devices in intensive care units observe a patient’s health status and trigger an alarm in critical situations. The alarm rules in commercially available monitoring systems are usually based on simple thresholds set by the clinical staff. Though there are some more advanced alarm rules integrated in modern monitoring devices, even for those, the false alarm rate is very high. Decision trees have proven suitable for alarm classification and false alarm reduction. Random forests which are ensembles of trees can improve the accuracy compared to single trees in many situations. In intensive care, the probability of misclassifying a situation in which an alarm is needed has to be controlled. Subject to this constraint the probability of misclassifying a situation in which no alarm should be given has to be minimized - an analogy to a hypothesis test for testing “situation is alarm relevant” vs. “situation is non alarm relevant” based on an ensemble of trees. This yields a classification rule for any given significance level, which is the probability of misclassifying alarm relevant situations. We apply this procedure to annotated physiological data recorded at an intensive care unit and generate rules for false alarm reduction.

The theories of traditional Chinese medicine (TCM) originated from experiences doctors had with patients in ancient times. We ask the question whether aspects of TCM theories can be reconstructed through modern day data analysis. We have recently analyzed a TCM data set using a machine learning method and found that the resulting statistical model matches the relevant TCM theory well. This is an exciting discovery because it shows that, contrary to common perception, there are scientific truths in TCM theories. It also suggests the possibility of laying a statistical foundation for TCM through data analysis and thereby turning it into a modern science.

This paper presents a novel method integrating gene-gene interaction information and Gene Ontology (GO) for the construction of new gene sets that are potentially enriched. Enrichment of a gene set is determined by Gene Set Enrichment Analysis. The experimental results show that the introduced method improves over existing methods, i.e. that it is capable to find new descriptions of the biology governing the experiments, not detectable by the traditional methods of evaluating the enrichment of predefined gene sets, defined by a single GO term.

This paper discusses variable selection for medical decision making; in particular decisions regarding which treatment to provide a patient. Current variable selection methods were designed for use in prediction applications. These techniques often leave behind small but important interaction variables that are critical when the goal is decision making rather than prediction. This paper presents a new method designed to find variables that aid in decision making and demonstrates the method on data from a clinical trial for treatment of depression.

This paper analyzes two different approaches to the detection of supporting factors used in descriptive induction. The first is based on the statistical comparison of the pattern properties relative to the properties of the entire negative and the entire positive example sets. The other approach uses artificially generated random examples that are added into the original training set. The methodology is illustrated in the analysis of patients suffering from brain ischaemia.

The present work aims at refining and expanding user’s queries thanks to association rules. We adapted the A-Close algorithm to a medical corpus indexed by MeSH descriptors. The originality of our approach lies in the use of the association rules in the information retrieval process and the exploitation of the structure of the domain knowledge to evaluate the association rules. The results show the usefulness of this query expansion approach. Based on observations, new knowledge is modelled as expert rules.

The growing availability of measurement devices in the operating room enables the collection of a huge amount of data about the state of the patient and the doctors’ practice during a surgical operation. This paper explores the possibilities of generating, from these data, decision support rules in order to support the daily anesthesia procedures. In particular, we focus on machine learning techniques to design a decision support tool. The preliminary tests in a simulation setting are promising and show the role of computational intelligence techniques in extracting useful information for anesthesiologists.