Catálogo de publicaciones - libros

As demand for health care increases, a high efficiency on limited resources is necessary for affordable high patient service levels. Here, we present an adaptive approach to efficient resource usage by automatic optimization of resource calendars. We describe a precise model based on a case study at the radiology department of the Academic Medical Center Amsterdam (AMC). We model the properties of the different groups of patients, with additional differentiating urgency levels. Based on this model, we develop a detailed simulation that is able to replicate the known scheduling problems. In particular, the simulation shows that due to fluctuations in demand, the allocations in the resource calendar must be flexible in order to make efficient use of the resources. We develop adaptive algorithms to automate iterative adjustments to the resource calendar. To test the effectiveness of our approach, we evaluate the algorithms using the simulation. Our adaptive optimization approach is able to maintain overall target performance levels while the resource is used at high efficiency.

In diagnostic decision-support systems, a test-selection facility serves to select tests that are expected to yield the largest decrease in the uncertainty about a patient’s diagnosis. For capturing diagnostic uncertainty, often an information measure is used. In this paper, we study the Shannon entropy, the Gini index, and the misclassification error for this purpose. We argue that for a large range of values, the first derivative of the Gini index can be regarded as an approximation of the first derivative of the Shannon entropy. We also argue that the differences between the derivative functions outside this range can explain different test sequences in practice. We further argue that the misclassification error is less suited for test-selection purposes as it is likely to show a tendency to select tests arbitrarily. Experimental results from using the measures with a real-life probabilistic network in oncology support our observations.

In this paper, we discuss efforts to apply a novel Bayesian network (BN) structure learning algorithm to a real world epidemiological problem, namely the Nasopharyngeal Carcinoma (NPC). Our specific aims are : (1) to provide a statistical profile of the recruited population, (2) to help indentify the important environmental risk factors involved in NPC, and (3) to gain insight on the applicability and limitations of BN methods on small epidemiological data sets obtained from questionnaires. We discuss first the novel BN structure learning algorithm called Max-Min Parents and Children Skeleton (MMPC) developped by Tsamardinos et al. in 2005. MMPC was proved by extensive empirical simulations to be an excellent trade-off between time and quality of reconstruction compared to most constraint based algorithms, especially for the smaller sample sizes. Unfortunately, MMPC is unable to deal with datasets containing approximate functional dependencies between variables. In this work, we overcome this problem and apply the new version of MMPC on Nasopharyngeal Carcinoma data in order to shed some light into the statistical profile of the population under study.

Most test-selection algorithms currently in use with probabilistic networks select variables myopically, that is, test variables are selected sequentially, on a one-by-one basis, based upon expected information gain. While myopic test selection is not realistic for many medical applications, non-myopic test selection, in which information gain would be computed for all combinations of variables, would be too demanding. We present three new test-selection algorithms for probabilistic networks, which all employ knowledge-based clusterings of variables; these are a myopic algorithm, a non-myopic algorithm and a semi-myopic algorithm. In a preliminary evaluation study, the semi-myopic algorithm proved to generate a satisfactory test strategy, with little computational burden.

We present the ProCarSur system for prognostic reasoning in the domain of cardiac surgery. The system has a three-tiered architecture consisting of a Bayesian network, a task layer, and a graphical user interface. In contrast to traditional prognostic tools, that are usually based on logistic regression, ProCarSur implements a dynamic, process-oriented view on prognosis. The system distinguishes between the various phases of peri-surgical care, explicates the scenarios that lead to different clinical outcomes, and can be used to update predictions when new information becomes available. To support users in their interaction with the Bayesian network, a set of predefined prognostic reasoning tasks is implemented in the task layer. The user communicates with the system through an interface that hides the underlying Bayesian network and aggregates the results of probabilistic inferences.

As the number of health-related web sites in various languages increases, so does the need for control mechanisms that give the users adequate guarantee on whether the web resources they are visiting meet a minimum level of quality standards. Based upon state-of-the-art technology in the areas of semantic web, content analysis and quality labelling, the MedIEQ project, integrates existing technologies and tests them in a novel application: the automation of the labelling process in health-related web content. MedIEQ provides tools that crawl the web to locate unlabelled health web resources, to label them according to pre-defined labelling criteria, as well as to monitor them. This paper focuses on content collection and discusses our experiments in the English language.

The automated differentiation between benign and malignant abnormalities is a difficult problem in the breast cancer domain. While previous studies consider a single Bayesian network approach, in this paper we propose a novel perspective based on Bayesian network decomposition. We consider three methods that allow for different (levels of) network topological or structural decomposition. Through examples, we demonstrate some advantages of Bayesian network decomposition for the problem at hand: (i) natural and more intuitive representation of breast abnormalities and their features (ii) compact representation and efficient manipulation of large conditional probability tables, and (iii) a possible improvement in the knowledge acquisition and representation processes.

The influence diagram (ID) is a powerful tool for modelling medical decision-making processes, like the optimal application of diagnostic imaging. In this area, where safety and efficacy are determined by a number of aspects varying in nature and importance, it is difficult for humans to relate all available pieces of evidence and consequences of choices. IDs are well suited to provide evidence-based diagnostic pathways. However, medical specialists cannot be expected to be familiar with IDs and their output can be difficult to interpret. To overcome these shortcomings, a methodology is developed to automatically extract the optimal pathways from an ID and represent these in a tree shaped flow diagram. It imposes a few general rules on the structure of the model, which determine the relation between decisions to perform imaging and the availability of test results. Extracting the optimal pathways requires post-processing the results of an ID, leaving out the sub optimal choices and irrelevant scenarios. Predictive value of tests are vital information in medical protocols, they are at hand in the ID for each relevant scenario. The methodology is illustrated by the problem of diagnosing acute chest pain.

Drug-induced lung disease (DILD), often suspected in pneumology, is still a diagnostic challenge because of the ever increasing number of pneumotoxic drugs and the large diversity of observed clinical patterns. As a result, DILD can only be evoked as a plausible diagnosis after the exclusion of all other possible causes. PneumoDoc is a computer-based decision support that formalises the evaluation process of the drug-imputability of a lung disease. The knowledge base has been structured as a two-level decision tree. Patient-specific chronological and semiological criteria are first examined leading to the assessment of a qualitative intrinsic DILD plausibility score. Then literature-based data including the frequency of DILD with a given drug and the frequency of the observed clinical situation among the clinical patterns reported with the same drug are evaluated to compute a qualitative extrinsic DILD plausibility score. Based on a simple multimodal qualitative model, extrinsic and intrinsic scores are combined to yield an overall DILD plausibility score.

We describe a system that facilitates speech therapists to administer cognitive rehabilitation exercises and to evaluate treatment outcomes. We started by augmenting a commercial tool with a more user-friendly interface, meeting the needs of the healthcare professionals involved. Then we integrated, into the same tool, a new type of exercise, that is particularly patient-tailored, being based on the recognition of familiar images within a picture (such as a relative, a domestic animal, a home object, etc). Segmentation techniques are used to elaborate an input picture and individuate areas including objects, that will be semi-automatically linked to text and sound. The picture and associated information are then stored in the system database and may be subsequently used as objects for the new exercise. Any number of images may be elaborated, personalised and stored for each patient. The performance has been tested on voluntary subjects with good results.