Ebook: Artificial Intelligence in Medicine

This volume contains the proceedings of the 4th International Conference on Artificial Intelligence in Medicine, which will be held in Munich, Germany, 3 - 6 October 1993. The conference is organized by AlME, the European Society for Artificial Intelligence in Medicine Europe. Previous international conferences were held in Marseille (1987), London (1989), and Maastricht (1991).

AIME was established in 1986 to foster fundamental and applied research in artificial intelligence and in symbolic information processing techniques for medical care and medical research Artificial Intelligence in medicine is a special field of Medical Informatics and deals with the enhancement of medical information systems by knowledge based components and new paradigms.

AIME 93 is hosted by the German National Centre for Environmental and Health Research (GSF) and supported by both the working group on “Expert Systems” of the German Society for Medical Informatics, Biometry, and Epidemiology (GMDS) and the national research project on Medical Knowledge Bases MEDWIS. The GSF-MEDIS Institute and the GSF conference service have taken most of the organizational burden on their shoulders.

AIME 93 will show recent scientific results presented in oral contributions, posters and demonstrations. The informatical topics range from “A priori probabilities” to “Validation” and the clinical topics from “Abdomen” to “Ventilation”. The tutorials, which are both methodological and user-oriented, will constitute and important element of the conference.

The present proceedings contain 60 contributions to AlME 93 and cover all presentations which are part of the scientific programme, among the 2 keynotes, 26 oral presentations with an average of 12 pages, 38 poster presentations and demonstrations with an average of 4 pages. The papers were each selected and reviewed by at least two members of the international programme committee. We are very thankful for the trouble they have taken. Their recommendations were sent to the authors and incorporated in the final texts. They also helped us shape the structure of the conference.

The papers are mainly grouped around the two subject areas of “Knowledge acquisition and representation” and “Strategies for medical reasoning” which form the backbone of most of the theoretical work in the field. It was encouraging to see that many papers also describe applications of AI systems in medicine and that many of these have reached a remarkable degree of maturity. AI technology is passing the threshold between research-oriented feasibility studies and recognized clinical applications. AIME 93 will show both the research of understanding the principles of medical knowledge and medical reasoning and the development of methodologies and tools for professional design, realization, and implementation of knowledge-based systems.

Many people - too many to name all of them here - have contributed to the solid preparation of this conference. We wish to thank them all as they have done this job in addition to their usual work, especially Volker Böhm, Lydia Janke, and Wolfgang Moser, whose considerable efforts have paved the way for AIME 93.

Steen Andreassen, Rolf Engelbrecht and Jeremy Wyatt (Editors)

October 1993

An analysis is made of the reasons why clinical support systems in health care (e.g., imaging systems, laboratory systems) are in routine use whereas decision support systems have not been widely accepted. The conclusions of this review are illustrated by the results of a large evaluation study of systems for ECG interpretation.

The explanation power of a medical knowledge based system (MKBS) can be highly enhanced by the availability of deep pathophysiological knowledge. This kind of knowledge is well represented through mathematical models whose formulation, either quantitative or qualitative, requires both domain specific and domain independent knowledge. This paper describes QCMF, a tool which makes easier the acquisition, through an interactive graphical interface, of the domain knowledge and embeds the needed methodological knowledge to formulate a qualitative pathophysiological model. Such a model is directly coded in QSIM and is simulated in order to predict the system behavior under the effects of a variety of pathogenetic mechanisms and therapeutical treatments. Moreover, different models of a given pathophysiological system are stored in a library so that they can be efficiently retrieved and exploited within a MKBS.

Magnetic Resonance (MR) imaging is a medical technique which permits the visualization of a variety of tumors, lesions, and abnormalities present within the soft biological tissues of the body. Segmentation of medical image data is the process of assigning anatomically-meaningful labels to each component of the image. This paper presents a method of knowledge-based interpretation of MR images of the head. In particular, the method is designed for the segmentation, or detection, of multiple sclerosis (MS) lesions of the brain. Knowledge of brain anatomy is represented in the form of a tissue probability model which provides a priori probabilities of brain tissue distribution per unit voxel in a standardized 3D ‘brain space’. As 90-95% of 118 lesions occur in white matter tissue [1], the model was used to confine the search for MS lesions to plausible white matter locations. Use of the model to segment multiple sclerosis lesions reduced the number of false positive lesions by 50-80% in comparison to non-model-based (i.e. purely data-driven) segmentation.

The performance of the statistical minimum distance and Bayesian classifiers (which both represent learned tissue classification rules as mathematical formulae) are compared to that of pruned and unpruned decision tree classifiers (which represent classification rules symbolically, as flowchart-like tree structures). Although each classifier performed at about the same level of accuracy, the concise and symbolic representation of the rules generated by the pruned decision tree classifier may be preferred for its greater human interpretability.

This paper describes ongoing basic research in the context of reasoning about naive physiology. A qualitative representation of commonsense physiology at various levels of abstraction and detail and high levels of interrelationships is described. This representation is then used to derive the behaviour of a physiological process from the behaviour of its subprocesses and the relationships between them. The model is then perturbed (possibly representing pathophysiology) and a method to derive the effect of such a perturbation by propagating it throughout the model is presented.

An extension of a causal probabilistic network, modelling the humane glucose metabolism, making it possible to estimate patient specific parameters from data from multiple days, is presented. The importance of preserving information on both the patient specific means and standard deviations of the parameters is described. The approach is illustrated by examples of the role of the patient specific insulin sensitivity.

In medical problem solving two main subproblems can be identified, namely the selection of possible hypotheses from some initial data, and the evaluation of these hypothesis, possibly taking into account some additional evidence. In a number of model-based systems, more superficial knowledge is used for generating diagnostic hypotheses, whereas deeper knowledge is used to assess them, but these systems are usually domain specific. In this paper we present a generic methodology for representing causal knowledge, which enables diagnostic reasoning to be made at the appropriate knowledge depth. In particular, to obtain candidate diagnoses from easily observable findings (superficial knowledge) and to assess them subsequently via less easily observable findings (e.g. signs and results from complementary tests, usually associated with deeper domain entities).

Some metabolic processes are described by compartmental models. They are characterized by a cyclic graph and a system. of differential equations. We introduce a general method to transform such a system into a sequence of acyclic causal probabilistic networks (CPN). This method is based on the solution of the system of differential equations and a stochastification procedure. The result of this transformation is a realistic model of the metabolic process which may be simulated and adapted to new evidence very easily using the shell HUGIN. In contrary to that, the adaption of the deterministic process of the compartmental model to new evidence in general is not possible. Furthermore, such a stochastified description of a metabolic process easily may be modified by adding additional nodes for additional effects to the CPN. We apply this method of transforming a compartmental model into a sequence of CPNs to a simple glucose - insulin model.

We are developing a decision support system for the diagnosis of polyneuropathies. The system is based on a causal probabilistic network (CPN) which models the polyneuropathies, pathophysiological changes and symptoms. An evaluation of the system shows that the system is able to correctly diagnose a polyneuropathy and find its cause.

Real time expert systems can be helpful to the anesthetist, being in charge of monitoring signals coming from the patient during an operation. The large stream of data can be organized, validated and interpreted by such an intelligent system. A physical model of the patient which is able to simulate the signals that are measured at an actual patient, combined with actual anesthesia equipment, can provide the signals, originating from known faults, from which the knowledge which forms the core of the expert system can be derived. Instead of physical patient models and actual equipment, mathematical models can be used for this task. Mathematical models are more flexible and cheaper than physical models. Our approach, in which we combine knowledge based methods with mathematical simulation models is described in this paper. The medical problem domain will be artificial ventilation of a patient during anesthesia.

This paper reports on the EPISTOL project, set up to to provide a perspective on how are knowledge based systems systems going to be used in the health sector 5 to 10 years from now, and how should this expected use influence the planning of research and development work up to that period. The results of the project are aimed to aid the planning of future programmes concerned with research and development in health telematics, namely the fourth Framework Programme of the Comission of the European Communities.

Ad hoc decision support is required for rare diseases. The construction of full-scale expert systems would be effective, but, too expensive. Co-operative problem-solving of a physician and a knowledge-based advisor combines their strengths efficiently. The architecture of a knowledge-based advisor is presented. The medical application domain is the acute radiation syndrome.

Macro-operations, as conveyors of strategy-directed diagnostic reasoning, are considered: they, as identifiable task-specific prescriptions, are to extend the expressiveness of knowledge representation and surpass the limitations of universal inferential operations.

The promise of computers in medicine, including their role in decision support, can only be fulfilled if the overall effectiveness of the system is not constrained by a lack of coordination between the members of a group all sharing the same goal, namely patient care. Cooperative medical decision making involves a continuous process, assessing the validity of data, information and knowledge acquired and inferred by the colleagues, that is, a shared knowledge space must be transparent. The ACCORD methodology provides an interpretation framework for the mapping of domain facts - constituting the world model of the expert - onto conceptual models, which can be expressed in formal representations. The Maccord framework allows a stepwise and inarbitrary reconstruction of the problem solving competence of medical experts as a prerequisite for an appropriate architecture of both medical knowledge bases and the “reasoning device”.

The GAMES method for medical knowledge based decision support system construction is based on a select-and-test model of a medical decision making task. This paper reports the approach to clinical validation of the model which has been undertaken by applying it to a range of diagnostic and treatment planning tasks in asthma, cancer and intensive care. The potential applicability of the select-and-test model to patient monitoring is also discussed.

A new class of computer intellectual systems for assisting high-skilled specialists is offered. Its destructive feature is an increased attention to stimulation methods of a specialist's intellectual activity. Characteristics of the class are demonstrated on the example of DINAR-2.2 system supporting decision making by the physician- reanimatologist on duty of the Regional Children Reanimation Consultative Centre (RCC).

In this paper a technical realization of an attentional mechanism extracting salient regions in face images is presented. Using the knowledge of the eye movement strategies of the human visual system the attentional mechanism localizes prominent regions in natural images. Elementary visual features are derived from high-resolution digitized images and represented in a so called topographical attention map. A top-down algorithm analyzes this map implemented as a multi-scale representation by searching the most salient features in each represented scale. The scale information is propagated from coarse scales to the next higher resolutions while the extent of the fixated region is adapted. After the extraction the selected and now locally limited regions are analyzed and interpreted. The proposed attentional mechanism is part of an image processing system identifying faces with dysmorphic signs. To improve the classification of a local analysis module, it is intended to examine these characteristic areas with specialized algorithms. This knowledge based approach shows that an artificial attention control system is capable of localizing the prominent facial areas (eyes, mouth, nose, etc.) in face images only based on elementary image features. The proposed algorithm can also be used to detect other attentive regions in any images based on appropriate features.

Drug dosage must be adjusted to kidney function to avoid toxic overdosage in patients with impaired renal function. These clinical demands must be attained within a hospital information system. The PHARMNEPH system is part of such a project. It is based on published pharmacokinetic knowledge standardized by nonparametric meta-analysis. A bayesian parameter estimate from population derived kinetics and sparce individual data will be implemented. Algorithms for a dosage calculation will be provided combining individual bayesian parameters with pharmacodynamic characteristics of each drug.

Detailed representation of intermediate pathophysiological concepts allows step-by-step diagnostic thinking which, starting from broad pathophysiologic terms, incrementally sharpens the focus of the working hypothesis and finally arrives at a specific diagosis. Rather than driving the diagnostic work-up with a single reasoning strategy, we let several - mainly forward-chained flow charts and hypothesize-and-test - compete in determining its course, i.e. the sequence of diagnostic tests.