Lothar Gierl

Case-based reasoning for antibiotics therapy advice: an investigation of retrieval algorithms and prototypes

We have developed an antibiotics therapy advice system called ICONS for patients in an intensive care unit (ICU) who have caught an infection as additional complication. Since advice for such critically ill patients is needed very quickly and as the actual pathogen still has to be identified by the laboratory, we use an expected pathogen spectrum based on medical background knowledge and known resistances. The expected pathogen spectra and the resistance information are periodically updated from laboratory results. To speed up the process of finding suitable therapy recommendations, we have applied case-based reasoning (CBR) techniques. As all required information should always be up to date in medical expert systems, new cases should be incrementally incorporated into the case base and outdated ones should be updated or erased. For reasons of space limitations and of retrieval time an indefinite growth of the case base should be avoided. To fulfill these requirements we propose that specific single cases should be generalised to more general prototypical ones and that subsequent redundant cases should be erased. In this paper, we present evaluation results of different generation strategies for generalised cases (prototypes). Additionally, we compare measured retrieval times for two indexing retrieval algorithms: simple indexing, which is appropriate for small and medium case bases, and tree-hash retrieval, which is advantageous for large case bases.

CBR in Medicine

Medicine differs from other knowledge domains by the interaction of research and practice. The objects are the patients - very complex organisms with high biological variance and a lot of interactive vital processes. The knowledge of these processes and their interactions is often weak. It mostly depends on a high number of sometimes even contradicting signs and symptoms. Furthermore, the individual vital processes are affected by changes of the environmental situations (e.g., new resistances, diseases or pathogens). To discover new medical knowledge, traditional research is based on disease case descriptions, case collections, and biostatistical case studies.

Paper: Integrating consultation and semi-automatic knowledge acquisition in a prototype-based architecture: Experiences with dysmorphic syndromes

The paper describes an application of cognitive theories of Tversky and Rosch to prototype similarity of dysmorphic syndromes cases. The knowledge-based system supports diagnostic consultation and research in dysmorphic syndromes. It has been used routinely for many years. The knowledge base is semi-automatically generated from known cases of an outpatient clinic. Some results of the evaluation process of the systems achievements are shown. General conclusions based on the experience with this successful system are discussed.

Peroxisome proliferator - activated receptor γ overexpression inhibits pro-fibrogenic activities of immortalised rat pancreatic stellate cells

Pancreatic stellate cells (PSCs) play a key role in the development of pancreatic fibrosis, a constant feature of chronic pancreatitis and pancreatic cancer. In response to pro‐fibrogenic mediators, PSCs undergo an activation process that involves proliferation, enhanced production of extracellular matrix proteins and a phenotypic transition towards myofibroblasts. Ligands of the peroxisome proliferator‐activated receptor gamma (PPARγ), such as thiazolidinediones, are potent inhibitors of stellate cell activation and fibrogenesis in pancreas and liver. The effects of PPARγ ligands, however, however, are at least in part mediated through PPARγ‐independent pathways. Here, we have chosen a different approach to study regulatory functions of PPARγ in PSCs. Using immortalised rat PSCs, we have established a model of tetracycline (tet)‐regulated PPARγ over‐expression. Induction of PPARγ expression strongly inhibited proliferation and enhanced the rate of apoptotic cell death. Furthermore, PPARγ‐overexpressing cells synthesised less collagen than controls. To monitor effects of PPARγ on PSC gene expression, we employed Affymetrix microarray technology. Using stringent selection criteria, we identified 21 up‐ and 19 down‐regualated genes in PPARγ‐overexpressing cells. Most of the corresponding gene products are either involved in lipid metabolism, play a role in signal transduction, or are secreted molecules that regulate cell growth and differentition. In conclusion, our data suggest an active role of PPARγ in the induction of a quiescent PSC phenotype. PPARγ‐regulated genes in PSCs may serve as novel targets for the development of antifibrotic therapies.

Temporal Abstractions and Case-Based Reasoning for Medical Course Data: Two Prognostic Applications

We have developed a method for analysis and prognosis of multiparametric kidney function courses. The method combines two abstraction steps (state abstraction and temporal abstraction) with Case-based Reasoning. Recently we have started to apply the same method in the domain of Geomedicine, namely for the prognosis of the temporal spread of diseases, mainly of influenza, where just one of the two abstraction steps is necessary, that is the temporal one. In this paper, we present the application of our method in the kidney function domain, show how we are going to apply the same ideas for the prognosis of the spread of diseases, and summarise the main principles of the method.

Medical multiparametric time course prognoses applied to kidney function assessments

In this paper, we describe an approach to utilize case-based reasoning methods for trend prognoses for the monitoring of the kidney function in an Intensive Care Unit (ICU) setting. Since using conventional methods for reasoning over time does not fit for course predictions with poor medical knowledge of typical course patterns, we have developed abstraction methods suitable for integration into our case-based reasoning system ICONS. These methods combine medical experience with prognoses of multiparametric courses. On the ICU, the monitoring system NIMON provides a daily report based on current measured and calculated kidney function parameters. Subsequently, we generate course-characteristic trend descriptions of the renal function over the course of time. Using case-based reasoning retrieval methods, we search in the case base for courses similar to the current trend descriptions. Finally, we present the current course together with similar courses as comparisons and as probable prognoses to the user. We applied case-based reasoning methods in a domain which seemed reserved for statistical methods and conventional temporal reasoning.

Experiences with Case-Based Reasoning Methods and Prototypes for Medical Knowledge-Based Systems

In this paper we discuss the importance to create prototypes automatically within Case-Based Reasoning systems. We present some general ideas about prototypes deduced from analyses of our experiences with prototype designs in domain specific medical CBR systems. Four medical Case-Based Reasoning systems are described. As they use prototypes for different purposes, the gained improvement is different as well. Furthermore, we claim that the generation of prototypes is an adequate technique to learn the intrinsic case knowledge, especially if the domain theory is weak.

Abstractions of Data and Time for Multiparametric Time Course Prognoses

In this paper, we describe an approach to utilize Case-Based Reasoning methods for trend prognoses for medical problems. Since using conventional methods for reasoning over time does not fit for course predictions without medical knowledge of typical course pattern, we have developed abstraction methods suitable for integration into our Case-Based Reasoning system ICONS. These methods combine medical experience with prognoses of multiparametric courses. We have chosen the monitoring of the kidney function in an Intensive Care Unit (ICU) setting as an example for diagnostic problems. On the ICU, the monitoring system NIMON provides a daily report based on current measured and calculated kidney function parameters. We subsequently generate course-characteristic trend descriptions of the renal function over the course of time. Using Case-Based Reasoning retrieval methods, we search in the case base for courses similar to the current trend descriptions. Finally, we present the current course together with similar courses as comparisons and as possible prognoses to the user. We applied Case-Based Reasoning methods in a domain which seemed reserved for statistical methods and conventional temporal reasoning.

Representation of individual gene expression in completely pooled mRNA samples

Designing microarray experiments, scientists are often confronted with the question of pooling due to financial constraints, but discussion of the validity of pooling tends toward a sub-pooling recommendation. Since complete pooling protocols can be considered part of sub-pooling designs, gene expression data from three complete pooling experiments were analyzed. Data from complete pooled versus individual mRNA samples of rat brain tissue were compared to answer the question whether the pooled sample represents individual samples in small-sized experiments. Our analytic approach provided clear results concerning the Affymetrix® MAS 5.0 signal and detection call parameters. Despite a strong similarity of arrays within experimental groups, the individual signals were evidently not appropriately represented in the pooled sample, with slightly more than half of all the genes considered. Our analysis reveals problems in cases of small complete pooling designs with less than six subjects pooled.

Experiences with Prototype Designs and Retrieval Methods in Medical Case-Based Reasoning Systems

In the recent years we have developed three medical Case-Based Reasoning systems for diagnosis of dysmorphic syndromes, for antibiotics therapy advice and for course analysis and prognosis of the kidney function. In this paper we summarise our experience with these medical CBR applications. Special focuses are on the retrieval methods and on the importance and the design of typical cases (prototypes) which we use as a knowledge generalisation step to fill the gap between single cases and general knowledge.