Jürgen Paetz

Knowledge-based approach to septic shock patient data using a neural network with trapezoidal activation functions

In this contribution we present an application of a knowledge-based neural network technique in the domain of medical research. We consider the crucial problem of intensive care patients developing a septic shock during their stay at the intensive care unit. Septic shock is of prime importance in intensive care medicine due to its high mortality rate. Our analysis of the patient data is embedded in a medical data analysis cycle, including preprocessing, classification, rule generation and interpretation. For classification and rule generation we chose an improved architecture based on a growing trapezoidal basis function network for our metric variables. Our results extend those of a black box classification and give a deeper insight in our patient data. We evaluate our results with classification and rule performance measures. For feature selection we introduce a new importance measure.

Reducing the number of neurons in radial basis function networks with dynamic decay adjustment

Classification is a common task for supervised neural networks. A specific radial basis function network for classification is the so-called RBF network with dynamic decay adjustment (RBFN-DDA). Fast training and good classification performance are properties of this network. RBFN-DDA is a dynamically growing network, i.e. neurons are inserted during training. A drawback of RBFN-DDA is its greedy insertion behavior. Too many superfluous neurons are inserted for noisy data, overlapping data or for outliers. We propose an online technique to reduce the number of neurons during training. We achieve our goal by deleting neurons after each training of one epoch. By using the improved algorithm on benchmark data and current medical data, the number of neurons is reduced clearly (up to 93.9% less neurons). Thus, we achieve a network with less complexity compared to the original RBFN-DDA.

Septic Shock Diagnosis by Neural Networks and Rule Based Systems

In intensive care units physicians are aware of a high lethality rate of septic shock patients. In this contribution we present typical problems and results of a retrospective, data driven analysis based on two neural network methods applied on the data of two clinical studies.

Data quality aspects of a database for abdominal septic shock patients

Since many years, medical researchers have investigated the mechanisms that may cause a septic shock. Despite many approaches that analyzed smaller parts of the relevant data or single variables, respectively, no larger database with all the possible relevant data existed. Our work was to bridge this gap. We built a large database for abdominal septic shock patients. While building it, we were confronted with many problems concerning the database realization and the data quality. Thus, we will demonstrate how we built our database and how we assured data quality. This is of interest for all medical or computer scientists who are concerned with building medical databases with retrospective data, e.g. for data mining purposes.

Review of A Large Clinical Series: Predicting Death for Patients With Abdominal Septic Shock

This paper reports the result of the MEDAN project that analyzes a multicenter septic shock patient data collection. The mortality prognosis based on 4 scores that are often used is compared with the prognosis of a trained neural network. We built an alarm system using the network classification results. Method. We analyzed the data of 382 patients with abdominal septic shock who were admitted to the intensive care unit (ICU) from 1998 to 2002. The analysis includes the calculation of daily sepsis-related organ failure assessment (SOFA), Acute Physiological and Chronic Health Evaluation (APACHE) II, simplified acute physiology score (SAPS) II, multiple-organ dysfunction score (MODS) scores for each patient and the training and testing of an appropriate neural network. Results. For our patients with abdominal septic shock, the analysis shows that it is not possible to predict their individual fate correctly on the day of admission to the ICU on the basis of any current score. However, when the trained network computes a score value below the threshold during the ICU stay, there is a high probability that the patient will die within 3 days. The trained neural network obtains the same outcome prediction performance as the best score, the SOFA score, using narrower confidence intervals and considering three variables only: systolic blood pressure, diastolic blood pressure and the number of thrombocytes. We conclude that the currently best available score for abdominal septic shock may be replaced by the output of a trained neural network with only 3 input variables.

About the Analysis of Septic Shock Patient Data

In intensive care medicine doctors are aware of a high mortality rate of septic shock patients. In this contribution we present the problems and the results of a retrospective, data driven analysis of two studies made in Frankfurt am Main between 1993 and 1997. Our approach includes the necessary steps of data mining, i.e. building up a data base, cleaning and preprocessing the data and finally choosing an adequate analysis for the medical patient data. We chose an architecture mainly based on a supervised neural network. The patient data is classified into two classes (survived and deceased). The importance of this classification for an early warning system is discussed.

A Frequent Patterns Tree Approach for Rule Generation with Categorical Septic Shock Patient Data

In abdominal intensive care medicine letality of septic shock patients is very high. In this contribution we present results of a data driven rule generation with categorical septic shock patient data, collected from 1996 to 1999. Our descriptive approach includes preprocessing of data for rule generation and application of an efficient algorithm for frequent patterns generation. Performance of generated rules is rated by frequency and confidence measures. The best rules are presented. They provide new quantitative insight for physicians with regard to septic shock patient outcome.

A neuro-fuzzy approach to virtual screening in molecular bioinformatics

Molecular bioinformatics is a transdisciplinary working area. One hot topic is the design of drugs using computers and intelligent algorithms. This is known as in silico approach. We use a new in silico approach for separating active ligand molecules from inactive ones for different drug targets. This kind of retrospective virtual screening is performed by using encoded molecule data and a neuro-fuzzy methodology for classification, feature selection, and rule generation. We generate rules in a retrospective screening process that identify regions, where clearly more active compounds can be found compared to their a priori probability. We show that our approach is superior to a common descriptor-based standard technique.

Supervised neuro-fuzzy clustering for life science applications

Classification, clustering and rule generation are important tasks in multidimensional data analysis. The combination of clustering or classification with rule generation gives an explanation for the achieved results. Especially in life science applications experts are interested in explanations to understand the underlying data. The usage of supervised neuro-fuzzy systems is a suitable approach for this combined task. Not always classification labels are available for the data when considering new problem areas in life science. Since we had already used a supervised neuro-fuzzy system for some applications, our aim in the case studies was to use the same neuro-fuzzy classifier for clustering, generating understandable rules also for clusters. To do so, we added Monte-Carlo random data to the original data and performed the clustering task with the present classifier in the medical, chemical, and biological domain.

Evolution of Real Valued Weights for RBF-DDA Networks

The radial basis function network with dynamic decay adjustment is a fast adaptive classifier. Its specific property is the utilization of integers as weights, counting those training samples that are elements of a certain radial region in the data space. In our experimentation we allow real valued weights instead of integers during their evolution. With real values the application of an evolutionary algorithm increases classification performance. An additional study shows the effects of pruning neurons with weights, that evolved to values lower than one. Overall, the model with real valued weights performs better or its network topology becomes less complex.