Methods Inf Med

Randomized observational studies on the economics of therapies--biometrical experience of two trials.

Year 1998
Hogel J. Rodloff AC. Buchele G. Gaus W.
Department of Biometry and Medical Documentation, University of Ulm.
Economic studies in medicine are intended to investigate costs, associated with a particular problem dealing with the indication, diagnosis or therapy, for instance, whether the high costs involved in a highly intensive or innovative therapy could be balanced by the eventual savings made, due to the shorter periods of treatment. In such situations a randomized controlled trial is necessary to find out which therapy or which therapeutical strategy is least expensive in the long run. Economic studies do, however, present some specific problems. Making a list of all the cost-relevant treatment items can be very laborious, but the use of flat rates and lump sums alone cannot lead to a complete cost analysis. Often, costs between hospitals vary more than between treatment regimens. Early and sudden deaths incur low costs and may bias the results. Furthermore, costs are distributed with a long and heavy upper tail including extreme outliers. This does, in fact, complicate the estimation of the sample size. In this article, these problems are outlined and, with the help of the data obtained from two randomized economic trials in health care, solutions are proposed and discussed.

Diagnosis of acute appendicitis in two databases. Evaluation of different neighborhoods with an LVQ neural network.

Year 1998
Pesonen E. Ohmann C. Eskelinen M. Juhola M.
Department of Computer Science and Applied Mathematics, University of Kuopio, Finland. erkki.pesonen@uku.fi
The use of an artificial neural network system was studied in the diagnosis of acute abdominal pain, especially acute appendicitis, with patients from Finland and Germany. Separate Learning Vector Quantization (LVQ) neural networks were trained with a training set from each database and also with a combined database. Each neural network was evaluated separately with a test set of cases from each database. With the combined database different neighborhood methods were compared to find the optimal choice for this decision-making problem. The acute appendicitis cases of the Finnish test data set were classified well with all the networks, but the cases of the German test set were difficult to classify for the Finnish network. The use of larger neighborhoods increased the sensitivity of the classification by nearly 10%. The differences in the results of the Finnish and German databases suggest that there are differences in the data collection or patient populations between centers. Therefore, care must be taken when using decision-support systems which have been developed in other centers. Neural networks offer a method to evaluate differences between databases. With the use of larger neighborhoods, the effects of the differences on the accuracy of the classification can be partly diminished.