Adaptive signal analysis of immunological data

This paper aims to investigate whether both supenised and unsupervised signal analysis can contribute to the interpretation of immunological data. For this purpose a data base was set up containing cellular data from bronchoalveolar lavage fluid which was obtained from 37 children with pulmonary diseases. The children were di-chotomized into two groups: 20 children suffered from chronic bronchitis whereas 17 children had an interstitial lung disease. A self-organizing map (SOM) and linear in-dependent component analysis were utilized to test higher- order correlations between cellular subsets and the patient groups. Furthermore, a supervised approach with a perceptron trained to the patients' diagnosis was applied. The SOM confirmed the results that were expected from previous statistical analyses. The results of the ICA were rather weak, which lies presumably in the fact that a linear mixing model of independent sources does not hold; nevertheless, we could find parameters of high diagnosis influence that were confirmed by the perceptron. The supervised perceptron learning after principal component analysis for dimension reduction turned out to be highly successful by linearly separating the patients into two groups with different diagnoses. The simplicity of the perceptron made it easy to extract diagnosis rules, which partly were blown already and can now readily be tested on larger data sets. In conclusion, neural network signal analysis provides promising tools for the analysis of highly complex immunological data.