Classification of plasma cortisol patterns in normal subjects and in Cushing's syndrome.

Abstract

The 24-h pattern of half-hourly sampled plasma cortisol in normal human subjects shows a 24-h (circadian) period, which may be variably distorted in patients who suffer from autonomous hypercortisolism (Cushing's syndrome). We have developed a pattern recognition system for computer classification of cortisol time series into the normal class and subclasses of Cushing's syndrome with different etiology ("pituitary" designating pituitary tumor, "adrenal" designating adrenal tumor, and "ectopic" designating tumor elsewhere). Discriminatory features were extracted from Fourier analysis and Karhunen-Loeve expansion coefficients of cortisol time series. Decision functions were trained by the LMSE algorithm and tested by the jack-knife test procedure on a data-base of 90 normal and patient patterns. The classification accuracy for normal, "pituitary," "adrenal," and "ectopic" classes was 100, 98.1, 98.3, and 100%, respectively. Hence this pattern recognition system may be useful as an aid in the differential diagnosis of Cushing's syndrome.