Cluster analysis: a simple, versatile, and robust algorithm for endocrine pulse detection

TLDR

Endocrine signaling is a key physiological communication mechanism, and many endocrine signals display episodic or pulsatile patterns. The authors aim to provide a versatile, statistically based algorithm for investigating the regulation of endocrine pulse signals. The algorithm defines a pulse as a statistically significant increase in a cluster of hormone values followed by a significant decrease in a second cluster, judged relative to experimental error and allowing the operator to specify cluster sizes for peaks and nadirs. The method is largely insensitive to unstable baseline hormone concentrations, varying pulse amplitudes, widths, or configurations, is minimally dependent on a priori secretion or disappearance rates, has been validated for false‑positive errors across 4–52 % CV, and performs well on LH, FSH, GH, prolactin, ACTH, cortisol, and non‑pulsatile proteins.

Abstract

Endocrine signaling provides one critical means of physiological communication within an organism. Many endocrine signals exhibit an episodic or pulsatile configuration. In an effort to provide a versatile and statistically based algorithm for investigating the regulation of endocrine pulse signals, we have formulated a computerized algorithm in which a pulse is defined as a statistically significant increase in a "cluster" of hormone values followed by a statistically significant decrease in a second cluster of values. The increase or decrease is judged in relation to the actual experimental error expressed by the replicates in the presumptive nadir and peak data clusters. The program permits the operator to specify the cluster sizes of test peaks and pre- and postpeak nadirs. This method is largely insensitive to unstable base-line hormone concentrations and is not adversely affected by varying pulse amplitudes, widths, or configurations within the endocrine series. In addition, the simple statistical basis for this algorithm renders it minimally dependent on explicit or a priori assumptions about rates of hormone secretion or disappearance. The program has been validated for false-positive errors against a wide range of intraseries coefficients of variation (4-52%). We have illustrated its performance for profiles of luteinizing hormone, follicle-stimulating hormone, growth hormone, prolactin, adrenocorticotropic hormone, and cortisol and compared these episodic patterns with those of stable serum constituents (total serum protein and calcium), which do not exhibit pulsatile fluctuation.