Self-organized segmentation of time series: separating growth hormone secretion in acromegaly from normal controls.
The pulsatile pattern of growth hormone (GH) secretion was assessed by sampling blood every 10 min over 24 h in healthy subjects (n = 10) under normal food intake and under fasting conditions (n = 6) and in patients with a GH-producing tumor (acromegaly, n = 6), before and after treatment with the somatostatin analog octreotide. Using autocorrelation, we found no consistent separation in the temporal dynamics of GH secretion in healthy controls and acromegalic patients. Time series prediction based on a single neural network has recently been demonstrated to separate the secretory dynamics of parathyroid hormone in healthy controls from osteoporotic patients. To better distinguish the differences in GH dynamics in healthy subjects and patients, we tested time series predictions based on a single neural network and a more refined system of multiple neural networks acting in parallel (adaptive mixtures of local experts). Both approaches significantly separated GH dynamics under the various conditions. By performing a self-organized segmentation of the alternating phases of secretory bursts and quiescence of GH, we significantly improved the performance of the multiple network system over that of the single network. It thus may represent a potential tool for characterizing alterations of the dynamic regulation associated with diseased states.