PREDICTIVE NEURAL NETWORKS FOR LEARNING THE TIME-COURSE OF BLOOD-GLUCOSE LEVELS FROM THE COMPLEX INTERACTION OF COUNTERREGULATORY HORMONES

Diabetes mellitus is a widespread disease associated with an impaired hormonal regulation of normal blood glucose levels. Patients with insu lin-dependent diabetes mellitus (IDDM) who practice conventional insul in therapy are at risk of developing hypoglycemia (low levels of blood glucose), which can lead to severe dysfunction of the central nervous system. In large retrospective studies, up to approximately 4% of dea ths of patients with IDDM have been attributed to hypoglycemia (Cryer, Fisher, & Shamoon, 1994; Tunbridge, 1981; Deckert, Poulson, & Larsen, 1978). Thus, a better understanding of the complex hormonal interacti on preventing hypoglycemia is crucial for treatment. Experimental data from a study on insulin-induced hypoglycemia in healthy subjects are used to demonstrate that feedforward neural networks are capable of pr edicting the time course of blood glucose levels from the complex inte raction of glucose counterregulatory (glucose-raising) hormones and in sulin. By simulating the deficiency of single hormonal factors in this regulatory network, we found that the predictive impact of glucagon, epinephrine, and growth hormone secretion, but not of cortisol and nor epinephrine, were dominant in restoring normal levels of blood glucose following hypoglycemia.