Applying neural networks to adjust insulin-pump doses

Programming appropriate insulin-dose levels is a common diabetic pump-user problem. We developed a neural-network advisory system that suggests the appropriate next-time insulin dose based on short historical discontinuous blood-glucose measurements and insulin doses settings. Diabetologists' high level decision taking process have been successfully learned. Our database consists of 25000 recorded data from 747 insulin-pump users under medical supervision. The efficient data concept is introduced. Training with efficient learning data allowed us to achieve very good generalisation. A portable neural-network controlled insulin-pump device is designed. A complete insulin advisory system including our algorithm is currently under clinical test. Preliminary results demonstrate that the performances of the neural-networks are equivalent to those of the physician.