Insulin pumps and continuous glucose monitors enable automatic control of blood glucose (BG) levels for patients with type 1 diabetes. Such controllers should carefully assess the likely future BG levels before injecting insulin, since the effects of insulin are prolonged, potentially deadly, and irreversible. Meals pose a strong challenge to this assessment as they create large, fast disturbances. Fortunately, meals have consistent and predictable effects, if their size and start time are known. We present a predictive algorithm that embeds meal detection and estimation into BG prediction. It uses a multiple hypothesis fault detector to identify meal occurrences, and linear Kalman filters to estimate meal sizes. It extrapolates and combines the state and state covariance estimates to form a prediction of BG values and uncertainties. These inputs enable controllers to assess and trade off the acute risks of low and chronic risks of high BG levels. We evaluate the predictor on simulated and clinical data.

This content is only available via PDF.
You do not currently have access to this content.