Multi-lag stacking for blood glucose level prediction

This work investigates blood glucose level prediction for type 1 diabetes in two horizons of 30 and 60 minutes. Initially, three conventional regression tools—partial least square regression (PLSR), multilayer perceptron, and long short-term memory—are deployed to create predictive models. They are trained once on 30 minutes and once on 60 minutes of historical data resulting in six basic models for each prediction horizon. A collection of these models are then set as base-learners to develop three stacking systems; two uni-lag and one multi-lag. One of the uni-lag systems uses the three basic models trained on 30 minutes of lag data; the other uses those trained on 60 minutes. The multi-lag system, on the other hand, leverages the basic models trained on both lags. All three stacking systems deploy a PLSR as meta-learner. The results obtained show: i) the stacking systems outperform the basic models, ii) among the stacking systems, the multi-lag shows the best predictive performance with a root mean square error of 19.01 mg/dl and 33.37 mg/dl for the prediction horizon of 30 and 60 minutes, respectively.