Background and Aims

Blood Glucose (BG) prediction models need to be improved to ameliorate BG regulation. Neural Networks (NN) and fractional order calculus are powerful tools for black box modeling. This work combines both approaches to propose the first NN model with fractional order learning algorithm to improve BG prediction.

Methods

The NN model (Fig. 1A), which uses BG, Insulin on board, and carbohydrates on board as inputs, consists of a three-layer NN with 2-2-1 neurons in the input, hidden and output layers, respectively. The NN was trained by a learning algorithm using the Grünwald-Letnikov fractional derivative (2).

Evaluation is performed on 10 T1D patients (aged > 18 years). Training is performed using 10 days of data and test is performed using 5 days of data unseen by the NN during training. RMSE is computed to evaluate accuracy on BG prediction 30-, and 60-min-ahead.

Results

Table 1 displays results reached by the proposed model and results reported in the literature. Fig. 1B shows RMSE reached by the proposed NN-based model on the 10 validation subsets (10 patients, 5 days). As expected, RMSE ± std increases when prediction horizon increases.

Conclusions

This work presented a NN-based BG prediction model trained by a fractional order learning algorithm. The proposed model reached the best performance on predictions 30- and 60-min-ahead, compared with results reported in the literature. Furthemore, the proposed model is the simplest of the compared approaches. In future works, the possibility to predict hypo- and hyperglycemia events by the proposed model will be evaluated.