Table 2: Results of testing differences in prediction accuracies between the created ANN, and MLR models
ANN model | ANN architecture | MSE on the test sample | SMAPE on the test sample (%) |
Fasting glucose variability | 2 hidden layers | 5.23 | 23.69 |
Postprandial glucose variability | 1 hidden layer | 30.58 | 33.24 |
Increased HbA1c | 1 hidden layer | 0.17 | 1.47 |
ANN: Artificial neural networks; HbA1c: glycosylated hemoglobin A1c; MSE: mean squared error; SMAPE: symmetric mean average percentage error.
MLR model | Model's size | MSE on the test sample | SMAPE on the test sample (%) |
Fasting glucose variability | number of explanatory variables = 14** | 2.47 | 69.50 |
Postprandial glucose variability | number of explanatory variables = 10 | 6.79 | 99.64 |
Increased HbA1c | number of explanatory variables = 16 | 1.66 | 13.44 |
MLR: Multiple linear regression; **the variables "Dg of the upper gastrointestinal tract disorders" and "Antidiabetic drugs in use > 6 months" was split into the two components.