Table 3.

Prediction accuracy of the proposed deep learning models along with other comparing methods based on the ±10% range of the actual glucose level value.

Patient #	LSTM-NN-TF-DTWa (with transfer learning) (%)	LSTM-NN-TF-ALLb (with transfer learning) (%)	LSTM-NNc (without transfer learning) (%)	ANNd (%)	KNNe regression (%)	Ridge regression (%)	Kernel ridge regression (%)	Moving average (last 3 days) (%)
1	76.67	73.33	73.33	73.33	46.67	76.67	43.33	73.33
2	86.67	86.67	83.33	86.67	53.33	86.67	50.00	76.67
3	60.00	60.00	60.00	53.33	50.00	53.33	43.33	53.33
4	85.71	85.71	85.71	71.43	42.86	71.43	71.43	71.43
5	63.33	46.67	16.67	13.33	36.66	66.67	66.67	76.66
6	46.00	36.67	33.33	26.67	10.00	36.67	26.67	43.33
7	33.33	26.67	33.33	30.00	33.33	26.67	26.67	20.00
8	63.33	56.67	66.67	56.67	46.67	60.00	43.33	80.00
9	60.00	16.67	60.00	20.00	26.67	40.00	20.00	40.00
10	73.33	70.00	73.33	73.33	56.67	56.67	63.33	63.33

^aLSTM-NN-TF-DTW: second transfer learning strategy.

^bLSTM-NN-TF-ALL: first transfer learning strategy.

^cLSTM-NN: without transfer learning.

^dANN: artificial neural network

^eKNN: k-nearest neighbors.