Table 3.
Quality metrics (%) of the ML models for NH prediction.
| PH | Sampling/Parameters | RF | LogRLasso | ANN | ||||
|---|---|---|---|---|---|---|---|---|
| CGM | CGM + Clinical Data | CGM | CGM + Clinical Data | CGM | CGM + Clinical Data | |||
| 15 min | OS |
Se
Sp AUC |
93.6 (3.4) 90.1 (2.4) 0.958 (0.011) |
90.9 (2.8) 91.8 (2.3) 0.953 (0.012) |
93.6 (1.9) 91.9 (2.2) 0.962 (0.010) |
93.0 (3.0) 93.0 (2.0) 0.968 (0.014) |
90.5 (5.9) 91.4 (1.6) 0.946 (0.032) |
90.8 (2.5) 89.1 (4.5) 0.935 (0.029) |
| NS |
Se
Sp AUC |
91.8 (1.2) 91.1 (3.9) 0.959 (0.020) |
94.5 (2.6) 91.4 (3.3) 0.97 (0.017) |
93.6 (3.4) 91.2 (2.5) 0.957 (0.021) |
92.4 (2.5) 92.3 (3.7) 0.958 (0.025) |
88.6 (3.6) 92.6 (3.1) 0.934 (0.032) |
90.3 (3.1) 91.0 (1.6) 0.935 (0.027) |
|
| US |
Se
Sp AUC |
88.2 (5.2) 92.7 (2.1) 0.953 (0.023) |
92.3 (3.4) 90.6 (1.3) 0.956 (0.009) |
90.5 (6.7) 91.4 (1.4) 0.947 (0.036) |
90.8 (4.7) 91.2 (2.4) 0.947 (0.018) |
90.0 (4.7) 90.2 (2.8) 0.947 (0.033) |
91.9 (3.7) 88.9 (3.6) 0.945 (0.017) |
|
| 30 min | OS |
Se
Sp AUC |
87.6 (1.9) 88.9 (3.1) 0.927 (0.03) |
86.6 (3.6) 87.0 (2.6) 0.911 (0.019) |
90.4 (1.7) 87.5 (2.2) 0.932 (0.06) |
91.0 (3.5) 87.7 (3.7) 0.94 (0.012) |
87.6 (3.9) 88.0 (4.0) 0.918 (0.031) |
84.6 (5.2) 87.2 (5.5) 0.881 (0.034) |
| NS |
Se
Sp AUC |
87.1 (4.6) 87.1 (6.0) 0.92 (0.036) |
90.4 (4.7) 87.4 (1.6) 0.942 (0.028) |
87.1 (4.0) 90.8 (1.9) 0.928 (0.012) |
86.9 (4.0) 90.3 (1.9) 0.933 (0.012) |
86.6 (3.2) 88.7 (2.2) 0.924 (0.018) |
83.3 (4.2) 86.3 (2.8) 0.881 (0.049) |
|
| US |
Se
Sp AUC |
89.5 (3.6) 86.5 (2.8) 0.912 (0.031) |
92.4 (3.1) 85.3 (1.2) 0.923 (0.021) |
85.1 (5.6) 89.5 (1.8) 0.913 (0.027) |
90.3 (3.2) 86.7 (1.9) 0.92 (0.03) |
85.1 (5.3) 87.5 (2.7) 0.908 (0.028) |
85.2 (3.6) 84.8 (2.2) 0.901 (0.023) |
|
The SD values of the estimates obtained with cross-validation process are shown in the parentheses. The highest AUC values for each PH and ML algorithm are highlighted in bold. Abbreviations: ANN, Artificial Neural Networks; AUC, area under the curve; CGM, continuous glucose monitoring; LogRLasso, Logistic Linear Regression with Lasso regularization; NP, nocturnal hypoglycemia; PH, prediction horizon; RF, Random Forest; OS, oversampling; NS, no sampling; US, undersampling; Se, sensitivity; Sp, specificity.