Table 4.
Description of the parameters related to the model training.
| Parameter | Description |
|---|---|
| Regularization mechanisms | 52/81 (64.2) |
| Dropout | 41/52 (78.9) |
| L1 or L2 regularization | 23/52 (45.1) |
| Others | 21/52 (40.4) |
| Simultaneous regularization techniques (n) | [1, 5], 1.6 (0.8), 1 [1-2] |
| Optimizer | |
| Adam | 39/81 (48.1) |
| Stochastic gradient descent (SGD) | 5/81 (6.2) |
| Adadelta | 5/81 (6.2) |
| Others | 7/81 (8.6) |
| Not reported | 25/81 (30.9) |
| Internal validation | |
| Random split | 54/81 (66.7) |
| Cross-validation | 15/81 (18.5) |
| Others | 3/81 (3.7) |
| Not reported | 9/81 (11.1) |
| Training set size (%) | [14, 90], 73.8 (10.7), 75 [70-80], (92.6) |
| Validation set size (%)a | [4.4, 50], 14.5 (7.5), 10 [10-16.9], (16.1), (12.3) |
| Test set size (%)a | [10, 80], 18.1 (9.5), 16.7 [15-20], (6.2), (16.1) |
| Measure of performance variabilityb | 44/81 (54.3) |
| External validation | 3/81 (3.7) |
| Comparison with | |
| Simpler models | 70/81 (86.4) |
| State-of-the-art models | 37/81 (45.7) |
Categorical parameters are described as N (%), while quantitative parameters as [min, max], mean (SD), median [Q1-Q3], (% studies reported).
a
Stands for the percentages of studies that did not report the information and that did not define those sets.
b
Use of techniques like cross-validation or bootstrapping to quantify the variability of the performance of the model in the internal validation.