Model calibration
| Model | Calibration measures* | APO predicted risk† | |||||
| Brier‡ | Reliability‡ | P value§ | 1 (low) | 2 | 3 | 4 (high) | |
| Regression models | Actual APO rate* | ||||||
| Stepwise-selection (LR-S) | 0.14 | 0.013 | <0.001 | 0.11 | 0.35 | 0.38 | 0.64 |
| Penalised (LASSO) | 0.13 | 0.007 | 0.13 | 0.11 | 0.40 | 0.39 | 0.74 |
| Neural networks (NN) | |||||||
| One hidden layer (NN-1) | 0.16 | 0.033 | <0.001 | 0.12 | 0.27 | 0.40 | 0.49 |
| Two hidden layers (NN-2) | 0.18 | 0.054 | <0.001 | 0.11 | 0.29 | 0.32 | 0.43 |
| Tree-based | |||||||
| Random forest (RF) | 0.13 | 0.004 | 0.55 | 0.09 | 0.35 | 0.47 | 1.00¶ |
| Gradient boosting (GB) | 0.14 | 0.009 | 0.19 | 0.13 | 0.33 | 0.38 | 0.83 |
| Support vector machine (SVM) | |||||||
| SVM-RBF | 0.13 | 0.005 | 0.62 | 0.10 | 0.40 | 0.61 | 0.61 |
| Ensemble | |||||||
| SuperLearner (SL) | 0.12 | 0.003 | 0.82 | 0.09 | 0.40 | 0.60 | 0.75 |
*Average across five independent, 10-fold cross-validations.
†APO predicted risk: 1: ≤25%, 2: 26%–50%, 3: 51%–75%, 4: >75%.
‡Agreement between predicted and observed APO; low scores indicate better agreement.
§P<0.05 indicates lack of fit using Spiegelhalter goodness-of-fit test.
¶Only one individual (who experienced an APO) had a prediction >75%.
APO, adverse pregnancy outcome; LASSO, least absolute shrinkage and selection operator; LR-S, logistic regression wih stepwise selection.