Performance of predictive models for preterm birth (n=4357)*
| Predictive models | Model calibration | Risk stratification capacity n (%) | Model discrimination† | |||||||
| Predicted probability of PTB (%) | Observed PTB n (%) 95% CI | Sensitivity (%) | Specificity (%) | Classification accuracy (%) | PPV (%) | NPV (%) | LR+ (%) | LR− (%) | ||
| Conventional logistic regression model with individual level predictors to that is, parity, ethnicity, body mass index, smoking, depression, and household income | <5 | 42 (4.81) 3.43 to 6.03 | 873 (20.04) | – | – | – | – | – | – | |
| ≥5–10 | 197 (6.96) 6.02 to 7.81 | 2832 (65.00) | 85.76 | 22.43 | 26.54 | 7.66 | 95.44 | 1.10 | 0.63 | |
| ≥10–15 | 77 (12.56) 9.99 to 15.96 | 613 (14.07) | 20.12 | 89.42 | 84.58 | 12.43 | 93.70 | 1.90 | 0.89 | |
| ≥15 | 4 (10.26) 2.82 to 24.37 | 39 (0.90) | 1.55 | 99.14 | 92.31 | 8.82 | 93.03 | 1.80 | 0.99 | |
| Multi-level logistic regression model with neighbourhood deprivation index and individual level predictors | <5 | 26 (2.22) 1.50 to 3.22 | 1177 (27.01) | – | – | – | – | – | – | – |
| ≥5–10 | 197 (7.30) 6.40 to 8.37 | 2690 (61.74) | 91.80 | 28.50 | 33.09 | 9.12 | 97.80 | 1.28 | 0.29 | |
| ≥10–15 | 75 (17.24) 13.97 to 21.09 | 435 (9.98) | 29.40 | 90.20 | 85.83 | 19.00 | 94.20 | 3.00 | 0.78 | |
| ≥15 | 18 (32.73) 21.60 to 46.20 | 55 (1.26) | 5.70 | 99.10 | 92.30 | 32.80 | 93.10 | 6.22 | 0.95 | |
| Multi-level logistic regression model with neighbourhood median personal income and individual-level predictors | <5 | 31 (2.64) 1.86 to 3.73 | 1174 (26.97) | – | – | – | – | – | – | – |
| ≥5–10 | 192 (7.16) 6.24 to 8.19 | 2683 (61.58) | 90.30 | 28.30 | 33.13 | 8.95 | 97.40 | 1.26 | 0.34 | |
| ≥10–15 | 81 (18.08) 14.78 to 21.92 | 448 (10.28) | 29.40 | 89.90 | 85.85 | 18.60 | 94.20 | 2.92 | 0.78 | |
| ≥15 | 12 (23.08) 13.52 to 36.53 | 52 (1.19) | 3.80 | 99.00 | 92.20 | 23.10 | 92.10 | 3.84 | 0.97 | |
*Model performance was assessed in the original sample (study sample).
†Model discriminatory was calculated using cumulative row values as different cut-offs to define high risk, for example, if all women with a model predicted probability of a preterm birth of 5% or higher are considered to have a positive test, model with deprivation index and individual level predictors would have a sensitivity of 91.80% and specificity of 28.50%.
LR+, positive likelihood ratio; LR–, negative likelihood ratio; NPV, negative predictive value; PPV, positive predictive value; PTB, preterm birth.