Table 4.
Model performance for identifying a SpO2 < 93% (validation dataset).
| Hypoxemia risk (n/N)a | Crude OR (95% CI) | Mean predicted hypoxemia, % | C-statistic (adjusted for optimism;b 95% CI) | Sensitivity | Specificity | PPV | NPV | |
|---|---|---|---|---|---|---|---|---|
| IMCI guideline model | ||||||||
| Non-case | 316/3,460 | Ref | 9.1% | 0.563 (0.564; 0.528, 0.600) | 20.6% | 92.1% | 23.2% | 90.9% |
| Case | 82/353 | 3.011 (2.292, 3.955) | 23.2% | – | ||||
| Total | 398/3,813 | – | – | |||||
| WHO-composite model | ||||||||
| Non-case | 172/2,857 | Ref | 6.0% | 0.677 (0.677; 0.635, 0.712) | 56.8% | 78.6% | 23.6% | 94.0% |
| Case | 226/956 | 4.833 (3.900, 5.989) | 23.6% | – | ||||
| Total | 398/3,813 | – | – | |||||
| Independent LASSO model | ||||||||
| −2 to −1 | 46/1,124 | Ref | 4.1% | 0.745 (0.742; 0.696, 0.779) | 46.7% | 90.7% | 36.9% | 93.6% |
| 0 | 57/1,179 | 1.191 (.800, 1.771) | 4.8% | – | ||||
| 1–2 | 109/1,006 | 2.848 (1.995, 4.065) | 10.8% | – | ||||
| 3–13 | 186/504 | 13.707 (9.697, 19.376) | 36.9% | – | ||||
| Total | 398/3,813 | – | ||||||
| Composite LASSO model | ||||||||
| −1 | 44/1,117 | Ref | 3.9% | 0.752 (0.750; 0.701, 0.787) | 49.0% | 89.5% | 35.3% | 93.8% |
| 0 | 53/1,164 | 1.163 (.773, 1.750) | 4.6% | – | ||||
| 1 | 42/458 | 2.462 (1.589, 3.814) | 9.2% | – | ||||
| 2 | 64/522 | 3.408 (2.286, 5.079) | 12.3% | – | ||||
| 3–11 | 195/552 | 13.320 (9.402, 18.871) | 35.3% | – | ||||
| Total | 398/3,813 | |||||||
SpO2 indicates peripheral arterial oxyhemoglobin saturation; OR, odds ratio; CI, confidence interval; PPV, positive predictive value; NPV, negative predictive value; IMCI, integrated management of childhood illnesses; WHO, World Health Organization; LASSO, least absolute shrinkage and selection operator reduction method for selection; all models represented here were fit using mixed effects logistic regression with a random effect for country (after selection and refitting for the independent and composite LASSO models).
Hypoxemia risk: numerator (number of hypoxemia cases) and denominator (total number of children with the relevant clinical score).
The independent and composite LASSO models AUC compared using Delong method with Bonferroni correction (p = 0.480).
n is the number of observed hypoxemic participants; N is the total number of participants.
C-statistics are adjusted for optimism.