The Effect of Imputation of PaO2/FIO2 From SpO2/FIO2 on the Performance of the Pediatric Index of Mortality 3.

Objectives: To investigate if the performance of Pediatric Index of Mortality 3 is improved by including imputed values for the PaO2/FIO2 ratio where measurements of PaO2 or FIO2 are missing.

Design: A prospective observational study.

Setting: A bi-national pediatric intensive care registry.

Patients: The records of 37,983 admissions of children less than 16 years old admitted to 19 ICUs.

Interventions: None.

Measurements And Main Results: Seven published equations describing an association between PaO2/FIO2 and oxygen saturation measured by pulse oximetry (SpO2)/FIO2 were used to derive an alternative variable d100 × FIO2/PaO2 for the Pediatric Index of Mortality 3 variable 100 × FIO2/PaO2. Six equations exclude SpO2/FIO2 values if SpO2 is greater than 96-98%. 100 × FIO2/PaO2 was missing in 72% of patient records primarily due to missing PaO2, d100 × FIO2/PaO2 was missing in 71% of patient records if values of SpO2greater than 97% were excluded or in 17% of patient records if all measurements of SpO2 were included. Univariable analysis supported the inclusion of SpO2 values greater than 97%. Compared to the standard Pediatric Index of Mortality 3 model, two alternative models imputing 100 × FIO2/PaO2 from d100 × FIO2/PaO2 only if 100 × FIO2/PaO2 was missing, or using d100 × FIO2/PaO2 values exclusively, resulted in a small but statistically significant improvements in discrimination of Pediatric Index of Mortality 3 (area under the receiver operator curve 0.9068 [0. 8965-0. 9171]; 0.9083 [0.8981-0.9184]; 0.9087 [0.8987-0.9188], respectively).

Conclusions: Imputation of the PaO2/FIO2 ratio in cases where arterial sampling was not performed resulted in a large reduction in the rate of missing data if all values of SpO2 were included. The imputation technique improved the discrimination of Pediatric Index of Mortality 3; however, the magnitude of the increment in overall model performance was small. A possible benefit of the approach is reducing the potential for bias resulting from variation in practice for invasive monitoring of oxygenation.