Quantifying Oxygen Exposure for Predicting Treatment-Requiring Retinopathy of Prematurity

Abstract

Purpose : Oxygen exposure and extreme prematurity are among two of the strongest risk factors for severe retinopathy of prematurity (ROP), however current screening criteria focuses on birth weight and gestational age (GA) at birth. Improved risk models may reduce the screening burden for low-risk infants and improve disease detection for infants with the most severe ROP, including aggressive posterior ROP (APROP). The purpose of this study was to evaluate the additive predictive value of quantifying oxygen exposure in early life for detection of treatment-requiring (TR-) ROP and APROP.

Methods : Demographics and oxygen exposure parameters were manually extracted from the electronic health record for each week of life (WOL) for 244 infants, 33 of whom eventually developed TR-ROP and 5 of whom developed APROP. Cumulative minimum, maximum, and total fraction of inspired oxygen (FiO2) were calculated by summing values per WOL. Using 5-fold cross-validation, models using various combinations of birthweight, GA and FiO2 were trained using random forest tuned with randomized grid search for prediction of future TR-ROP. Performance was evaluated using mean area under the receiver operating curve (AUROC) and precision-recall curve (AUPRC). To test the predictive value of oxygen exposure for APROP, cumulative minimum FiO2 exposure was also plotted against eventual ROP outcome (no treatment, TR-ROP without APROP, or APROP) and an AUROC score was generated.

Results : On 5-fold cross-validation, the models trained on GA + cumulative minimum FiO2 exposure had slightly higher performance than the models trained on GA alone (Figure 1, mean AUROC = 0.93±0.06 vs. 0.91±0.06, AUPRC = 0.76±0.08 vs. 0.74±0.13 respectively) for TR-ROP. For APROP, the AUROC of cumulative minimum FiO2 exposure was 0.92 with clear dose response between oxygen exposure and level of ROP (Figure 2).

Conclusions : Quantitative oxygen exposure variables can be extracted and used to augment the identification of high-risk infants for developing TR-ROP, including APROP. Future work should focus on prospectively evaluating models that account for oxygen exposure.

This is a 2021 ARVO Annual Meeting abstract.

 

5-Fold cross-validation curves for combinations of random forest models accounting for gestational age and oxygen exposure

 

Box plot of oxygen exposure vs. ROP outcome