A Multimodal Deep Learning Approach Towards Predicting Severe Retinopathy of Prematurity

Abstract

Purpose : Retinopathy of prematurity (ROP) is a leading cause of visual disability in children worldwide. The risk of severe ROP is primarily related to two factors: lower gestational age (GA) and the consequences of physiologic hyperoxia on retinal development. Oxygen is now tightly regulated to minimize the risk of severe ROP, with frequent adjustments to keep oxygen saturations in range. The purpose of this study was to evaluate whether time-series oxygen data in the electronic health record (EHR) was associated with the onset of type 2 or treatment-requiring (TR-) ROP.

Methods : Demographics, GA and birthweight (BW), and time-series oxygen data for the first month of birth were extracted from the EHR for 230 infants who were born at OHSU, 87 of whom eventually developed type 2 or worse ROP (48 with TR-ROP). Time-series oxygen data included daily mean, maximum, minimum, and coefficient of variance (CV) of inspired oxygen (FiO2) and oxygen saturation (SpO2). We developed two machine learning (ML) models: a random forest and a support vector machine (SVM) using various combinations of birthweight, GA, and oxygen data for prediction of future type 2 or worse ROP. 5-fold cross-validation and recursive feature elimination were used to tune and optimize the prediction models. Also, to explore the predictive value of time-series oxygen data, we developed a multimodal long short-term memory (LSTM) model using demographics and daily oxygen data for 1 month. Mean area under the receiver operating curve (AUROC) and F1 score were used to evaluate the model performance.

Results : With 5-fold cross-validation, the multimodal-LSTM models (Figure 1) had higher performance than the best ML models (SVM using GA and 3 average FiO2 features) and SVM models trained on GA alone (Figure 2, mean AUROC = 0.89±0.04 vs. 0.86±0.05 vs. 0.83±0.04). Also, the multimodal-LSTM models showed the highest F1 score (0.85±0.06) followed by the SVM with 4 variables (0.82±0.07) and SVM with GA alone (0.80±0.06).

Conclusions : Time-series daily oxygen saturation and exposure features can be used to improve the detection of high-risk infants for developing severe ROP and may help improve our understanding of disease pathophysiology.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Figure 1: Architecture of multimodal-LSTM prediction model

 

Figure 2: ROC curves of multimodal-LSTM, SVM with 4 variables, and SVM with GA alone