Retinopathy of Prematurity in Late Preterm Twins with a Birth Weight Discordance: Can it be Predicted by Artificial Intelligence?

Abstract

Objectives: The objective is to predict the development of retinopathy of prematurity (ROP) in discordant twins using a machine learning approach.

Methods: The records of 640 twin pairs born at 32-35 weeks gestational age (GA) with birth weight (BW) discordance were evaluated retrospectively. The infants' gender, GA, postmenstruel age at examination, BW, discordance rate, ROP Stages and Zones, and treatment options were recorded. The variables were used to develop a model to predict the development of ROP. Machine learning models were used for algorithm training and 10-fold cross-validation (CV) was applied for validation. The main measures were reported as sensitivity, specificity, receiver operating characteristic curve, and the area under the curve.

Results: A total of 640 twin pairs underwent ophthalmic examination, of which 55 (4.3%) were ROP. The infants' GA was 33.56±1.01 weeks (32-35 weeks) and BW was 1996±335 g (1000-3400 g). The mean discordance rate of the infants was 11.8±9.7% (0.0-53.9%). Using operating points, the Decision Tree algorithm detected ROP prediction with 71% sensitivity and 80% specificity in CV, while the Multi-Layer Perceptron algorithm detected 70% sensitivity and specificity. In addition, the X-Tree and Random Forest algorithms detected ROP prediction with 84% and 80% specificity, respectively.

Conclusion: The results of this study support that BW discordance may be effective in the development of ROP in preterm twins and that artificial intelligence models can predict the development of ROP in accordance with clinical findings.

Keywords: Artificial intelligence; machine learning; preterm; retinopathy of prematurity; twin.

Figure 1

A summary of the study procedure. (a) One ophthalmic center records were included. Only twin pairs born at 32–35 weeks of gestation with birth weight (BW) discordance were included in the study. (b) The variables such as gender, gestational age, BW, discordance rate, and postmenstruel age at examination were used to develop a model. (c) We used machine learning models for algorithm training and 10-fold cross-validation for validation. (d) Model evaluation was reported as sensitivity and specificity and was graphically described via the receiver operating characteristic curve and the quantitative performance of the model was summarized by the area under the curve.

Figure 2

A receiver operating characteristic curve and area under the curve of the predictive models.