ROP is a proliferative vitreoretinopathy affecting preterm infants, which is a predominant etiology of childhood blindness and vision impairment globally. In 2010, an estimated 184,700 infants of 14.9 million premature infants developed any stage of ROP worldwide, 20,000 of whom suffered blindness or sever visual impairment from ROP [21] Timely screening of ROP is crucial, whereas barriers exist to its effective implementation, including scarce health resources in low- and middle-income regions, and unnecessary examination under varying screening criteria across the world [1]. Currently, noninvasive prediction of ROP has been highlighted, aiming to safely ease the screening burden without missing severe cases, especially for regions where the number of ophthalmologists is limited and the conventional ROP screening is not readily accessible.
A recent study reported an individual predictive model based on gestational age, birth weight, and sex for infants with a gestational age of 24 to 30 weeks, enabling early risk estimation for receiving ROP treatment [22]. The predictive ability of this novel model was confirmed by test with excellent AUCs varying from 0.84 to 0.94. Furthermore, the input variables were simple, facilitating the general use of the model. However, infants born at gestational age < 24 weeks or > 31 weeks could not be included. In China, more mature and larger infants are also at risk of ROP [23]. Therefore, the inclusion criteria in our study were broader than the screening guidelines recommended in the USA (gestational age < 30 weeks, birth weight < 1500 g) and UK (gestational age < 32 weeks, birth weight < 1500 g) [24, 25]. Totally, 591 infants with mean gestational age and birth weight of 31.0 ± 2.5 weeks and 1.5 ± 0.4 kg, respectively, were included.
The pathogenesis of ROP is still not fully unveiled, which appears to be multi-factorial and is associated with a number of epigenetic modifications by perinatal factors [26]. In the present study, 12 risk factors of ROP were selected, including gestational age, birth weight, caesarean delivery, Apgar score at five minutes, bronchopulmonary dysplasia, respiratory distress syndrome, intracranial hemorrhage, hypoxic ischemic encephalopathy, mechanical ventilation, blood transfusions, necrotizing enterocolitis, and patent ductus arteriosus. Previous studies have examined the interrelationships of the various independent risk factors in the presence of ROP. Giapros et al. demonstrated that ROP was independently related to both low gestational age and the presence of chronic lung disease, which were interrelated in the development of ROP. The combined risk effect on ROP was found to be larger than the sum of the individual risk effects, revealing that there was a positive interaction on an additive scale [27]. In the study of Chen et al., sepsis, oxygen exposure, and low gestational age, were found to be independently associated with a significantly increased risk of ROP with significant interrelationships [28]. These findings might imply that artificial neural network, which simulates the function and structure of biological neural network to establish non-linear mathematical models could perform better when constructing predictive models of ROP with various pathogenic factors and complicated interrelationships among them [16].
Nowadays, BP neural network algorithm, a multi-layer feed forward network, has been widely used in the field of medicine, especially to assist diagnosis and clinical decision-making, with its strong fault tolerance, adaptiveness, nonlinear comprehensive reasoning ability, and the powerful ability to solve co-linearity and interactions between variables [19]. Yao et al. proposed a BP neural network for predicting diabetic retinopathy based on biochemical and metabolic parameters, which achieved an AUC of 0.84 with a sensitivity of 0.73 and a specificity of 0.83 [29]. Nevertheless, further improvement is required concerning its limited study cohort and unsatisfactory sensitivity before its application in screening practice. In the present study, based on the risk factors extraction, predictive models were established using BP neural networks, and GA-BP neural networks, respectively, and displayed encouraging performance on detection of ROP with reasonable AUCs. In comparison to the logistic regression model, the BP neural network model exhibited superior performance for ROP prediction, which was further improved by applying GA-BP neural network with higher sensitivity, accuracy and AUC. We also validated and compared the prediction performances of ROP score prediction model with our prediction models, the AUC, sensitivity, specificity and accurancy of our prediction models and ROP Score were comparable in predicting ROP.
A potential criticism of our study is the small sample size, which may bring the doubt of the optimal cut-off value and predictive ability of the models owing to insufficient machine learning. Artificial neural networks would be better trained using large-scale datasets [16]. The proposed models need to be further validated in larger cohorts. Another limitation of our study is that ROP risk-stratification for optimization of timing and frequency of the screening examinations and accurate treatment cannot be realized by our predictive models, which might be explored in our future study to better represent the current clinical scenario and achieve the goal of precision medicine.
In conclusion, we developed predictive models for noninvasive ROP detection using BP neural network, and GA-BP neural network, respectively, with the hope to safely spare unnecessary screening. GA-BP neural network exhibited superior predictive ability for ROP when dealing with its non-linear clinical data.