Early detection, prevention, and intervention of ROP are key to reducing incidence and mortality and alleviating the huge socio-economic burden caused by ROP. Screening out the factors with significant statistical differences from numerous risk factors and exploring the best combination of risk factors are an effective strategy for predictive diagnostic, targeted prevention and personalized treatment of ROP [13, 29]. Low-risk preterm infants should receive fewer examinations to alleviate scarce medical resources, while high-risk infants (premature infants with ROP who require treatment) should be treated with medication or laser surgical interventions to reduce the risk of blindness, which effectively promotes the paradigm shift from reactive medicine to the advanced approach by utilizing the 3 PM framework. The active promotion of the PPPM framework has been applied in many fields, making great contributions to the health quality improvement including fundus diseases [32], diabetes [33], cancers [34], etc. In this study, we retrospectively evaluated an AI system for ROP risk and TP-ROP risk prediction developed for preterm infants in Chinese population based on the i-ROP cohort study dataset from SZEH. First, statistical single-factor analysis was used to determine the risk factors with statistically significant differences among the five risk factors for ROP and TR-ROP. Then, based on the risk factors with significant statistical differences found in the single-factor analysis, a binary logistic regression model was used for multi-factor analysis to determine the final risk prediction factors. Finally, with statistically significant risk factors as input, two machine learning methods of logistic regression and decision tree and a deep learning method of MLP were used to explore the optimal prediction method for ROP and TR-ROP risk prediction.

We observed that among the three AI-based ROP risk prediction models, logistic regression has better comprehensive performance. Only using BW and GA can identify approximately 23% of infants who are at risk of ROP before being diagnosed with ROP, while simultaneously excluding 97.75% of those at low risk (Table 3, Fig. 3-A). When other risk factors were added to the combination of GA + BW, both logistic regression and MLP showed improvement in specificity, which indicated that more low-risk people could be excluded. In MLP, adding gender can achieve better sensitivity compared to the combination GA + BW (Fig. 3). In addition, we knew that in the results of single-factor analysis and multi-factor analysis for risk factors (Tables 1 and 2), MD is a risk factor with significant statistical difference, but there is no significant improvement after adding MD to the ROP risk prediction methods. Instead, a certain performance decline was observed in all three methods (AUC and AUCPR exhibited a decrease in Fig. 2). On the other hand, although there is no statistically significant difference in the risk factor MB, the model’s specificity was improved by adding MB to logistic regression (Fig. 3). Among the three AI-based TR-ROP risk prediction models, logistic regression demonstrated relatively superior performance. Using the risk factor combination GA + BW + Gender + MD + MB, 75.36% of children who need treatment could be identified, while excluding nearly 75% of infants who did not require treatment (Table 3, Fig. 5-D). Moreover, we found that there was no significant difference in AUC between the two methods of logistic regression and MLP in the four risk factor combinations (Table 4).

The comprehensive confusion matrix for each model’s five-fold cross-validation results in ROP risk prediction. A: GA + BW, B: GA + BW + Gender, C: GA + BW + Gender + MD, D: GA + BW + Gender + MD + MB, for logistic regression. E: GA + BW, F: GA + BW + Gender, G: GA + BW + Gender + MD, H: GA + BW + Gender + MD + MB, for decision tree. I: GA + BW, J: GA + BW + Gender, K: GA + BW + Gender + MD, L: GA + BW + Gender + MD + MB, for MLP

The comprehensive ROC curve and PR curve for each model’s five-fold cross-validation results in TR-ROP risk prediction. A, ROC curve for logistic regression. B, ROC curve for decision tree. C, ROC curve for MLP. D, PR curve for logistic regression. E, PR curve for decision tree. F, PR curve for MLP

The comprehensive confusion matrix for each model’s five-fold cross-validation results in TR-ROP risk prediction. A: GA + BW, B: GA + BW + Gender, C: GA + BW + Gender + MD, D: GA + BW + Gender + MD + MB, for logistic regression. E: GA + BW, F: GA + BW + Gender, G: GA + BW + Gender + MD, H: GA + BW + Gender + MD + MB, for decision tree. I: GA + BW, J: GA + BW + Gender, K: GA + BW + Gender + MD, L: GA + BW + Gender + MD + MB, for MLP

Nomograms of prevention models. A: Nomogram of ROP risk prediction, B: Nomogram of TR-ROP risk prediction

It can be seen from the result analysis that the comprehensive performance of logistic regression is better, BW + GA risk factor combination can obtain the best performance in ROP risk prediction, and the risk factor combination GA + BW+ Gender + MD + MB has the best performance in TR-ROP risk prediction. We established a new ROP and TP-ROP prediction model based on the combination of these risk factors. The formula of ROP risk prediction model was as follows:

$$\textrm\ \textrm\ \textrm=\frac}^_1-0.0011_2}}$$

(1)

The formula of TR-ROP prediction model is as follows:

$$\textrm\ \textrm\ \textrm-\textrm=\frac}^_1-0.0012_2-0.3617_3+0.2714_4+0.1217_5}}$$

(2)

Where x1 represents GA, x2 represents BW, x3 represents gender, x4 represents MD, and x5 represents MB. In the context of PPPM [35], efforts to incorporate this risk prediction paradigm into current clinical workflows would be advantageous, as it has the potential to become a promising tool for early detection and treatment of ROP disorders in infants.

In terms of prevention [35], ROP was one of the main causes of blindness in children, and early detection can help prevent further deterioration of the condition in time and stop the progression of the disease. Our findings indicated that using the five easily available ROP risk factors for early ROP risk and TR-ROP risk prediction can help reduce the burden on ophthalmologists and improve screening efficiency, especially in low and middle-income regions. In this study, for a targeted ROP prevention model designed for a specific Chinese population, we further simplified the model into nomograms as shown in Fig. 6, so that it can be more convenient to enter the clinical application.

The above situation was obtained from the analysis of a specific Chinese population, and the model is unlikely to generalize well “out of the box” to populations different from the Chinese screening population. Depending on epidemiological and demographic risk factors, the prediction models need to be re-adjusted based on local disease epidemiology, and different risk factor combinations may have different performance because the incidence of ROP varies geographically and over time [29, 30]. We believed that these results prove that the AI-based risk prediction model can effectively reduce the number of ROP examinations and associated physiological stress in low-risk infants, thereby improving the efficiency of ROP screening, and can also be used as an epidemiological tool to monitor NICU-level ROP risk and TP-ROP risk prediction across regions and time. Finally, quantitative monitoring that combines GA and BW, the two strongest risk factors for developing ROP, with other risk factors could allow for earlier and more consistent diagnosis of TR-ROP, thereby minimizing the overall risk of adverse outcomes.

This model may also be easier to implement than the existing ROP risk prediction models. Coyner et al. [13] proposed a method for TR-ROP risk prediction using a combination of GA and Vascular Severity Score (VSS) factors and achieved good performance. However, in this method, an additional AI model needs to be trained to obtain the VSS based on the patient’s fundus images, which is relatively cumbersome in clinical practice because the images are not part of the standard of care and digital fundus cameras may be expensive, so in some low- or middle-income countries or regions, fundus images of infants are not easily available. Moreover, the best AUC obtained by this method was 0.82, while our prediction model could reach 0.83. The specificity of our proposed risk prediction model was much higher than that of the ROP prediction model of the Children’s Hospital of Philadelphia [36], which uses BW + GA + weight gain to predict the risk of type II ROP and TR-ROP, and the highest specificity was 53.4%, while in in our TR-ROP prediction model, the specificity reached 74.83%.

There are many risk factors for ROP [29], and this study was based on the analysis of a specific Chinese population. In practice, when the model was extended to a population different from the Chinese screening population, it was recommended to appropriately adjust the input parameters of the model according to the local disease epidemiology. Different combinations of risk factors may have different impacts on the performance of the risk prediction model. This paradigm of ROP and TR-ROP risk prediction by using ROP risk factors as input to the model will enable high-risk infants receive earlier and more accurate diagnosis and treatment, while minimizing the over-examination of low-risk infants, and better reduce or eliminate the occurrence of ROP-related blindness. In this study, the overall sensitivity of the ROP risk prediction model was relatively low, and it may be necessary to continue to integrate oxygen exposure, intraventricular hemorrhage, neonatal sepsis, necrotizing enterocolitis, thrombocytopenia, and other previously associated risk factors [29] to further improve the model’s sensitivity.

For the automatic identification of ROP, many deep learning (DL) algorithms were currently developed based on color fundus photography images. For example, Brown et al. [23] used two CNNs to realize the automatic diagnosis of plus disease in ROP. First, a U-Net architecture was used to segment blood vessels, and then an Inception network [37] was used to realize the classification task of plus disease in ROP. Furthermore, there were studies using DL algorithms to achieve quantitative assessment of ROP risk severity. For example, Taylor et al. [24] used a DL algorithm for vascular severity scoring to identify TR-ROP by quantifying clinical disease progression, while Kellyn et al. [38] used a vascular severity score based on DL to describe aggressive posterior retinopathy of prematurity quantitatively. In addition, there are also some studies to implement automatic classification tasks of ROP. For example, Zhang et al. [39] used a VGG-16 network to realize the automatic classification task of ROP. Wang et al. [40] developed an automatic ROP detection system using two deep neural networks, in which the Id-Net network was used to identify the ROP, the Gr-Net network was further used to grade the severity of ROP. With the popularity of smartphones, smartphone-based cameras may become a substitute for professional color fundus photography, and the acquisition of infant fundus images may become more convenient. According to fundus images, more clinical characteristics can be obtained, such as the development of blood vessels in different fundus zones. Whether combining that information with the factors previously involved can further improve the prediction performance of the model is worth exploring. If possible, this would not only greatly reduce the burden of ROP screening, but also enable high-risk infants to receive timely treatment and reduce the burden on society. In any case, the risk prediction model proposed in this study can lay a basic framework for a new ROP prediction model, which can at least achieve ROP risk prediction and TR-ROP risk prediction only by inputting GA and BW, without requiring a complete ophthalmoscopy. This paradigm allows low-risk infants to receive fewer examinations, while high-risk infants receive earlier, more precise diagnosis and treatment. This could allow rural areas and low- and middle-income countries to make better use of scarce resources.