From Eye Examination to Early Cognitive Evaluation for Preterm Newborns: An Explainable Machine Learning Approach

From Eye Examination to Early Cognitive Evaluation for Preterm Newborns: An Explainable Machine Learning Approach | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article From Eye Examination to Early Cognitive Evaluation for Preterm Newborns: An Explainable Machine Learning Approach Margaret Ming-Chih Ho, Man-Lin, Mai, Wei-Yang Lin, Yu-Shu Huang, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7172977/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Purpose To use machine learning (ML) to identify the important retinal features associated with cognitive development in newborns with prematurity in early childhood, with the goal of screening preterm babies with poor neurodevelopment for early intervention. Methods We retrospectively reviewed the medical charts of 163 infants (326 eyes) born at Chang Gung Memorial Hospital between 2011 and 2024 who underwent IQ testing and scored < 85 or ≥ 115. Essential characteristics considered predictors come from perinatal variables and eye examination for retinopathy of prematurity (ROP) with optical coherence tomography (OCT) images. The ML model was trained using Random Forest (RF) to collectively associate the predictors with the cognitive assessment outcome. SHAP (Shapley Additive exPlanations) was used to identify the clinical features most predictive of the mental development of a preterm baby. Results The overall accuracy of the ML model was 83.44%, combining all the clinical characteristics, including retinal thickness and retinal thickness difference in bilateral eyes obtained from OCT. The SHAP analysis reveals that lower birth weight, higher stage of ROP, and lower zone development are highly associated with lower IQ scores in our study cohort. Conclusions Through an ML approach, this study identified the BW and the stage of ROP and zone as the leading ocular features associated with cognitive outcomes in preterm newborns in early childhood. It paved the way for timely interventions to improve long-term neurodevelopmental outcomes. Ophthalmology Pediatrics Artificial Intelligence and Machine Learning cognitive outcome intelligence quotient machine learning prematurity of retinopathy preterm infant Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Developmental intellectual disability (DID) is the second most prevalent developmental disability among children under 20 years old, following hearing loss. 1 It has become a global concern, affecting 0.2–1.5% of children worldwide. 2 DID is a complex and multifactorial process influenced by genetic, environmental, and perinatal factors. Without early intervention, DID might become a lifelong problem and affect the children’s quality of life. Given the prevalence of DID and its significant impact on the quality of life of children and their families, it is crucial to identify children at high risk of DID for early intervention. Retinopathy of prematurity (ROP) is a retinal vascular disease that may lead to visual impairment and is associated with central nervous system (CNS) comorbidities, such as intraventricular hemorrhage (IVH) and restricted brain development, resulting in cognitive impairment 3 . Previous studies have shown that the presence and severity of ROP may be linked to poor neurodevelopmental outcomes. 4, 5 Other examinations, such as brain magnetic resonance imaging (MRI), were also used to evaluate the risk factors for poor cognitive outcomes in children in previous studies. 6, 7 However, limitations such as high cost, limited accessibility, and the need for sedation, make MRI less practical in routine neonatal care. The intelligence quotient (IQ) score is standardized to have a mean of 100 and a standard deviation of 15. 8 An IQ score of 85-115 is considered average, while scores below 85 are considered below-average and scores more than 115 indicate above-average. Ideally, a statistical threshold value should be placed at 85 to separate children needing early intervention from those of normal development. However, such a statistically determined threshold does not necessarily map to a clear cut in the feature space, as shown in Figure 1(a) : projecting the feature space to two dimensions using t-SNE, the group with IQ ≥ 85 spans the entire feature space, completely overlapping with the IQ < 85 group. This can pose challenges in identifying prognostic factors for children with poor cognitive development. To address this, we excluded children with a normal IQ (IQ between 85 and 115) and focused on those with below-average IQ (IQ < 85) and those with above-average IQ (IQ ≥ 115). Higher separability between the two groups is confirmed by Figure 1(b) . By comparing the clinical and ocular features of children with below-average IQ to those with above-average IQ, we can delineate ocular differences between the two groups and identify the clinical characteristics and ocular features that contribute to poor cognitive outcomes for early intervention. Recent advances in artificial intelligence (AI) and machine learning (ML) approaches profoundly impact various healthcare domains. Unlike conventional linear regression or correlation coefficient methods that typically analyze one feature at a time, the ML approach can efficiently analyze multiple risk factors and their interplay in one setting. Furthermore, the ML approach might offer novel insights into detecting factors that may not have been considered relevant if chosen manually for analysis with traditional methods. Prognostic factors reported in previous studies 6, 9 associated with poor cognitive outcomes in preterm infants included lower gestational age (GA), male sex, nonwhite race, lower level of parental education, poor intrauterine growth, diabetes, lower birth weight, and hospitalization. Nevertheless, none of these studies have discussed the degree of contribution or the interplay of these factors to the cognitive outcome in early childhood with ML techniques. By leveraging basic clinical features and ocular examinations, we aim to identify the key features associated with poor cognitive development in early childhood in newbornswith prematurity with an ML approach. To our knowledge, this study is the first to integrate ocular examination into evaluating the cognitive outcomes in preterm infants. It is also the first study to investigate the relative contributions of the key ocular features when all are considered in the prediction of cognitive outcomes with an ML approach. Methods This study was approved by the Institutional Review Board of Chang Gung Memorial Hospital in Linkou, Taiwan (IRB number: 202401555B0) and complied with the principles in the Declaration of Helsinki. Eligibility Criteria and Study Population This study retrospectively reviewed the electronic medical charts and medical images of 163 children (326 eyes) under 18 years old born between January 2011 and August 2024 at Chang Gung Memorial Hospital, Linkou branch, Taiwan. Basic characteristics at birth, including gender, GA, birth weight (BW), Apgar score at 5 min, and the Apgar score difference between 1 and 5 minutes, were included. Ocular findings included the stage, zone, and disease of ROP, as well as optical coherence tomography (OCT) features such as retinal thickness and differences between bilateral eyes. Since GA and BW are highly correlated, we focus on BW instead of GA as the attribute in our ML models. Patients with IQ scores between 85 and 115, those who did not undergo IQ testing, and those with incomplete medical records were excluded from this study. The mean age in our study cohort was 7.6 ± 2.4 years old, with 96 male babies and 67 female babies. Eighty-six children (172 eyes) fall into the IQ < 85 cohort, while 77 children (153 eyes) fall into the IQ ≥ 115 cohort. There were 48 term infants, 55 preterm infants without ROP, and 60 infants with ROP. Table 1 shows the detailed demographics of our study cohort. IQ Measurement The IQ in our study cohort was measured by the Chinese version of the Wechsler Preschool and Primary Scale of Intelligence, fourth edition (WPPSI-IV) 10 , or the Wechsler Intelligence Scale for Children, fourth edition or fifth edition (WISC-IV, V) 8, 11 according to the children’s age at the time of data acquisition. The Wechsler Preschool and Primary Scale of Intelligence (WPPSI) 10 and the Wechsler Intelligence Scale for Children (WISC) 8, 11 were designed to evaluate the cognitive ability of children of different ages. The WPPSI was the major instrument when assessing the cognitive ability of children ages 2 years, 6 months through 7 years, while the WISC was widely used when evaluating children between the ages of 6-16. In our study cohort, the IQ scores were collected retrospectively between the ages of 3 and 13. The IQ score remained relatively stable from early childhood to adulthood, leaving children with poor cognitive outcomes disadvantaged unless further interventions are implemented. 12 The WPPSI-IV was used in children between 3 and 6 years old, while the WISC-IV and WISC-V were used to evaluate the children’s IQ if the subject was between 6 and 13 years old. Pediatric psychiatrists performed all the cognitive tests without awareness of the patient’s clinical and ocular conditions. Machine Learning The machine learning model is designed to classify a child as below-normal IQ range or above-normal IQ range. We utilize a Random Forest (RF) model for this task, incorporating data from a single eye, including stage, zone, plus disease, the treatment times of ROP, and the essential characteristics, such as gender, birth weight (BW), Apgar score at 5 min, and the Apgar score difference between 1 and 5 minutes. The RF algorithm constructs multiple decision trees by training on different subsets of the training data and determines the final classification outcome through a majority voting mechanism, aggregating the predictions from all decision trees to reduce the variance of the model. Unlike basic models such as logistic regression, which assume linear relationships between variables and are sensitive to outliers, RF is well-suited for handling high-dimensional, complex datasets and diverse data types. We also applied the Synthetic Minority Oversampling Technique (SMOTE) to balance the number of samples between the two classes before training, as we wanted to prevent the model from favoring one class during learning. SMOTE generates new samples by examining the k-nearest neighbors of a sample in the minority class, continuing this process until the data between the two classes becomes balanced. Leave-One-Out Cross-Validation (LOOCV) was employed as the validation strategy to ensure the model's generalizability. In the original LOOCV, each data point is sequentially held out as a validation set while the model is trained on the remaining data. However, to align with our experiment's requirements, we slightly modified its definition. In this study, specific features, such as gender and eyes of the same patient, share common features, i.e. differences in retinal thickness between both eyes. Therefore, we modified validation to "Leave-One-Person-Out Cross-Validation". Specifically, we considered the two eyes of the same individual as a validation group, meaning that each validation process included two data points, one for each eye of the same person. In our dataset, which consists of n samples (eye data), we used N-2 samples as the training set and two samples from the same individual as the validation set. This process was repeated N/2 times, and the average of all results was computed to obtain the final evaluation score. This iterative process provides an unbiased estimate of model performance, as every data point is used for training and testing. Figure 2 shows the flowchart of our training process. Explainability in ML refers to the ability to better understand and interpret the output of a complicated model, which is crucial for ensuring transparency when applying the results in clinical settings. Shapley Additive exPlanations (SHAP) is a game-theoretic method that explains the output of ML models by assigning each feature a "contribution" value. This value quantifies the magnitude and direction of the feature's effect on the model's predictions, thereby clarifying the relationship between individual features and the target outcome. Additionally, dependence plots derived from SHAP provide a visual representation of how individual predictors affect the model’s output. By enabling such insights, SHAP enhances the interpretability of ML models and ensures their outputs are more accessible and transparent. In this study, SHAP were used to identify the clinical features most indicative of the cognitive outcomes in preterm infants in early childhood. Results Performance Analysis Table 2. provides the performance summary of our model. The overall AUC was highest in the group combining the clinical features and retinal thickness on OCT (83.86%), and the accuracy was highest in the group combining the clinical, retinal thickness, and retinal thickness difference in bilateral eyes (83.44 %). The ML model performed better in precision, recall, and F1-score in the group when combining OCT features than clinical features alone. In the group combining the clinical features and retinal thickness, our model achieved a precision of 0.88, a recall of 0.78, and an F1-score of 0.83 in predicting infants with an IQ score less than 85, and achieved a precision of 0.79, a recall of 0.88, and an F1-score of 0.83 in predicting infants with the IQ score more than 115. The results were similar in the group combining the clinical features, retinal thickness, and retinal thickness difference in bilateral eyes. In this group, our model achieved a precision of 0.88, a recall of 0.79, and an F1-score of 0.83 in predicting infants with an IQ score less than 85, and achieved a precision of 0.79, a recall of 0.88, and an F1-score of 0.83 in predicting infants with the IQ score more than 115. Feature Analysis The SHAP analysis highlights features from the model trained on the complete feature set to enhance its explainability. The features are ordered by their importance score. Figure 3a ranked clinical features (BW, stage, zone, Apgar score at 5 minutes, Apgar score difference, sex, plus disease, and ROP treatment times) based on their importance. We demonstrated that the lower birth weight, higher stage of ROP, and lower zone development were associated with lower IQ scores. These features consistently rank among the top three after adding OCT features (Figure 3b and 3c). As for the OCT features, the thickness of the foveawas found to be the most important feature besides the BW, the stage of ROP, and the zone in the SHAP analysis.The dependence plots demonstrate the detailed correlation between the clinical features and the IQ scores (Figure 4) . Discussion To the best of our knowledge, this is the first study to integrate ocular examination into evaluating cognitive outcomes in preterm infants. By assessing the clinical information and ocular examinations, we identify the important features associated with cognitive development in newbornswith prematurity in early childhood with an ML approach with an overall accuracy of 83.44%. Through SHAP analysis, we highlight the BW, the stage of ROP, and zone development as the primary parameters associated with cognitive development. Early intervention is crucial in optimizing cognitive outcomes for children at high risk of cognitive deficits. The Abecedarian Early Intervention Project 13 conducted a randomized controlled trial on 111 infants at risk of poor cognitive outcomes. It proved that infants receiving early high-quality educational intervention may have a higher IQ score than the control group. Early intervention programs may positively influence the cognitive and motor outcomes of preterm infants, which may also reduce the burden on the families. 14, 15 Our model identifies the crucial features related to poor mental development in preterm infants in early childhood, hoping to detect children with a high risk of poor neurodevelopmental outcomes for early intervention. Low BW is a well-established factor associated with impaired cognitive development in children. 16-18 According to the World Health Organization (WHO), around 20 million infants are born with low birth weight each year. 19 Low birth weight often reflects underlying complexities, including prematurity, malnutrition, underdeveloped brain structure, and a higher prevalence of comorbidities such as intraventricular hemorrhages, all of which can adversely impact cognitive outcomes. 16 Previous studies showed that the impact of low birth weight on adverse cognitive outcomes may persist into adolescence and early adulthood. 20 Through SHAP analysis, our study identified BW as one of the most influential factors contributing to adverse cognitive outcomes in preterm infants. Timely evaluation and intervention should be done to improve the cognitive outcome in children with low BW. Preterm infants with ROP are at a significantly increased risk of cognitive delays. Previous studies showed children with ROP may be more likely to experience motor, mental, and language delays compared with those without ROP. 21-23 Our study included the key elements of ROP, the stage, the zone, the presence of plus disease, and the treatment times of ROP. We found that the stage and zone of ROP are two of the leading factors, among others, that are related to poor cognitive development. Our findings emphasize the importance of ROP screening and monitoring to decrease the risk of long-term cognitive impairments. Retinal thickness was a potential biomarker associated with cognitive development in children in our study. Embryologically speaking, the ocular structure is an extension of the central nervous system. Previous studies have demonstrated a relationship between a thinner mean retinal nerve fiber layer (RNFL) and lower IQ scores, suggesting that the retina is a biological correlate of cognitive functioning and might be a potential indicator of overall brain function. 24, 25 Rothman et al. 26 identified cystoid macular edema as a biomarker of poor neurodevelopmental outcomes in preterm infants, further supporting the link between retinal pathology and developmental cognitive deficits. Our study showed that a thicker fovea was associated with worse cognitive performance. This finding aligns with prior observations that children with ROP may have a thicker fovea compared to those without ROP, and ROP itself is a risk factor for adverse cognitive outcomes. 21-23 Interestingly, we also observed that the asymmetrical retinal thickness may slightly improve the accuracy of our model. The difference in retinal thickness in bilateral eyes may indicate different severities of ROP in individual eyes. Although the exact mechanism is unclear, this finding may highlight the importance of detailed retinal assessments in evaluating children with high risk of cognitive deficits. Our ML model was a pioneer model incorporating OCT information in evaluating the cognitive development in preterm infants. OCT is a non-invasive, high-resolution imaging modality crucial in assessing retinal structures. The previous study 27 proved that the RNFL thickness was associated with cognitive functional decline in adult patients. Barrett-Young et al . 24 reported that the thinner RNFL and ganglion cell layer (GCL) were associated with lower IQ in childhood and the middle ages. In our ML model, we incorporated the OCT information, including the foveal and parafoveal thickness, and the fovea and parafoveal difference in the bilateral eyes, with clinical features to provide a more comprehensive and objective assessment. By adding the OCT information, we demonstrated a significant improvement in both accuracy and the AUC. While the presented ML model identifies the important features associated with cognitive development in newbornswith prematurity in early childhood, several limitations should be acknowledged. First, our training relied solely on basic clinical characteristics and eye examination data. Genetic information and external factors, such as parental educational status and environmental influences, were excluded from our model. While these factors provide additional insights, their exclusion could also be considered a strength. Since such information is not routinely recorded in clinical settings, our ML model demonstrates the feasibility of using readily available data. Despite this limitation, the model achieved a high overall accuracy exceeding 83%, underscoring its robustness and applicability in daily practice. Future studies incorporating genetic and environmental information could be conducted to enhance the model's comprehensiveness. In addition, since our study cohort excluded children with IQ scores between 85 and 115, this model is not suitable as a screening tool for early intervention. However, this study highlighted key ocular features associated with cognitive outcomes, providing crucial insights into ocular factors possibly linked to mental development. Finally, the lack of external validation on independent cohorts may limit the model's utility in real-world clinical settings. External validation with diverse populations across different races might improve the generalizability of our model and evaluate the model's performance in real-world settings. Conclusion This study is the first to integrate ocular examination into evaluating cognitive outcomes in preterm infants. By assessing the ML model, we identified key clinical features associated with cognitive development in preterm newborns in early childhood. To enhance the model's interpretability, we employed SHAP analysis, highlighting the significance of the BW, the stage of ROP, and zone development in influencing cognitive outcomes. This study demonstrates the potential of using ML models to predict IQ in children and identify those at high risk for poor cognitive outcomes, paving the way for timely interventions to enhance long-term neurodevelopmental outcomes. Declarations Grant Support: NSTC 113-2314-B-182A-040, CMRPG3P0051~2, CMRPG3P0191, PSC-CUNY, Award # 65406-00 53, NSTC 113-2221-E-194-034. Disclosure: The sponsors had no role in the design or conduct of this research. Acknowledgments Ethics approval and consent to participate The study adhered to the Declaration of Helsinki and was approved by the Institutional Review Board of Chang Gung Memorial Hospital (IRB number: 202401555B0), which granted a waiver of consent because patient anonymity was maintained by the data source. Availability of data and materials The data analyzed during this study are available on request from the corresponding authors, Wei-Chi, Wu, and Chia-Ling Tsai. Authors' contributions All authors have participated directly in the planning and execution of the work. MMH: interpreted data, drafting and writing the article; MLM, HYS: acquisition and analysis of data; WYL, RFC, CLT, WCW: design of the study, acquisition of data, final approval. All authors read and approved the final manuscript. References Olusanya BO, Smythe T, Ogbo FA, Nair MKC, Scher M, Davis AC (2023) Global prevalence of developmental disabilities in children and adolescents: A systematic umbrella review. Front Public Health 11:1122009 McKenzie K, Milton M, Smith G, Ouellette-Kuntz H (2016) Systematic Review of the Prevalence and Incidence of Intellectual Disabilities: Current Trends and Issues. Curr Dev Disorders Rep 3:104–115 Drost FJ, Keunen K, Moeskops P et al (2018) Severe retinopathy of prematurity is associated with reduced cerebellar and brainstem volumes at term and neurodevelopmental deficits at 2 years. Pediatr Res 83:818–824 Godugu S, Vudugula SA, Neupane B et al (2023) Association between severe retinopathy of prematurity (ROP) and poor motor neurodevelopmental outcome. Indian J Ophthalmol 71:2944–2946 Morken TS, Dammann O, Skranes J, Austeng D (2019) Retinopathy of prematurity, visual and neurodevelopmental outcome, and imaging of the central nervous system. Semin Perinatol 43:381–389 Linsell L, Malouf R, Morris J, Kurinczuk JJ, Marlow N (2015) Prognostic Factors for Poor Cognitive Development in Children Born Very Preterm or With Very Low Birth Weight: A Systematic Review. JAMA Pediatr 169:1162–1172 Anderson PJ (2023) Predicting neurodevelopmental outcome in children born very preterm – does neonatal MRI have a role? Pediatr Res 94:868–869 Wechsler D (2014) Wechsler Intelligence Scale for Children–Fifth Edition. WISC-V) APA PsycTests. Faramarzi R, Darabi A, Emadzadeh M, Maamouri G, Rezvani R (2023) Predicting neurodevelopmental outcomes in preterm infants: A comprehensive evaluation of neonatal and maternal risk factors. Early Hum Dev 184:105834 Wechsler D (1967) WPPSI: Wechsler preschool and primary scale of intelligence. Psychological Corporation, [1967] ©1967, New York, N.Y Wechsler D (2003) Wechsler Intelligence Scale for Children, Fourth Edition (WISC-IV). APA PsycTests Schneider W, Niklas F, Schmiedeler S (2014) Intellectual development from early childhood to early adulthood: The impact of early IQ differences on stability and change over time. Learn Individual Differences 32:156–162 Muennig P, Robertson D, Johnson G, Campbell F, Pungello EP, Neidell M (2011) The effect of an early education program on adult health: the Carolina Abecedarian Project randomized controlled trial. Am J Public Health 101:512–516 Spittle A, Orton J, Anderson PJ, Boyd R, Doyle LW (2015) Early developmental intervention programmes provided post hospital discharge to prevent motor and cognitive impairment in preterm infants. Cochrane Database Syst Rev 2015:Cd005495 Anderson PJ, Treyvaud K, Spittle AJ. Early developmental interventions for infants born very preterm – what works? Seminars in Fetal and Neonatal Medicine 2020;25 Gu H, Wang L, Liu L et al (2017) A gradient relationship between low birth weight and IQ: A meta-analysis. Sci Rep 7:18035 Breslau N, Dickens WT, Flynn JR, Peterson EL, Lucia VC (2006) Low birthweight and social disadvantage: Tracking their relationship with children's IQ during the period of school attendance. Intelligence 34:351–362 Eves R, Mendonça M, Baumann N et al (2021) Association of Very Preterm Birth or Very Low Birth Weight With Intelligence in Adulthood: An Individual Participant Data Meta-analysis. JAMA Pediatr 175:e211058 Organization WH (2019) UNICEF-WHO low birthweight estimates: levels and trends 2000–2015. World Health Organization Kormos CE, Wilkinson AJ, Davey CJ, Cunningham AJ (2014) Low birth weight and intelligence in adolescence and early adulthood: a meta-analysis. J Public Health (Oxf) 36:213–224 Ahn JH, Lee KM, Kim MJ et al (2022) Neurodevelopmental outcomes in very low birthweight infants with retinopathy of prematurity in a nationwide cohort study. Sci Rep 12:5053 Huang CC, Chu CH, Lin YK, Lin YC, Huang HM, Chang YC (2022) Retinopathy of Prematurity Is a Biomarker for Pathological Processes in the Immature Brain. Neonatology 119:727–734 Beligere N, Perumalswamy V, Tandon M et al (2015) Retinopathy of prematurity and neurodevelopmental disabilities in premature infants. Semin Fetal Neonatal Med 20:346–353 Barrett-Young A, Ambler A, Cheyne K et al (2022) Associations Between Retinal Nerve Fiber Layer and Ganglion Cell Layer in Middle Age and Cognition From Childhood to Adulthood. JAMA Ophthalmol 140:262–268 Rothman AL, Sevilla MB, Mangalesh S et al (2015) Thinner Retinal Nerve Fiber Layer in Very Preterm Versus Term Infants and Relationship to Brain Anatomy and Neurodevelopment. Am J Ophthalmol 160:1296–1308e1292 Rothman AL, Tran-Viet D, Gustafson KE et al (2015) Poorer neurodevelopmental outcomes associated with cystoid macular edema identified in preterm infants in the intensive care nursery. Ophthalmology 122:610–619 Ko F, Muthy ZA, Gallacher J et al (2018) Association of Retinal Nerve Fiber Layer Thinning With Current and Future Cognitive Decline: A Study Using Optical Coherence Tomography. JAMA Neurol 75:1198–1205 Tables Table 1. Demographics in Our Study Cohort Intelligence Quotient (IQ) IQ < 85 IQ ≥ 115 p -value Patients, (eye) 86 (172) 77 (154) Age at IQ evaluation, yr (SD) 7.2 (2.5) 8.1 (2.2) Gender, Male, (%) 55 (64.0) 41 (53.2) 0.167 GA, wk, (SD) 29.8 (4.9) 35.1 (4.5) <0.001 BW, g, (SD) 1359.7 (852.6) 2365.3 (918.8) <0.001 Apgar score at 1 min 5.7 (2.3) 7.9 (1.6) <0.001 Apgar score at 5 min 7.5 (2.1) 9.3 (1.0) <0.001 Term infant, n, (%) 8 (9.3) 40 (51.9) <0.001 Preterm infant without ROP, n, (%) 24 (27.9) 31 (40.2) 0.001 ROP, n, (%) 54 (62.8) 6 (7.7) <0.001 ROP stage, eye, (%) <0.001 1 20 (11.6) 0 (0) 2 20 (11.6) 5 (3.2) 3 64 (37.2) 7 (4.5) 4 1 (0.6) 0 (0) 5 3 (1.7) 0 (0) Plus disease, eye, (%) 63 (36.6) 6 (3.9) <0.001 Zone, eye, (%) <0.001 1 19 (11) 0 (0) 2 79 (45.9) 10 (6.5) 3 74 (43.0) 144 (93.5) ROP with treatment, eye Intravitreal anti-VEGF Injection 44 (25.6) 8 (5.2) <0.001 Laser Photocoagulation 3 (1.7) 0 (0) 0.146 Combine Intravitreal Injection and Laser Photocoagulation 22 (12.8) 0 (0) <0.001 Vitrectomy 4 (2.3) 0 (0) 0.076 Average treatment times, n 1.3 1 BW, birth weight; GA, gestational age; ROP, retinopathy of prematurity; SD: standard deviation; VEGF, vascular endothelial growth factor Table 2. The Performance Summary of Our Model Features AUC (%) Accuracy (%) IQ score Precision Recall F1-score Clinical features 79.55 78.83 < 85 0.86 0.72 0.78 ≧115 0.73 0.87 0.80 Clinical features, retinal thickness 83.86 83.13 < 85 0.88 0.78 0.83 ≧115 0.79 0.88 0.83 Clinical features, retinal thickness, retinal thickness difference in bilateral eyes 83.01 83.44 < 85 0.88 0.79 0.83 ≧115 0.79 0.88 0.83 AUC, area under curve, IQ, intelligence quotient Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7172977","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":488243007,"identity":"77387b28-81f9-4d41-a5b9-ed1f45e67eee","order_by":0,"name":"Margaret Ming-Chih Ho","email":"","orcid":"","institution":"Jen-Ai Hospital Dali Branch","correspondingAuthor":false,"prefix":"","firstName":"Margaret","middleName":"Ming-Chih","lastName":"Ho","suffix":""},{"id":488243008,"identity":"f18be603-58c9-49ff-9994-2fee009592dc","order_by":1,"name":"Man-Lin, Mai","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Mai","middleName":"","lastName":"Man-Lin","suffix":""},{"id":488243009,"identity":"3b0e9ca0-ef07-4a03-9ebf-d98c4baa4502","order_by":2,"name":"Wei-Yang Lin","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Wei-Yang","middleName":"","lastName":"Lin","suffix":""},{"id":488243010,"identity":"368e3cfb-98a8-46e8-b5f0-2a0a2c404f06","order_by":3,"name":"Yu-Shu Huang","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Yu-Shu","middleName":"","lastName":"Huang","suffix":""},{"id":488243011,"identity":"59802055-4acb-4e07-a449-b1c55bba5a62","order_by":4,"name":"Ruey-Feng Chang","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Ruey-Feng","middleName":"","lastName":"Chang","suffix":""},{"id":488243012,"identity":"533a30a6-722d-4ca8-803d-ed175f04739d","order_by":5,"name":"Chia-Ling Tsai","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABCElEQVRIiWNgGAWjYDACCcYGhgcGUE5CBUICKI5HSwJICxtIyxmitIBUMkC1MLYRoYV/dnPjh4SCO3YN8u0PPzycZy3PwH7G7DEPg43shgM4LLlzsFkiweBZcgMbj7FE4rZ0wwaeHHNjHoY0Y1xaDCQSG4BaDiczsPEwALUcTmCQ4DGT5mE4nIhHS/MPiBb2xz8S58C1/MenpQ1kix3Q+2ZAG+FaDuDUInEjsc0CqCWBjS3HzCLhWLphG09ameQcg2TjmTi08M9If3zjw5/D9vzMxx/f/FFjLc/PfnibxJsKO9k+HFpgIBEaI8zg+GHiMcCrGgzsGWBaQIDxB2Edo2AUjIJRMHIAANzvVm7i56NtAAAAAElFTkSuQmCC","orcid":"","institution":"","correspondingAuthor":true,"prefix":"","firstName":"Chia-Ling","middleName":"","lastName":"Tsai","suffix":""},{"id":488243013,"identity":"49e64e6b-fba1-46ed-b748-20a11af0da91","order_by":6,"name":"Wei-Chi Wu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA1UlEQVRIiWNgGAWjYFACxgZmBgMbOX72BiDHwIJoLWnGkj0HQFokiLOHmYHhUKLBjAQQmwgtBsebGz8XFBxIMJB8fnXDjwIJBv727gT8Ws4cbJaeYXAnz1w6p+xmD9BhEmfObsCrxexGYhszj8GzYsvZOWk3eIBaDCRyCWi5/xCk5XDihptn0m7+IUrLDUaolhvsx24TZYv9mcRmaR5wIOew3ZYxkOAh6BfJ9uMPP/P8AUXl8Wc334AY7b34tSABHgMwSaxyEGB/QIrqUTAKRsEoGEEAAO1YSaDj5xHMAAAAAElFTkSuQmCC","orcid":"","institution":"Chang Gung Memorial Hospital, Linkou branch","correspondingAuthor":true,"prefix":"","firstName":"Wei-Chi","middleName":"","lastName":"Wu","suffix":""}],"badges":[],"createdAt":"2025-07-21 03:55:46","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-7172977/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7172977/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":87665044,"identity":"79dc9600-7ab4-440c-a13a-ad42ee4a1f35","added_by":"auto","created_at":"2025-07-27 11:02:15","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":428715,"visible":true,"origin":"","legend":"\u003cp\u003e(a) The t-distributed Stochastic Neighbor Embedding (t-SNE) of children reveals that the group with IQ ≥ 85 spans the entire feature space, completely overlapping with the IQ \u0026lt; 85 group. \u0026nbsp;(b) When focusing on those with below-average IQ (IQ \u0026lt; 85) and those with above-average IQ (IQ ≥ 115), the t-SNE demonstrated clearer clustering and better separation, indicating stronger distinguishing features between these two groups.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7172977/v1/8fe4f3631f3bdbf0b61eb125.png"},{"id":87663741,"identity":"b1586950-d2f7-4181-814b-9e3707b19d98","added_by":"auto","created_at":"2025-07-27 10:54:15","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":55847,"visible":true,"origin":"","legend":"\u003cp\u003eThe flowchart of our training process.\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7172977/v1/26812495829da6d53d18da95.jpg"},{"id":87663745,"identity":"019bb9e1-970e-4b37-b499-1c163324de5a","added_by":"auto","created_at":"2025-07-27 10:54:15","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":402688,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003e(\u003c/strong\u003ea) The SHAP analysis demonstrated that the lower birth weight, higher stage of ROP, and lower zone development were associated with lower IQ scores. (b) When including OCT features, the thickness of the fovea was found to be the most important feature besides the BW, the stage of ROP, and the zone development. (c) The leading factors remained the same after adding the retinal thickness difference of bilateral eyes.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-7172977/v1/c17cd601671b890cfc0e3fbf.png"},{"id":87663749,"identity":"9e6edbac-9d73-45e7-918b-6595f0587141","added_by":"auto","created_at":"2025-07-27 10:54:15","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":968645,"visible":true,"origin":"","legend":"\u003cp\u003eThe dependence plots demonstrate that the low BW (a), stage 1-5 ROP (b), zone 1 and 2 ROP (c), and the higher foveal thickness on OCT (d)positively contribute to the model's predictions, increasing the likelihood that it classifies the child as having a lower IQ score.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-7172977/v1/c720c80532c5d6cc83aa0d5f.png"},{"id":87666038,"identity":"a268c583-3c6d-4547-8062-b7ae18144cb0","added_by":"auto","created_at":"2025-07-27 11:10:20","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2248089,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7172977/v1/6b618b30-ed57-42fd-8352-9521853b00fc.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eFrom Eye Examination to Early Cognitive Evaluation for Preterm Newborns: An Explainable Machine Learning Approach\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"Introduction","content":"\u003cp\u003eDevelopmental intellectual disability (DID) is the second most prevalent developmental disability among children under 20 years old, following hearing loss.\u003csup\u003e1\u003c/sup\u003e It has become a global concern, affecting 0.2–1.5% of children worldwide.\u003csup\u003e2\u003c/sup\u003e DID is a complex and multifactorial process influenced by genetic, environmental, and perinatal factors. Without early intervention, DID might become a lifelong problem and affect the children’s quality of life. Given the prevalence of DID and its significant impact on the quality of life of children and their families, it is crucial to identify children at high risk of DID for early intervention.\u003c/p\u003e\n\u003cp\u003eRetinopathy of prematurity (ROP) is a retinal vascular disease that may lead to visual impairment and is associated with central nervous system (CNS) comorbidities, such as intraventricular hemorrhage (IVH) and restricted brain development, resulting in cognitive impairment\u003csup\u003e3\u003c/sup\u003e. Previous studies have shown that the presence and severity of ROP may be linked to poor neurodevelopmental outcomes.\u003csup\u003e4, 5\u003c/sup\u003e Other examinations, such as brain magnetic resonance imaging (MRI), were also used to evaluate the risk factors for poor cognitive outcomes in children in previous studies.\u003csup\u003e6, 7\u003c/sup\u003e However, limitations such as high cost, limited accessibility, and the need for sedation, make MRI less practical in routine neonatal care.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe intelligence quotient (IQ) score is standardized to have a mean of 100 and a standard deviation of 15.\u003csup\u003e8\u003c/sup\u003e An IQ score of 85-115 is considered average, while scores below 85 are considered below-average and scores more than 115 indicate above-average. Ideally, a statistical threshold value should be placed at 85 to separate children needing early intervention from those of normal development. However, such a statistically determined threshold does not necessarily map to a clear cut in the feature space, as shown in \u003cstrong\u003eFigure 1(a)\u003c/strong\u003e: projecting the feature space to two dimensions using t-SNE, the group with IQ ≥ 85 spans the entire feature space, completely overlapping with the IQ \u0026lt; 85 group. This can pose challenges in identifying prognostic factors for children with poor cognitive development. To address this, we excluded children with a normal IQ (IQ between 85 and 115) and focused on those with below-average IQ (IQ \u0026lt; 85) and those with above-average IQ (IQ ≥ 115). Higher separability between the two groups is confirmed by\u003cstrong\u003e\u0026nbsp;Figure 1(b)\u003c/strong\u003e. By comparing the clinical and ocular features of children with below-average IQ to those with above-average IQ, we can delineate ocular differences between the two groups and identify the clinical characteristics and ocular features that contribute to poor cognitive outcomes for early intervention.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eRecent advances in artificial intelligence (AI) and machine learning (ML) approaches profoundly impact various healthcare domains. Unlike conventional linear regression or correlation coefficient methods that typically analyze one feature at a time, the ML approach can efficiently analyze multiple risk factors and their interplay in one setting. Furthermore, the ML approach might offer novel insights into detecting factors that may not have been considered relevant if chosen manually for analysis with traditional methods.\u003c/p\u003e\n\u003cp\u003ePrognostic factors reported in previous studies \u003csup\u003e6, 9\u003c/sup\u003e associated with poor cognitive outcomes in preterm infants included lower gestational age (GA), male sex, nonwhite race, lower level of parental education, poor intrauterine growth, diabetes, lower birth weight, and hospitalization. Nevertheless, none of these studies have discussed the degree of contribution or the interplay of these factors to the cognitive outcome in early childhood with ML techniques. By leveraging basic clinical features and ocular examinations, we aim to identify the key features associated with poor cognitive development in early childhood in newbornswith prematurity with an ML approach. To our knowledge, this study is the first to integrate ocular examination into evaluating the cognitive outcomes in preterm infants. It is also the first study to investigate the relative contributions of the key ocular features when all are considered in the prediction of cognitive outcomes with an ML approach.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eThis study was approved by the Institutional Review Board of Chang Gung Memorial Hospital in Linkou, Taiwan (IRB number: 202401555B0) and complied with the principles in the Declaration of Helsinki. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eEligibility Criteria and Study Population\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThis study retrospectively reviewed the electronic medical charts and medical images of\u0026nbsp;163 children (326 eyes) under 18 years old born between January 2011 and August 2024 at Chang Gung Memorial Hospital, Linkou branch, Taiwan. Basic characteristics at birth, including gender, GA, birth weight (BW), Apgar score at 5 min, and the Apgar score difference between 1 and 5 minutes, were included. Ocular findings included the stage, zone, and disease of ROP, as well as optical coherence tomography (OCT) features such as retinal thickness and differences between bilateral eyes. Since GA and BW are highly correlated, we focus on BW instead of GA as the attribute in our ML models. Patients with IQ scores between 85 and 115, those who did not undergo IQ testing, and those with incomplete medical records were excluded from this study.\u003c/p\u003e\n\u003cp\u003eThe mean age in our study cohort was 7.6 \u0026plusmn; 2.4 years old, with 96 male babies and 67 female babies. Eighty-six children (172 eyes) fall into the IQ \u0026lt; 85 cohort, while 77 children (153 eyes) fall into the IQ \u0026ge; 115 cohort. There were 48 term infants, 55 preterm infants without ROP, and 60 infants with ROP. \u003cstrong\u003eTable 1\u003c/strong\u003e shows the detailed demographics of our study cohort.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eIQ Measurement\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe IQ in our study cohort was measured by the Chinese version of the Wechsler Preschool and Primary Scale of Intelligence, fourth edition (WPPSI-IV)\u003csup\u003e10\u003c/sup\u003e, or the Wechsler Intelligence Scale for Children, fourth edition or fifth edition (WISC-IV, V)\u003csup\u003e8, 11\u003c/sup\u003e according to the children\u0026rsquo;s age at the time of data acquisition. The Wechsler Preschool and Primary Scale of Intelligence (WPPSI)\u003csup\u003e10\u003c/sup\u003e and the Wechsler Intelligence Scale for Children (WISC)\u003csup\u003e8, 11\u003c/sup\u003e were designed to evaluate the cognitive ability of children of different ages. The WPPSI was the major instrument when assessing the cognitive ability of children ages 2 years, 6 months through 7 years, while the WISC was widely used when evaluating children between the ages of 6-16.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn our study cohort, the IQ scores were collected retrospectively between the ages of 3 and 13. The IQ score remained relatively stable from early childhood to adulthood, leaving children with poor cognitive outcomes disadvantaged unless further interventions are implemented.\u003csup\u003e12\u003c/sup\u003e The WPPSI-IV was used in children between 3 and 6 years old, while the WISC-IV and WISC-V were used to evaluate the children\u0026rsquo;s IQ if the subject was between 6 and 13 years old. Pediatric psychiatrists performed all the cognitive tests without awareness of the patient\u0026rsquo;s clinical and ocular conditions.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eMachine Learning\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe machine learning model is designed to classify a child as below-normal IQ range or above-normal IQ range. We utilize a Random Forest (RF) model for this task, incorporating data from a single eye, including stage, zone, plus disease, the treatment times of ROP, and the essential characteristics, such as gender, birth weight (BW), Apgar score at 5 min, and the Apgar score difference between 1 and 5 minutes. The RF algorithm constructs multiple decision trees by training on different subsets of the training data and determines the final classification outcome through a majority voting mechanism, aggregating the predictions from all decision trees to reduce the variance of the model. Unlike basic models such as logistic regression, which assume linear relationships between variables and are sensitive to outliers, RF is well-suited for handling high-dimensional, complex datasets and diverse data types. We also applied the Synthetic Minority Oversampling Technique (SMOTE) to balance the number of samples between the two classes before training, as we wanted to prevent the model from favoring one class during learning. SMOTE generates new samples by examining the k-nearest neighbors of a sample in the minority class, continuing this process until the data between the two classes becomes balanced.\u003c/p\u003e\n\u003cp\u003eLeave-One-Out Cross-Validation (LOOCV) was employed as the validation strategy to ensure the model\u0026apos;s generalizability. In the original LOOCV, each data point is sequentially held out as a validation set while the model is trained on the remaining data. However, to align with our experiment\u0026apos;s requirements, we slightly modified its definition. In this study, specific features, such as gender and eyes of the same patient, share common features, i.e. differences in retinal thickness between both eyes. Therefore, we modified validation to \u0026quot;Leave-One-Person-Out Cross-Validation\u0026quot;. Specifically, we considered the two eyes of the same individual as a validation group, meaning that each validation process included two data points, one for each eye of the same person. In our dataset, which consists of n samples (eye data), we used N-2 samples as the training set and two samples from the same individual as the validation set. This process was repeated N/2 times, and the average of all results was computed to obtain the final evaluation score. This iterative process provides an unbiased estimate of model performance, as every data point is used for training and testing. \u003cstrong\u003eFigure 2\u0026nbsp;\u003c/strong\u003eshows the flowchart of our training process.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eExplainability in ML refers to the ability to better understand and interpret the output of a complicated model, which is crucial for ensuring transparency when applying the results in clinical settings. Shapley Additive exPlanations (SHAP) is a game-theoretic method that explains the output of ML models by assigning each feature a \u0026quot;contribution\u0026quot; value. This value quantifies the magnitude and direction of the feature\u0026apos;s effect on the model\u0026apos;s predictions, thereby clarifying the relationship between individual features and the target outcome. Additionally, dependence plots derived from SHAP provide a visual representation of how individual predictors affect the model\u0026rsquo;s output. By enabling such insights, SHAP enhances the interpretability of ML models and ensures their outputs are more accessible and transparent. In this study, SHAP were used to identify the clinical features most indicative of the cognitive outcomes in preterm infants in early childhood.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cem\u003ePerformance Analysis\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2.\u003c/strong\u003e provides the performance summary of our model. The overall AUC was highest in the group combining the clinical features and retinal thickness on OCT (83.86%), and the accuracy was highest in the group combining the clinical, retinal thickness, and retinal thickness difference in bilateral eyes (83.44 %). The ML model performed better in precision, recall, and F1-score in the group when combining OCT features than clinical features alone. In the group combining the clinical features and retinal thickness, our model achieved a precision of 0.88, a recall of 0.78, and an F1-score of 0.83 in predicting infants with an IQ score less than 85, and achieved a precision of 0.79, a recall of 0.88, and an F1-score of 0.83 in predicting infants with the IQ score more than 115. The results were similar in the group combining the clinical features, retinal thickness, and retinal thickness difference in bilateral eyes. In this group, our model achieved a precision of 0.88, a recall of 0.79, and an F1-score of 0.83 in predicting infants with an IQ score less than 85, and achieved a precision of 0.79, a recall of 0.88, and an F1-score of 0.83 in predicting infants with the IQ score more than 115.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eFeature Analysis\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe SHAP analysis highlights features from the model trained on the complete feature set to enhance its explainability. The features are ordered by their importance score. \u003cstrong\u003eFigure 3a\u0026nbsp;\u003c/strong\u003eranked clinical features (BW, stage, zone, Apgar score at 5 minutes, Apgar score difference, sex, plus disease, and ROP treatment times) based on their importance. We demonstrated that the lower birth weight, higher stage of ROP, and lower zone development were associated with lower IQ scores. These features consistently rank among the top three after adding OCT features \u003cstrong\u003e(Figure 3b and 3c).\u0026nbsp;\u003c/strong\u003eAs for the OCT features, the thickness of the foveawas found to be the most important feature besides the BW, the stage of ROP, and the zone in the SHAP analysis.The dependence plots demonstrate the detailed correlation between the clinical features and the IQ scores \u003cstrong\u003e(Figure 4)\u003c/strong\u003e.\u0026nbsp;\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eTo the best of our knowledge, this is the first study to integrate ocular examination into evaluating cognitive outcomes in preterm infants. By assessing the clinical information and ocular examinations, we identify the important features associated with cognitive development in newbornswith prematurity in early childhood with an ML approach with an overall\u0026nbsp;accuracy of 83.44%. Through SHAP analysis, we highlight the BW, the stage of ROP, and zone development as the primary parameters associated with cognitive development.\u003c/p\u003e\n\u003cp\u003eEarly intervention is crucial in optimizing cognitive outcomes for children at high risk of cognitive deficits. The Abecedarian Early Intervention Project\u003csup\u003e13\u003c/sup\u003e conducted a randomized controlled trial on 111 infants at risk of poor cognitive outcomes. It proved that infants receiving early high-quality educational intervention may have a higher IQ score than the control group. Early intervention programs may positively influence the cognitive and motor outcomes of preterm infants, which may also reduce the burden on the families.\u003csup\u003e14, 15\u003c/sup\u003e Our model identifies the crucial features related to poor mental development in preterm infants in early childhood, hoping to detect children with a high risk of poor neurodevelopmental outcomes for early intervention.\u003c/p\u003e\n\u003cp\u003eLow BW is a well-established factor associated with impaired cognitive development in children.\u003csup\u003e16-18\u003c/sup\u003e According to the World Health Organization (WHO), around 20 million infants are born with low birth weight each year.\u003csup\u003e19\u003c/sup\u003e Low birth weight often reflects underlying complexities, including prematurity, malnutrition, underdeveloped brain structure, and a higher prevalence of comorbidities such as intraventricular hemorrhages, all of which can adversely impact cognitive outcomes.\u003csup\u003e16\u003c/sup\u003e Previous studies showed that the impact of low birth weight on adverse cognitive outcomes may persist into adolescence and early adulthood.\u003csup\u003e20\u003c/sup\u003e Through SHAP analysis, our study identified BW as one of the most influential factors contributing to adverse cognitive outcomes in preterm infants. Timely evaluation and intervention should be done to improve the cognitive outcome in children with low BW.\u003c/p\u003e\n\u003cp\u003ePreterm infants with ROP are at a significantly increased risk of cognitive delays. Previous studies showed children with ROP may be more likely to experience motor, mental, and language delays compared with those without ROP.\u003csup\u003e21-23\u003c/sup\u003e Our study included the key elements of ROP, the stage, the zone, the presence of plus disease, and the treatment times of ROP. We found that the stage and zone of ROP are two of the leading factors, among others, that are related to poor cognitive development. Our findings emphasize the importance of ROP screening and monitoring to decrease the risk of long-term cognitive impairments.\u003c/p\u003e\n\u003cp\u003eRetinal thickness was a potential biomarker associated with cognitive development in children in our study. Embryologically speaking, the ocular structure is an extension of the central nervous system. Previous studies have demonstrated a relationship between a thinner mean retinal nerve fiber layer (RNFL) and lower IQ scores, suggesting that the retina is a biological correlate of cognitive functioning and might be a potential indicator of overall brain function.\u003csup\u003e24, 25\u003c/sup\u003e Rothman \u003cem\u003eet al.\u003c/em\u003e\u003csup\u003e26\u003c/sup\u003e identified cystoid macular edema as a biomarker of poor neurodevelopmental outcomes in preterm infants, further supporting the link between retinal pathology and developmental cognitive deficits. Our study showed that a thicker fovea was associated with worse cognitive performance. This finding aligns with prior observations that children with ROP may have a thicker fovea compared to those without ROP, and ROP itself is a risk factor for adverse cognitive outcomes.\u003csup\u003e21-23\u003c/sup\u003e\u0026nbsp; Interestingly, we also observed that the asymmetrical retinal thickness may slightly improve the accuracy of our model. The difference in retinal thickness in bilateral eyes may indicate different severities of ROP in individual eyes. Although the exact mechanism is unclear, this finding may highlight the importance of detailed retinal assessments in evaluating children with high risk of cognitive deficits.\u003c/p\u003e\n\u003cp\u003eOur ML model was a pioneer model incorporating OCT information in evaluating the cognitive development in preterm infants. OCT is a non-invasive, high-resolution imaging modality crucial in assessing retinal structures. The previous study\u003csup\u003e27\u003c/sup\u003e proved that the RNFL thickness was associated with cognitive functional decline in adult patients. Barrett-Young \u003cem\u003eet al\u003c/em\u003e.\u003csup\u003e24\u003c/sup\u003e reported that the thinner RNFL and ganglion cell layer (GCL) were associated with lower IQ in childhood and the middle ages. In our ML model, we incorporated the OCT information, including the foveal and parafoveal thickness, and the fovea and parafoveal difference in the bilateral eyes, with clinical features to provide a more comprehensive and objective assessment. By adding the OCT information, we demonstrated a significant improvement in both accuracy and the AUC.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWhile the presented ML model identifies the important features associated with cognitive development in newbornswith prematurity in early childhood, several limitations should be acknowledged. First, our training relied solely on basic clinical characteristics and eye examination data. Genetic information and external factors, such as parental educational status and environmental influences, were excluded from our model. While these factors provide additional insights, their exclusion could also be considered a strength. Since such information is not routinely recorded in clinical settings, our ML model demonstrates the feasibility of using readily available data. Despite this limitation, the model achieved a high overall accuracy exceeding 83%, underscoring its robustness and applicability in daily practice. Future studies incorporating genetic and environmental information could be conducted to enhance the model's comprehensiveness. In addition, since our study cohort excluded children with IQ scores between 85 and 115, this model is not suitable as a screening tool for early intervention. However, this study highlighted key ocular features associated with cognitive outcomes, providing crucial insights into ocular factors possibly linked to mental development. Finally, the lack of external validation on independent cohorts may limit the model's utility in real-world clinical settings. External validation with diverse populations across different races might improve the generalizability of our model and evaluate the model's performance in real-world settings.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis study is the first to integrate ocular examination into evaluating cognitive outcomes in preterm infants. By assessing the ML model, we identified key clinical features associated with cognitive development in preterm newborns in early childhood. To enhance the model's interpretability, we employed SHAP analysis, highlighting the significance of the BW, the stage of ROP, and zone development in influencing cognitive outcomes. This study demonstrates the potential of using ML models to predict IQ in children and identify those at high risk for poor cognitive outcomes, paving the way for timely interventions to enhance long-term neurodevelopmental outcomes.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eGrant Support:\u0026nbsp;\u003c/strong\u003eNSTC 113-2314-B-182A-040, CMRPG3P0051~2, CMRPG3P0191, PSC-CUNY, Award # 65406-00 53, NSTC 113-2221-E-194-034.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDisclosure:\u0026nbsp;\u003c/strong\u003eThe sponsors had no role in the design or conduct of this research.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study adhered to the Declaration of Helsinki and\u0026nbsp;was approved by the Institutional Review Board of Chang Gung Memorial Hospital (IRB number: 202401555B0), which granted a waiver of consent because patient anonymity was maintained by the data source.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data analyzed during this study are available on request from the corresponding authors, Wei-Chi, Wu, and Chia-Ling Tsai.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors' contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll authors have participated directly in the planning and execution of the work.\u003c/p\u003e\n\u003cp\u003eMMH: interpreted data, drafting and writing the article; MLM, HYS: acquisition and analysis of data; WYL, RFC, CLT, WCW: design of the study, acquisition of data, final approval. All authors read and approved the final manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eOlusanya BO, Smythe T, Ogbo FA, Nair MKC, Scher M, Davis AC (2023) Global prevalence of developmental disabilities in children and adolescents: A systematic umbrella review. Front Public Health 11:1122009\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMcKenzie K, Milton M, Smith G, Ouellette-Kuntz H (2016) Systematic Review of the Prevalence and Incidence of Intellectual Disabilities: Current Trends and Issues. Curr Dev Disorders Rep 3:104\u0026ndash;115\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDrost FJ, Keunen K, Moeskops P et al (2018) Severe retinopathy of prematurity is associated with reduced cerebellar and brainstem volumes at term and neurodevelopmental deficits at 2 years. Pediatr Res 83:818\u0026ndash;824\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGodugu S, Vudugula SA, Neupane B et al (2023) Association between severe retinopathy of prematurity (ROP) and poor motor neurodevelopmental outcome. Indian J Ophthalmol 71:2944\u0026ndash;2946\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMorken TS, Dammann O, Skranes J, Austeng D (2019) Retinopathy of prematurity, visual and neurodevelopmental outcome, and imaging of the central nervous system. Semin Perinatol 43:381\u0026ndash;389\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLinsell L, Malouf R, Morris J, Kurinczuk JJ, Marlow N (2015) Prognostic Factors for Poor Cognitive Development in Children Born Very Preterm or With Very Low Birth Weight: A Systematic Review. JAMA Pediatr 169:1162\u0026ndash;1172\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAnderson PJ (2023) Predicting neurodevelopmental outcome in children born very preterm \u0026ndash; does neonatal MRI have a role? Pediatr Res 94:868\u0026ndash;869\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWechsler D (2014) Wechsler Intelligence Scale for Children\u0026ndash;Fifth Edition. WISC-V) APA PsycTests.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFaramarzi R, Darabi A, Emadzadeh M, Maamouri G, Rezvani R (2023) Predicting neurodevelopmental outcomes in preterm infants: A comprehensive evaluation of neonatal and maternal risk factors. Early Hum Dev 184:105834\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWechsler D (1967) WPPSI: Wechsler preschool and primary scale of intelligence. Psychological Corporation, [1967] \u0026copy;1967, New York, N.Y\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWechsler D (2003) Wechsler Intelligence Scale for Children, Fourth Edition (WISC-IV). \u003cem\u003eAPA PsycTests\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSchneider W, Niklas F, Schmiedeler S (2014) Intellectual development from early childhood to early adulthood: The impact of early IQ differences on stability and change over time. Learn Individual Differences 32:156\u0026ndash;162\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMuennig P, Robertson D, Johnson G, Campbell F, Pungello EP, Neidell M (2011) The effect of an early education program on adult health: the Carolina Abecedarian Project randomized controlled trial. Am J Public Health 101:512\u0026ndash;516\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSpittle A, Orton J, Anderson PJ, Boyd R, Doyle LW (2015) Early developmental intervention programmes provided post hospital discharge to prevent motor and cognitive impairment in preterm infants. Cochrane Database Syst Rev 2015:Cd005495\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAnderson PJ, Treyvaud K, Spittle AJ. Early developmental interventions for infants born very preterm \u0026ndash; what works? \u003cem\u003eSeminars in Fetal and Neonatal Medicine\u003c/em\u003e 2020;25\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGu H, Wang L, Liu L et al (2017) A gradient relationship between low birth weight and IQ: A meta-analysis. Sci Rep 7:18035\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBreslau N, Dickens WT, Flynn JR, Peterson EL, Lucia VC (2006) Low birthweight and social disadvantage: Tracking their relationship with children's IQ during the period of school attendance. Intelligence 34:351\u0026ndash;362\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eEves R, Mendon\u0026ccedil;a M, Baumann N et al (2021) Association of Very Preterm Birth or Very Low Birth Weight With Intelligence in Adulthood: An Individual Participant Data Meta-analysis. JAMA Pediatr 175:e211058\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOrganization WH (2019) UNICEF-WHO low birthweight estimates: levels and trends 2000\u0026ndash;2015. World Health Organization\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKormos CE, Wilkinson AJ, Davey CJ, Cunningham AJ (2014) Low birth weight and intelligence in adolescence and early adulthood: a meta-analysis. J Public Health (Oxf) 36:213\u0026ndash;224\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAhn JH, Lee KM, Kim MJ et al (2022) Neurodevelopmental outcomes in very low birthweight infants with retinopathy of prematurity in a nationwide cohort study. Sci Rep 12:5053\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHuang CC, Chu CH, Lin YK, Lin YC, Huang HM, Chang YC (2022) Retinopathy of Prematurity Is a Biomarker for Pathological Processes in the Immature Brain. Neonatology 119:727\u0026ndash;734\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBeligere N, Perumalswamy V, Tandon M et al (2015) Retinopathy of prematurity and neurodevelopmental disabilities in premature infants. Semin Fetal Neonatal Med 20:346\u0026ndash;353\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBarrett-Young A, Ambler A, Cheyne K et al (2022) Associations Between Retinal Nerve Fiber Layer and Ganglion Cell Layer in Middle Age and Cognition From Childhood to Adulthood. JAMA Ophthalmol 140:262\u0026ndash;268\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRothman AL, Sevilla MB, Mangalesh S et al (2015) Thinner Retinal Nerve Fiber Layer in Very Preterm Versus Term Infants and Relationship to Brain Anatomy and Neurodevelopment. Am J Ophthalmol 160:1296\u0026ndash;1308e1292\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRothman AL, Tran-Viet D, Gustafson KE et al (2015) Poorer neurodevelopmental outcomes associated with cystoid macular edema identified in preterm infants in the intensive care nursery. Ophthalmology 122:610\u0026ndash;619\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKo F, Muthy ZA, Gallacher J et al (2018) Association of Retinal Nerve Fiber Layer Thinning With Current and Future Cognitive Decline: A Study Using Optical Coherence Tomography. JAMA Neurol 75:1198\u0026ndash;1205\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003e\u003cstrong\u003eTable 1.\u0026nbsp;\u003c/strong\u003eDemographics in Our Study Cohort\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eIntelligence Quotient (IQ)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eIQ \u0026lt; 85\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" style=\"width: 180px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eIQ \u0026ge; 115\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u003cem\u003ep\u003c/em\u003e\u003c/strong\u003e\u003cstrong\u003e-value\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003ePatients, (eye)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e86\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(172)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e77\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(154)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003eAge at IQ evaluation, yr (SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e7.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(2.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e8.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(2.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003eGender, Male, (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(64.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e41\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(53.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e0.167\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003eGA, wk, (SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e29.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(4.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e35.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(4.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003eBW, g, (SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e1359.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(852.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e2365.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(918.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003eApgar score at 1 min\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e5.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(2.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e7.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(1.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003eApgar score at 5 min\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e7.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(2.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e9.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(1.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003eTerm infant, n, (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(9.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(51.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003ePreterm infant without ROP, n, (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e24\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(27.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e31\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(40.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003eROP, n, (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(62.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(7.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003eROP stage, eye, (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003e\u0026nbsp; 1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(11.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003e\u0026nbsp; 2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(11.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(3.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003e\u0026nbsp; 3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e64\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(37.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(4.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003e\u0026nbsp; 4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(0.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003e\u0026nbsp; 5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(1.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003ePlus disease, eye, (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e63\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(36.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(3.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003eZone, eye, (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003e\u0026nbsp; 1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(11)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003e\u0026nbsp; 2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e79\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(45.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(6.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003e\u0026nbsp; 3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e74\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(43.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e144\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(93.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003eROP with treatment, eye\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003eIntravitreal anti-VEGF Injection\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e44\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(25.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(5.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003eLaser Photocoagulation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(1.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e0.146\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003eCombine Intravitreal Injection\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eand Laser Photocoagulation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(12.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003eVitrectomy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e(2.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e(0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e0.076\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 293px;\"\u003e\n \u003cp\u003eAverage treatment times, n\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e1.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 85px;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 123px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eBW, birth weight; GA, gestational age; ROP, retinopathy of prematurity; SD: standard deviation; VEGF, vascular endothelial growth factor\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2.\u0026nbsp;\u003c/strong\u003eThe Performance Summary of Our Model\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"794\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 142px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eFeatures\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 93px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAUC (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eIQ score\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePrecision\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecall\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eF1-score\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width: 142px;\"\u003e\n \u003cp\u003eClinical features\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" style=\"width: 93px;\"\u003e\n \u003cp\u003e79.55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" style=\"width: 115px;\"\u003e\n \u003cp\u003e78.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026lt; 85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e0.86\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e0.72\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122px;\"\u003e\n \u003cp\u003e0.78\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e≧115\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e0.73\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e0.87\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122px;\"\u003e\n \u003cp\u003e0.80\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width: 142px;\"\u003e\n \u003cp\u003eClinical features, retinal thickness\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" style=\"width: 93px;\"\u003e\n \u003cp\u003e83.86\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" style=\"width: 115px;\"\u003e\n \u003cp\u003e83.13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026lt; 85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e0.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122px;\"\u003e\n \u003cp\u003e0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e≧115\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e0.79\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122px;\"\u003e\n \u003cp\u003e0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width: 142px;\"\u003e\n \u003cp\u003eClinical features,\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eretinal thickness, retinal thickness difference in bilateral eyes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" style=\"width: 93px;\"\u003e\n \u003cp\u003e83.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" style=\"width: 115px;\"\u003e\n \u003cp\u003e83.44\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e\u0026lt; 85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e0.79\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122px;\"\u003e\n \u003cp\u003e0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 105px;\"\u003e\n \u003cp\u003e≧115\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e0.79\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122px;\"\u003e\n \u003cp\u003e0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eAUC, area under curve, IQ, intelligence quotient\u003c/p\u003e\n"}],"fulltextSource":"","fullText":"","funders":[{"identity":"7d0b90fe-06fe-4270-995b-e420365bd360","identifier":"10.13039/501100004192","name":"National Science and Technology Development Agency","awardNumber":"113-2314-B-182A-040, 113-2221-E-194-034","order_by":0},{"identity":"ec7cd6fb-421f-4114-a35d-666db36cb829","identifier":"10.13039/100012553","name":"Chang Gung Memorial Hospital","awardNumber":"G3P0051~2, CMRPG3P0191","order_by":1},{"identity":"df8a4207-8eaa-4255-8cea-42f712fc1db6","identifier":"10.13039/100006462","name":"City University of New York","awardNumber":"Award # 65406-00 53","order_by":2}],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Linkou Chang Gung Memorial Hospital","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"cognitive outcome, intelligence quotient, machine learning, prematurity of retinopathy, preterm infant","lastPublishedDoi":"10.21203/rs.3.rs-7172977/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7172977/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003ePurpose\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo use machine learning (ML) to identify the important retinal features associated with cognitive development in newborns with prematurity in early childhood, with the goal of screening preterm babies with poor neurodevelopment for early intervention.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe retrospectively reviewed the medical charts of 163 infants (326 eyes) born at Chang Gung Memorial Hospital between 2011 and 2024 who underwent IQ testing and scored \u0026lt; 85 or ≥ 115. Essential characteristics considered predictors come from perinatal variables and eye examination for retinopathy of prematurity (ROP) with optical coherence tomography (OCT) images. The ML model was trained using Random Forest (RF) to collectively associate the predictors with the cognitive assessment outcome. SHAP (Shapley Additive exPlanations) was used to identify the clinical features most predictive of the mental development of a preterm baby.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe overall accuracy of the ML model was 83.44%, combining all the clinical characteristics, including retinal thickness and retinal thickness difference in bilateral eyes obtained from OCT. The SHAP analysis reveals that lower birth weight, higher stage of ROP, and lower zone development are highly associated with lower IQ scores in our study cohort.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThrough an ML approach, this study identified the BW and the stage of ROP and zone as the leading ocular features associated with cognitive outcomes in preterm newborns in early childhood. It paved the way for timely interventions to improve long-term neurodevelopmental outcomes.\u003c/p\u003e","manuscriptTitle":"From Eye Examination to Early Cognitive Evaluation for Preterm Newborns: An Explainable Machine Learning Approach","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-27 10:54:11","doi":"10.21203/rs.3.rs-7172977/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"2cb10248-df20-46b9-856d-c331f6bfcd47","owner":[],"postedDate":"July 27th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":51823832,"name":"Ophthalmology"},{"id":51823833,"name":"Pediatrics"},{"id":51823834,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2025-07-27T10:54:11+00:00","versionOfRecord":[],"versionCreatedAt":"2025-07-27 10:54:11","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7172977","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7172977","identity":"rs-7172977","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}