Congenital ophthalmologic malformations (COMs) encompass a variety of anomalies that can severely compromise patients’ quality of life and still relevant causes of evitable childhood blindness.⁽¹⁾ These malformations include congenital cataracts, congenital glaucoma, and colobomas, which often have genetic origins or result from teratogenic exposures during embryonic development.⁽²⁾ These conditions pose significant diagnostic and therapeutic challenges, especially in regions with limited access to specialized services.

Recent studies have explored the application of Artificial Intelligence (AI) in the diagnosis of ophthalmological congenital malformations, demonstrating great potential in the analysis of medical images and early identification of structural anomalies.^(3–5) Deep learning models, such as convolutional neural networks, have been used to diagnose congenital cataracts with levels of accuracy comparable to that of human experts.⁽⁵⁾ Other applications include the automated detection of retinopathy of prematurity⁽¹⁾ and the use of supervised learning algorithms to classify pediatric conditions such as strabismus and refractive errors.⁽⁶⁾ Despite these advances, the proposed methods until now depend on the acquisition of high-quality images and/or the availability of high computing power, making them difficult to use in scenarios with fewer resources.

Reviews reinforce AI's transformative potential in ophthalmology, particularly for early detection of sight-threatening diseases such as cataract, glaucoma, and retinopathy of prematurity. AI has been shown to deliver diagnostic accuracy comparable – or even superior – to experienced ophthalmologists, with the added advantage of scalability to underserved populations.^(7,8) A recent meta-analysis confirmed that AI significantly improves diagnostic accuracy across ophthalmic diseases.⁽⁹⁾ Furthermore, novel implementations, such as real-time AI systems integrated into ultrasonography, have shown efficacy in detecting fetal intracranial malformations, expanding the potential of AI in prenatal screening.⁽¹⁰⁾

Most recently, AI-driven cloud-based collaborative platforms have been validated in multicenter trials for congenital cataracts, showing comparable diagnostic and treatment decision performance to expert ophthalmologists while facilitating wider access to care.⁽⁵⁾ Reviews and prospective analyses highlight the critical role AI can play in expanding access to early ophthalmic diagnosis, optimizing treatment strategies, and supporting ophthalmic care in low-resource environments.^(11,12)

Therefore, this study aims to evaluate whether an AI model based on easily accessible perinatal variables can predict COMs, particularly those whose late detection may lead to preventable blindness.

This study stands as a proof of concept that applying AI and traditional LR analysis with the aim of predicting children with high risk for COMs that may be hard to diagnose without proper ophthalmological evaluation can yield great accuracy even with limited number of characteristics available.

The prediction models obtained considerable and similar performance when subjected to general analysis with LR AUC of 0.833 and SVM with AUC of 0.834. In the sub-analyses, the ML model obtained better performance in conditions of difficult detection in the physical examination and the traditional mathematical model in conditions where early screening is potentially beneficial. The variables that weighed most in the decision of the models were weight, Apgar, mother's age, gestational age and the number of non-living children. These data are consistent, since the association between congenital malformations and the factors mentioned is documented in the scientific literature.⁽¹⁶⁾

AI models have already been validated as an important tool to assist ophthalmologists in various diseases such as diabetic retinopathy, retinopathy of prematurity and retinoblastoma,^(1–3) with most studies training prediction models on optical coherence tomography (OCT) images or fundus images. However, few studies evaluate physical examination characteristics, symptoms and other clinical and perinatal characteristics in COMs, such as our study that evaluated the latter cluster. Another example would be the study by Long et al., 2017, one of the largest multicenter pilot studies that applied AI to differentiate congenital cataracts from normal eyes from ocular images with a sensitivity of 100% and 97.67% specificity.⁽⁵⁾ Another large case-control study on congenital cataract that used the Random Forest prediction model and other models assessing clinical characteristics also evaluated in our study, such as gender of the newborn and preterm births, demonstrates a similar sensitivity of 80% and a high specificity of 90%, highlighting important factors such as parental education, family history, and the presence of comorbidities.⁽¹⁷⁾

In addition to congenital cataract, the study by Marouf et al. evaluated a set of eye diseases, including primary congenital glaucoma through symptoms and changes in the physical examination, also demonstrating a high accuracy of the SVM model with 99.11% and LR of 98.58% and highlighted the clinical variables cloudy, blurry or foggy vision, double vision, and problem in identifying color.⁽¹⁸⁾ Comparable advances have been reported in congenital systemic conditions. For example, AI models integrating perinatal data have successfully predicted neonatal outcomes, congenital abnormalities, and other risks, further validating the importance of easily accessible predictors in precision medicine.^(19,20)

The extension of these predictive approaches to ophthalmology has the potential to reshape neonatal screening paradigms by complementing, rather than replacing, existing diagnostic workflows. A fundamental implication lies in the opportunity to transform perinatal and early-life health data – routinely collected and readily available – into actionable insights for ophthalmic risk stratification. Unlike imaging-based models, which demand expensive infrastructure, skilled personnel, and high-quality data acquisition, perinatal-based models leverage universally recorded variables such as birth weight, gestational age, and Apgar scores. This accessibility provides a foundation for democratizing ophthalmic prediction and ensuring that early risk identification is not limited to highly resourced centers.

Moreover, the integration of such AI-driven tools into routine neonatal care has significant potential to enhance healthcare equity. In many countries, ophthalmologic care is disproportionately centralized, with specialized pediatric ophthalmologists clustered in major urban areas. These results occur in delayed diagnosis for infants in remote or underserved settings, where congenital malformations may remain undetected until irreversible visual impairment. By embedding AI-supported screening algorithms into primary care settings, maternity wards, or even mobile health applications, frontline healthcare providers could be empowered to recognize high-risk newborns and initiate timely referrals. Such decentralization of risk detection aligns with global health priorities aimed at reducing preventable childhood blindness and addressing disparities in access to care.⁽¹²⁾

From a clinical management perspective, predictive models could also support the personalization of follow-up protocols. High-risk neonates identified by AI-based tools could be placed under intensified surveillance, ensuring earlier diagnostic imaging and specialist evaluation, while low-risk infants could continue under standard follow-up schedules. This stratified approach would optimize the use of limited ophthalmological resources, reducing unnecessary specialist consultations while simultaneously minimizing the risk of delayed detection in vulnerable groups. Additionally, by highlighting the relative importance of perinatal variables in prediction, these models provide clinicians with a more nuanced understanding of systemic risk factors that may be overlooked in routine practice, thus bridging epidemiological evidence with individualized patient care.

Another important implication is the integration of AI models into telemedicine platforms. Teleophthalmology has already been promising in expanding access to eye care, particularly for conditions such as diabetic retinopathy and retinopathy of prematurity. When combined with AI-driven perinatal risk prediction, telemedicine could create a two-tiered system: infants flagged as high risk could undergo remote consultations supported by image sharing and cloud-based AI diagnostics, ensuring rapid triage and management even in geographically isolated regions. Such synergy between predictive modeling and telehealth infrastructures would be particularly valuable in low- and middle-income countries, where both technological constraints and specialist shortages remain pressing barriers.⁽²¹⁾

At the public health level, predictive AI models could inform population-based screening policies. For example, risk scores derived from perinatal data could be used by ministries of health to guide resource allocation, directing screening campaigns or mobile ophthalmology units to regions with higher predicted prevalence of COMs. Furthermore, predictive analytics could be integrated into maternal and child health registries, providing policymakers with real-time epidemiological insights into the burden of congenital ocular disease and enabling evidence-based planning of preventive strategies.

Finally, these implications also raise ethical and governance considerations. The adoption of AI into neonatal and ophthalmic care requires robust frameworks to ensure transparency, accountability, and explainability. LR offers interpretability that can be directly communicated to clinicians and families, while more complex models such as SVM may demand additional interpretability tools to build trust and prevent the so-called "black box" dilemma. Equally important is the need to ensure that AI-based tools do not exacerbate inequities by being restricted to technologically advanced hospitals. Instead, their true transformative potential lies in being adapted to resource-limited environments, where their predictive value can have the greatest clinical and societal impact.⁽²²⁾

Despite the promising findings, several limitations must be acknowledged. First, although the dataset analyzed was large and multicentric, the exclusion of newborns with missing perinatal variables reduced the effective sample size, particularly in the sub-analyses. This may have influenced the sensitivity of the models and restricted the generalizability of the results. Second, all data originated from a single-country registry, which, although diverse in terms of geography, may not fully capture variations in genetic, environmental, and healthcare factors present in other populations. External validation in different cultural and epidemiological contexts will therefore be essential before broader implementation.

Third, the reliance on routinely collected perinatal data represents both strength and limitation. While these variables are widely available and facilitate scalability, they do not replace direct ophthalmologic assessment or imaging-based diagnostics. The models should thus be seen as triage tools rather than definitive diagnostic systems. Fourth, model explainability remains a broader challenge in adoption of AI. Although LR provides interpretable outputs, the performance of SVM, while slightly superior in certain subgroups, comes with limited transparency, highlighting the tension between accuracy and interpretability in clinical AI applications.

Finally, as with most retrospective studies, the analysis is limited by potential biases inherent to the quality and completeness of registry data. The presence of unmeasured confounders—such as parental education, genetic predispositions, or prenatal exposures—may have influenced outcomes but could not be fully accounted for in our models. Prospective data collection and inclusion of broader clinical and sociodemographic variables may enhance the robustness and predictive value of future iterations of these models.

This study provides a proof of concept that Artificial Intelligence and traditional statistical approaches, when applied to perinatal and clinical variables, can effectively predict congenital ophthalmologic malformations that are otherwise difficult to detect in routine neonatal care. Both logistic regression and support vector machine models demonstrated comparable accuracy, underscoring the feasibility of using readily available, non-invasive predictors to stratify risk in newborns. The identification of birth weight, Apgar scores, maternal age, and gestational age as key determinants reinforces previously established associations between these variables and congenital anomalies, while demonstrating their utility within predictive modeling frameworks.

The broader implications of these findings suggest that Artificial Intelligence-enhanced prediction tools could complement current ophthalmic screening practices, particularly in low-resource settings where access to specialized services is limited. By enabling earlier referral of high-risk infants, such models have the potential of reducing the burden of preventable childhood blindness and optimize resource allocation. Nevertheless, limitations regarding data completeness, generalizability, and model interpretability must be addressed before clinical integration.

Future research should prioritize multicenter and cross-population validation, the incorporation of additional sociodemographic and genetic data, and the development of explainable AI systems to ensure clinical trust. With these refinements, predictive models based on perinatal features could evolve into scalable, equitable, and impactful tools for neonatal ophthalmology, bridging the gap between population-level screening and individualized risk assessment.