Every year in the United States, approximately 18,000 very preterm infants are affected by bronchopulmonary dysplasia (BPD), a chronic lung disease that often leads to prolonged hospitalizations and long-term neurodevelopmental challenges. For many of these infants, the situation is further complicated by pulmonary hypertension (PH), a condition where high blood pressure in the lungs can increase the risk of mortality by 4.6 times compared to those with BPD alone.
Identifying these conditions early is critical, but current methods are often either insufficiently precise or overly invasive. However, groundbreaking research is suggesting that the “window to the soul”—the eye—might actually be a window to an infant’s cardiopulmonary health.
What is Oculomics?
The field of “oculomics” is based on the emerging ability of deep learning (DL) algorithms to diagnose systemic, non-ocular diseases by analyzing retinal images. Because the retina’s blood vessels are easily visible and share characteristics with vessels throughout the body, they can reveal signs of cardiovascular, kidney, and neurological disorders.
In a recent multi-institutional study, researchers applied this concept to the most vulnerable patients: premature infants in the neonatal intensive care unit (NICU).
The Study: Retinal Imaging Beyond Eye Disease
Preterm infants already undergo routine screening for retinopathy of prematurity (ROP) using digital fundus cameras. Researchers analyzed 1,357 retinal images from 493 infants to see if the AI could detect features associated with BPD and PH before they were clinically diagnosed.
The team compared three different types of predictive models:
- Demographics-only: Based on birth weight (BW) and gestational age (GA).
- Imaging-only: Based solely on features extracted from retinal photographs by an AI model called ResNet18.
- Multimodal: A combination of both retinal imaging and demographic data.
Breakthrough Results
The study found that AI could predict these life-threatening conditions with impressive accuracy, often outperforming traditional methods:
- For BPD: The multimodal model achieved an accuracy (AUC) of 0.82, proving more effective than using demographics or imaging alone.
- For PH: The results were even more striking. Retinal imaging was highly diagnostic on its own, with an AUC of 0.91, significantly outperforming the demographics-only model, which scored just 0.68.
Crucially, these findings remained consistent even when the AI analyzed images that showed no visible signs of ROP. This suggests that the AI is identifying unique vascular biomarkers in the retina related specifically to heart and lung health, rather than just piggybacking on signs of eye disease.
Why This Matters for Neonatal Care
The implications of this “proof-of-concept” study are profound for the future of neonatal medicine:
- Earlier Intervention: Predicting BPD as early as 31 to 32 weeks’ postmenstrual age could allow neonatologists to implement more aggressive or targeted pulmonary management sooner.
- Reducing Invasive Testing: Currently, the “gold standard” for diagnosing PH is invasive cardiac catheterization. Retinal imaging could provide a non-invasive screening tool that prompts earlier echocardiography and potentially spares infants from riskier procedures.
- Seamless Integration: Since retinal imaging is already a standard part of ROP screening, this AI technology has a “built-in” path for clinical deployment without requiring new, expensive equipment.
The Path Ahead
While these results are “hypothesis-generating,” further validation in larger, more diverse populations is necessary. Researchers hope to eventually move from 2D photographs to 3D imaging (like OCTA) to provide even more detailed data on biomarkers for BPD and PH.
By turning routine eye exams into a comprehensive health check, AI is opening new doors to protecting the lives of the smallest and most fragile patients.
Reference: Singh, P., Kumar, S., Tyagi, R., et al. (2026). Deep Learning–Based Prediction of Cardiopulmonary Disease in Retinal Images of Premature Infants. JAMA Ophthalmology, 144(3), 230-237.
Interesting study!