#431 - [Journal Club] - 📌 Can Retinal Images Predict BPD and Pulmonary Hypertension?

In this Journal Club episode, Ben and Daphna explore an exciting new frontier in neonatology: oculomics. Reviewing a recent paper from JAMA Ophthalmology, they discuss how deep learning models applied to routine ROP screening images can predict the development of BPD and pulmonary hypertension in preterm infants. By combining visual features extracted via neural networks with standard demographic data, researchers achieved impressive predictive accuracy weeks before clinical diagnosis is typically made. Tune in to hear how the eyes might just be the window to the neonatal pulmonary vasculature!

----

Deep Learning-Based Prediction of Cardiopulmonary Disease in Retinal Images of Premature Infants. Singh P, Kumar S, Tyagi R, Young BK, Jordan BK, Scottoline B, Evers PD, Ostmo S, Coyner AS, Lin WC, Gupta A, Erdogmus D, Chan RVP, McCourt EA, Barry JS, McEvoy CT, Chiang MF, Campbell JP, Kalpathy-Cramer J.JAMA Ophthalmol. 2026 Jan 22:e255814. doi: 10.1001/jamaophthalmol.2025.5814. Online ahead of print.PMID: 41569552

Transcript

Ben Courchia MD (00:00.568) Hello everybody, welcome back to the incubator podcast. We're back today for another episode of Journal Club. It is Thursday and I have another interesting article for you, Daphna. This one is coming straight out of JAMA Ophthalmology. Not always, yeah. The paper is titled, I know, but this one has one of our good friends on the paper, Dr. James Berry.

Daphna Yasova Barbeau MD (00:06.894) Hmm.

Daphna Yasova Barbeau MD (00:13.703) That's not always on our radar.

Ben Courchia MD (00:25.038) And the paper is called Deep Learning Based Prediction of Cardio-Pulmonary Disease in Retinal Images of Preterm Infants. I know. Yesterday we spoke about BPD and pulmonary hypertension. And here, this is also going to be a center point of this particular research. The authors mentioned that it is a leading cause of...

Daphna Yasova Barbeau MD (00:30.808) Very cool. It's coming.

Ben Courchia MD (00:47.296) ...morbidity and mortality in premature infants and current prediction tools rely primarily on demographics such as gestational age at birth and birth weight, which are easy to obtain, but obviously are not the most sensitive or specific to be clinically actionable. Now they're mentioning that recent advances in deep learning have demonstrated the ability to diagnose non-ocular disease from retinal images, a field that is often referred to as oculomics. In this project, the authors analyzed a large dataset of fundus images obtained as part of routine ROP examinations. Their objective was to determine if these images could identify patients who would later develop BPD or pulmonary hypertension. Crucially, to ensure that this was predictive rather than just diagnostic, they limited their analysis to the images obtained at 34 weeks post-menstrual age or less, specifically chosen to precede the standard clinical diagnosis window for BPD and pulmonary hypertension. Have I piqued your curiosity? You're nodding, but you're muted.

Daphna Yasova Barbeau MD (01:53.646) Yeah, as soon as that. Absolutely. I think it's super exciting. Plus it would be cool as the neonatologist to just see more images of what we're looking at. What do they say? The eyes are the window to the soul. To the whole body, I guess. But isn't that true? When we have a kid that we're like, hmm, something's up with this kid, genetics maybe. Let's get an eye exam.

Ben Courchia MD (01:57.964) Ahem.

Ben Courchia MD (02:04.845) Yeah.

Ben Courchia MD (02:08.898) The window to the soul, but it looks like the eyes are the window to your pulmonary vasculature.

Ben Courchia MD (02:20.558) Yeah. The study utilized data from the Multi-Institutional Imaging and Informatics in Retinopathy of Prematurity, the iROP study, collected between 2012 and 2020. The analysis included infants recruited from seven neonatal intensive care units. Infants were eligible if they met criteria for ROP screening, having a birth weight of 1500 grams or less...

Daphna Yasova Barbeau MD (02:22.198) Infection. Let's get an eye.

Ben Courchia MD (02:49.774) ...or a gestational age less than 31 weeks. The authors developed a multimodal model. First, they used a deep learning model with a ResNet 18 architecture to extract visual features from the retinal fundus photos. Then, as a second layer of that multimodal model, they combined these visual features with demographic data. They added birth weight, gestational age, and post-menstrual age. Third, these combined inputs were then fed to a support vector machine classifier to predict the final diagnosis. I usually would ask if you have any questions about that; please do not ask me any questions about that. I am reporting the paper, and I would not be able to explain the intricacies to you. I'm just going to be very honest.

Daphna Yasova Barbeau MD (03:43.214) I appreciate your honesty.

Ben Courchia MD (03:47.864) But that's also how I read a lot of these AI papers. I think that's an important point. I don't understand the intricacies of the ResNet 18 architecture and so on. Don't get bogged down with those particular details. Just keep reading through to understand how they're leveraging these tools, which is what I did. To ensure that the model was not simply detecting the severity of retinopathy of prematurity itself, they trained the secondary model using only the images that had no clinical signs of ROP as well. The definition of BPD that they were using was the oxygen requirement at 36 weeks, and pulmonary hypertension was diagnosed using echocardiography at 34 weeks, defined as a pulmonary artery pressure greater than one-half systemic pressures. Any questions about any of that? No, thank you. We're moving right along.

Daphna Yasova Barbeau MD (04:40.395) No, I think we're gonna go. We're moving right along.

Ben Courchia MD (04:46.518) In terms of the outcomes. The analysis included 493 infants in the BPD cohort. The mean gestational age was 25.7 weeks for infants who had BPD compared to 27.3 weeks for those who did not. The PH cohort was a subset of 184 infants from a single site out of these 400 and something patients. In terms of the prediction of BPD, the authors reported performance on a held-out test set using the area under the receiver operating characteristic curve, the AUC. They found that using the demographics alone, an AUC of 0.72 was achieved. Using an imaging-only model, they achieved the exact same type of AUC, 0.72. But when they were using the multimodal model using this SVM, the support vector machine, they were able to achieve a much higher area under the receiver operating characteristic curve of 0.82. This improvement was statistically significant compared to imaging alone and approached significance compared to demographics alone. When it came to the prediction of pulmonary hypertension, the demographics alone model performed modestly with an AUC of 0.68. The imaging-only model performed remarkably better with an AUC of 0.91. Using the multimodal model, it also achieved an AUC of 0.91, showing no additional benefit over imaging alone. Importantly, these results persisted even when the models were trained on images lacking any significant signs of ROP, suggesting that the algorithm was detecting identifying features independent of standard ROP pathology. So you didn't really need to have underlying ROP to basically make the prediction. It's a relatively small review of this paper. The authors conclude that retinal images obtained during ROP screening can predict the diagnosis of BPD and pulmonary hypertension in preterm infants. In their discussion, they offer a compelling hypothesis for why this works. This suggests that...

Ben Courchia MD (07:05.718) ...mechanical ventilation or CPAP may lead to subtle changes in the retinal or choroidal vasculature that a deep learning algorithm is able to detect even before overt clinical symptoms appear. The findings of the paper suggest that oculomics could lead to earlier diagnosis and potentially avoid the need for invasive diagnostic testing, such as cardiac catheterization in the future. I don't know about that, but I know that we're all trying to figure out a way to screen for pulmonary hypertension in the least invasive way. This doesn't seem like it's a big stretch using data that is already collected. So I thought this was quite interesting.

Daphna Yasova Barbeau MD (07:44.983) I think that is super cool, especially that it's not a separate eye exam. It's the same eye exam. They're going to get them anyways, the babies that are most at risk. And if they start getting them around 31 weeks and we can already have more information, that may manage how we wean babies off of ventilatory support. Do we offer them other medication support? I think it's super exciting.

Ben Courchia MD (07:53.74) Mm-hmm.

Ben Courchia MD (08:12.844) Yeah. I think it's only the beginning of things that we're going to see AI able to do a decent job at that otherwise was completely untapped potential. Who has ever looked at ROP screening images to actually look at BPD, pulmonary hypertension, and so on? This is completely a new frontier.

Daphna Yasova Barbeau MD (08:24.525) For sure.

Daphna Yasova Barbeau MD (08:34.241) Yeah, and it's not like we're great at predicting it anyway. So it's awesome. Very cool.

Ben Courchia MD (08:38.144) Anyway, yeah. All right, buddy. I will see you tomorrow for some new news. Bye.

Daphna Yasova Barbeau MD (08:45.559) Sounds good.