Hypertension
Poster Session 4
Liat Shenhav, PhD
Assistant Professor
New York University
New York, NY, United States
Andrew Laine, DSc
Professor
Columbia University
New York, NY, United States
Kaveri Thakoor, PhD
Assistant Professor
Columbia University
New York, NY, United States
Ye Tian, BA
Student
Columbia University
New York, NY, United States
Geoffrey Wu, BA
Student
Columbia University
New York, NY, United States
Zoilo Castillo, BA
Research Coordinator
Columbia University
New York, NY, United States
Yessenia Gutierrez, BA
Columbia University
New York, NY, United States
Lisa Hark, MBA, PhD
Columbia University Medical Center
New York, NY, United States
Ronald J. Wapner, MD (he/him/his)
Professor of OBGYN; Director of Reproductive Genetics
Columbia University Irving Medical Center
New York, NY, United States
Srilaxmi Bearelly, MD
Associate Professor
Columbia University
Stony Brook, NY, United States
Preeclampsia is one of the most severe pregnancy complications and a leading cause of maternal death. Notably, the initial vascular alterations leading to preeclampsia occur in the placenta and maternal vasculature well before symptoms develop. This observation is key to novel diagnostics. Here we evaluate, for the first time, changes in the retinal vasculature occurring prior to the patient becoming symptomatic and use an Artificial Intelligence (AI) framework to develop and test a vascularity based predictive algorithm.
Study Design:
Recent advancements in ocular imaging now provide high-resolution, cheap, and non-invasive imaging of the retina. Using the Optos Daytona machine, we collected retinal images from 300 pregnant women at 1st, 2nd and 3rd trimester visits. We further developed an AI framework, combining topological data analysis and deep learning termed DVT-Net (Deep Vascular Topology Network), suitable to characterize the topological and geometrical changes in the retinal vasculature to predict preeclampsia (Figure 1).
Results:
From the cohort of 300 pregnancies, 12 women were diagnosed with preeclampsia, 40 with gestational hypertension and 248 had neither. Using images from the first and second trimesters, our computational framework (DVT-Net) and a 10-fold cross validation, we achieved an accuracy of 0.813 (95% CI [0.76, 0.91]) in predicting the above-mentioned pregnancy-disorders.
Conclusion:
Our approach provides a nuanced description of the retinal vasculature topology in normal and hypertensive pregnancies and serves as an accessible and sensitive imaging surrogate of placental vasculature changes. This may allow early prediction of preeclampsia and gestational hypertension.