Category: Computers
Poster Session III
To assess if artificial intelligence (AI) can learn from ultrasound image characteristics to make reliable predictions about risk of preterm birth.
Study Design:
A proprietary AI software was written and trained on a set of de-identified ultrasound images from a retrospective cohort of women who delivered at the University of Kentucky from 2017 to 2021. The subpopulation of patients with outcome data consisted of 5,714 patients with 19,940 unique US exams and 877,141 total ultrasound images. Data from 4,505 patients (79% of this cohort) was used to train the AI to optimize prediction for preterm birth < 37 w (15,634 unique ultrasound studies and 757,050 total images). Specific anatomic locations (“hot spots”) were identified which provided discrimination for prediction. The date of the exam and ultrasound imaging alone were utilized to suggest whether that specific patient would be expected to deliver preterm ( < 37 w GA) or < 30 days from the examination.
The remaining 21% of the cohort was reserved for derivation and validation of test characteristics. Outcomes for this subgroup were blinded from the AI analysis process with 3rd party independent monitoring. 1,209 unique patients with 4,306 ultrasound studies and 120,091 total images were included. Predictions were made for each patient after each individual exam. Sensitivity, specificity, AUC, PPV and NPV of the AI’s predictions for PTB were made in comparison to the actual GA at delivery.
Results:
Preterm birth rates in the training set (18.4%) and validation set (18.6%) were similar. The test characteristics for the validation set are presented in Table 1 and Figure 1. Hot spots included the cervix, lower uterine segment, amniotic fluid, decidual thickness and vascularization, placental location, and select fetal anatomy.
Conclusion:
AI can predict preterm delivery via ultrasound. Test characteristics varied based on the endpoints selected. Information from hot spots incorporated into the prediction model are from a variety of sources, showing the AI simultaneously combines multiple interrogations into one prediction.
Neil Bharat Patel, MD
Maternal Fetal Medicine Fellow Physician
University of Kentucky
Lexington, Kentucky, United States
John O'Brien, MD
Chandler Medical Center, University of Kentucky
Lexington, Kentucky, United States
Robert Bunn, BS
Ultrasound AI
Highlands Ranch, Colorado, United States
Garrett K. Lam, MD
High Risk Pregnancy Center
Draper, Utah, United States