Patel Neil, O'Brien John, Bunn Robert, Schanbacher Brandon, Bauer John, Lam Garrett K
Department of Maternal Fetal Medicine, Ascension Sacred Heart, Pensacola, FL, USA.
Department of Obstetrics and Gynecology, Division of Maternal Fetal Medicine, University of Kentucky, Lexington, KY, USA.
J Matern Fetal Neonatal Med. 2025 Dec;38(1):2532099. doi: 10.1080/14767058.2025.2532099. Epub 2025 Jul 27.
To evaluate the ability of a proprietary artificial intelligence (AI) model to predict the number of days until delivery using ultrasound images alone and to assess the continuous improvement of prediction accuracy, particularly for preterm births, through model retraining.
An AI software was developed and trained using de-identified ultrasound images from a cohort of women who delivered at the University of Kentucky from 2017 to 2021. Initially, 5,714 pregnant women, with 19,940 unique ultrasound exams and 877,141 total ultrasound images were utilized from this timeframe. Images from 79% of this cohort (4,505 patients) trained the AI to estimate the number of days until delivery and secondarily optimize predictions related to preterm birth (<37 weeks gestational age). The output consisted of days until delivery which was subsequently categorized as either preterm or term birth.The remaining 21% of the cohort (1,209 patients) was reserved for derivation and validation of test characteristics. Delivery outcomes for this subgroup were blinded from the AI by an independent third-party data monitor. Unique predictions were made for each patient after each ultrasound exam, and the AI's performance was evaluated against the actual delivery date using metrics such as R values and mean absolute error (MAE) compared to actual days until delivery. After initial testing, the AI was retrained x3 more using the same data (Version 2, V2) and later with an additional 1,165,618 images obtained by extension of the study to include data from our center until 2023 (Version 3 (V3), Version 4 (V4)- consisted of retraining on V3).
Preterm birth rates were similar between the training (18.4%) and validation (18.6%) sets in the initial study set. The initial AI model exhibited a sensitivity of 39% and specificity of 93% for preterm birth prediction, with an AUC of 0.757. The AI's predictions of days to delivery versus actual in the validation set yielded R of 0.90 for term births, 0.88 for spontaneous preterm birth plus term births, and 0.48 for spontaneous preterm birth alone. The MAE in predicting the number of days until delivery showed similar accuracy across all trimesters that were assessed by image analysis. Finally, retraining with improvements in AI architecture and training methodology using additional images provided improved preterm birth prediction, with R values for all births increasing from 0.85 (V1) to 0.88 (V3) to 0.92 (V4). For spontaneous PTB, MAE was 19.99 days in V4.
AI can predict timing until delivery from ultrasound data alone. This technology can also predict preterm delivery with limited sensitivity. Retraining the AI with supervised and unsupervised learning has the potential to further improve performance.
评估一种专有的人工智能（AI）模型仅使用超声图像预测分娩前天数的能力，并通过模型再训练评估预测准确性的持续提高，特别是对于早产情况。
使用来自2017年至2021年在肯塔基大学分娩的一组女性的去识别超声图像开发并训练了一种AI软件。最初，在这个时间段内使用了5714名孕妇，有19940次独特的超声检查和总共877141张超声图像。该队列中79%（4505名患者）的图像用于训练AI以估计分娩前天数，并其次优化与早产（孕周<37周）相关的预测。输出结果为分娩前天数，随后被分类为早产或足月产。队列中其余21%（1209名患者）被保留用于测试特征的推导和验证。该亚组的分娩结果由独立的第三方数据监测员对AI进行盲法处理。每次超声检查后对每位患者进行独特的预测，并使用如R值和平均绝对误差（MAE）等指标将AI的表现与实际分娩日期进行比较，与实际分娩前天数相对比。在初始测试后，使用相同数据对AI进行了3次再训练（版本2，V2），后来又使用通过将研究扩展至包括我们中心直至2023年获得的另外1165618张图像进行再训练（版本3（V3）、版本4（V4）——包括在V3基础上的再训练）。
在初始研究组中，训练集（18.4%）和验证集（18.6%）的早产率相似。初始AI模型对早产预测的敏感性为39%，特异性为93%，AUC为0.757。在验证集中，AI对分娩天数的预测与实际情况相比，足月产的R值为0.90，自发早产加足月产的R值为0.88，仅自发早产的R值为0.48。通过图像分析评估，在预测分娩前天数时，MAE在所有孕期显示出相似的准确性。最后，使用额外图像并改进AI架构和训练方法进行再训练，提高了早产预测能力，所有分娩的R值从0.85（V1）提高到0.88（V3）再到0.92（V4）。对于自发早产，V4中的MAE为19.99天。
AI仅根据超声数据就能预测分娩时间。这项技术对早产的预测敏感性有限。使用有监督和无监督学习对AI进行再训练有可能进一步提高性能。
