Introduction

Early pregnancy loss (EPL), defined as pregnancy loss before 12 weeks of gestation, remains a common yet under-recognized challenge in obstetrics, carrying significant emotional and medical implications for patients and healthcare providers alike. Despite its high prevalence, many cases occur without warning signs, limiting opportunities for timely intervention.

Current risk assessment methods—based on clinical history, biochemical markers, and conventional ultrasound interpretation—are constrained by subjectivity and the inability to capture subtle biological changes at early stages. Recent advances in artificial intelligence (AI) applied to ultrasound imaging offer a transformative opportunity to address these limitations. By extracting micro-level indicators from ultrasound videos that are imperceptible to the human eye, AI models have demonstrated the capacity to identify pregnancies at high risk of EPL well before conventional methods.

One example is a hybrid AI model that automatically measures biometric parameters such as gestational sac area, yolk sac diameter, crown–rump length, and fetal heart rate from early pregnancy ultrasound videos. This approach reduces inter-observer variability and achieves high predictive performance, with area under the curve (AUC) values approaching 0.97—surpassing traditional clinical strategies. Integrating such tools into prenatal care could enable earlier, more accurate risk stratification, enhance patient counseling, and improve clinical outcomes during the critical therapeutic window.

Why AI and Ultrasound Represent a Frontier in Obstetrics

AI combined with ultrasound represents a paradigm shift in EPL risk prediction, moving from reactive detection to proactive, precision-based monitoring. Machine learning enables the identification of complex, multidimensional patterns, while deep learning algorithms process high-volume image and video data to uncover clinically relevant features beyond human perception.

A key strength of AI in this setting is its ability to analyze pixel-level temporal changes in ultrasound video clips, detecting micro-morphological and hemodynamic variations such as subtle irregularities in the yolk sac or gestational sac, variations in trophoblastic texture, micro-fluctuations in embryonic heart rate variability, and nuanced changes in uteroplacental blood flow patterns. These early indicators, often undetectable in static images, can provide valuable prognostic insights.

Deep learning architectures—particularly convolutional neural networks (CNNs)—are used to capture spatial detail in ultrasound frames, while temporal dynamics are modeled through recurrent neural networks (RNNs) or long short-term memory (LSTM) networks. This sequential analysis enables detection of evolving changes in embryonic development and the uterine environment, improving prediction accuracy. High-resolution ultrasound recordings, Doppler flow measurements, and relevant maternal clinical data can be combined in multimodal models, further enhancing predictive capability.

AI Model Development and Training

Developing AI systems for EPL risk assessment involves a carefully designed pipeline, beginning with robust data preprocessing. Image normalization ensures consistency across datasets, denoising removes ultrasound artifacts, and optimal frame selection captures the most informative moments within video sequences.

CNNs serve as the primary method for spatial feature extraction, isolating critical anatomical details, while RNNs or LSTMs handle temporal pattern recognition, modeling the evolution of embryonic and placental features over time. Training datasets typically comprise retrospectively collected ultrasound videos from pregnancies with confirmed outcomes, with rigorous labeling protocols to differentiate viable from non-viable cases. Balanced datasets and augmentation techniques are applied to minimize prediction bias.

Validation is essential to ensure clinical applicability, with performance evaluated across multiple independent datasets from different institutions. Key metrics include sensitivity, specificity, AUC, and positive predictive value (PPV). In one study, a hybrid AI model employing CNN-based biometric extraction followed by ensemble learning achieved a precision of 98% and an AUC of 0.969 from a dataset of 630 ultrasound videos. Similar models combining deep learning and traditional machine learning methods have reported accuracies exceeding 95% and precision above 97%.

While most research to date focuses on retrospective cohorts, emerging prospective evaluations and integration into clinical workflows suggest that such systems could soon be deployed in real-world obstetric practice. The challenge lies in ensuring model generalizability across diverse patient populations, imaging equipment, and healthcare settings.

Reference:

  1. Liu L, Zang Y, Zheng H, Li S, Song Y, Feng X, Zhang X, Li Y, Cao L, Zhou G, Dong T, Huang Q, Pan T, Deng J, Cheng D. An AI method to predict pregnancy loss by extracting biological indicators from embryo ultrasound recordings in early pregnancy. Sci Rep. 2025 Jul 17;15(1):25946. doi: 10.1038/s41598-025-11518-5. PMID: 40676105; PMCID: PMC12271456.

  2. Yan S, Xiong F, Xin Y, Zhou Z and Liu W (2024) Automated assessment of endometrial receptivity for screening recurrent pregnancy loss risk using deep learning-enhanced ultrasound and clinical data. Front. Physiol. 15:1404418. doi: 10.3389/fphys.2024.1404418

  3. Liu L, Liu B, Wu H, Gan Q, Huang Q and Li M (2025) Optimizing predictive features using machine learning for early miscarriage risk following single vitrified-warmed blastocyst transfer. Front. Endocrinol. 16:1557667. doi: 10.3389/fendo.2025.1557667

  4. Yaseen I, Rather RA. A Theoretical Exploration of Artificial Intelligence's Impact on Feto-Maternal Health from Conception to Delivery. Int J Womens Health. 2024 May 22;16:903-915. doi: 10.2147/IJWH.S454127. PMID: 38800118; PMCID: PMC11128252.

  5. Yan S, Xiong F, Xin Y, Zhou Z and Liu W (2024) Automated assessment of endometrial receptivity for screening recurrent pregnancy loss risk using deep learning-enhanced ultrasound and clinical data. Front. Physiol. 15:1404418. doi: 10.3389/fphys.2024.1404418

  6. Aguado AM, Jimenez-Perez G, Chowdhury D, Prats-Valero J, Sánchez-Martínez S, Hoodbhoy Z, Mohsin S, Castellani R, Testa L, Crispi F, Bijnens B, Hasan B and Bernardino G (2024) AI-enabled workflow for automated classification and analysis of feto-placental Doppler images. Front. Digit. Health 6:1455767. doi: 10.3389/fdgth.2024.1455767

  7. Togunwa TO, Babatunde AO and Abdullah K-u-R (2023) Deep hybrid model for maternal health risk classification in pregnancy: synergy of ANN and random forest. Front. Artif. Intell. 6:1213436. doi: 10.3389/frai.2023.1213436

  8. Yaseen I, Rather RA. A Theoretical Exploration of Artificial Intelligence's Impact on Feto-Maternal Health from Conception to Delivery. Int J Womens Health. 2024 May 22;16:903-915. doi: 10.2147/IJWH.S454127. PMID: 38800118; PMCID: PMC11128252.

  9. Salomon LJ, Alfirevic Z, Da Silva Costa F, Deter RL, Figueras F, Ghi T, et al. ISUOG Practice guidelines: ultrasound assessment of fetal biometry and growth. Ultrasound Obstet Gynecol. (2019) 53(6):715–23. doi: 10.1002/uog.20272