Multimodal Deep Learning Integrating Day 3 Embryo Images and Clinical Data Predicts Pregnancy Outcomes After Fresh Single Embryo Transfer

Abstract This retrospective cohort study evaluated whether a multimodal deep learning model integrating Day 3 (D3) embryo images and clinical parameters could improve predictions of clinical pregnancy (CP), live birth (LB), and miscarriage following fresh single embryo transfer (SET), compared to unimodal approaches. A total of 1,207 fresh D3 SET cycles performed between January 2024 and August 2025 at a single center were analyzed. We developed an ImprovedFusion model combining a convolutional neural network for image feature extraction and a multilayer perceptron for clinical data, alongside unimodal benchmarks (ClinicalMLP and ImageResNet). Model performance was assessed using area under the receiver operating characteristic curve (AUC), sensitivity, and specificity, with SHAP and Grad-CAM providing interpretability. The CP rate was 44.32% (535/1,207), the LB rate was 35.82% (423/1,181; 26 lost to follow-up), and pregnancy loss occurred in 25.17% (151/600) of biochemical pregnancies. ImprovedFusion significantly outperformed both unimodal models across all endpoints. For CP, the multimodal model achieved an AUC of 0.754 (95% CI: 0.683–0.826; sensitivity: 0.889, specificity: 0.436), versus 0.689 and 0.592 for unimodal models. For LB, the AUC was 0.721 (95% CI: 0.647–0.795; sensitivity: 0.750), versus 0.665 and 0.626. For miscarriage, the AUC was 0.702 (95% CI: 0.576–0.828), versus 0.679 and 0.553. SHAP identified endometrial thickness, maternal age, trigger-day hormones, and image features as top predictors, while Grad-CAM highlighted greater attention to blastomere regions over fragmentation, suggesting cell number predicts better. Following external validation, this framework may offer a cost-effective, interpretable tool to enhance embryo selection and reproductive outcomes.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-12
DOI
https://doi.org/10.21203/rs.3.rs-10710514/v1
Primary Topic
Assisted Reproductive Technology and Twin Pregnancy
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Multimodal Deep Learning Integrating Day 3 Embryo Images and Clinical Data Predicts Pregnancy Outcomes After Fresh Single Embryo Transfer

Jialin Wu, Tingting Li, Xiaoyan Liang, Yanfang Wang et al.
Research Square
Assisted Reproductive Technology and Twin Pregnancy
preprint

Multimodal Deep Learning Integrating Day 3 Embryo Images and Clinical Data Predicts Pregnancy Outcomes After Fresh Single Embryo Transfer

Jialin Wu, Tingting Li, Xiaoyan Liang, Yanfang Wang, Linan Xu, Rui Huang
preprint en

Abstract

Abstract This retrospective cohort study evaluated whether a multimodal deep learning model integrating Day 3 (D3) embryo images and clinical parameters could improve predictions of clinical pregnancy (CP), live birth (LB), and miscarriage following fresh single embryo transfer (SET), compared to unimodal approaches. A total of 1,207 fresh D3 SET cycles performed between January 2024 and August 2025 at a single center were analyzed. We developed an ImprovedFusion model combining a convolutional neural network for image feature extraction and a multilayer perceptron for clinical data, alongside unimodal benchmarks (ClinicalMLP and ImageResNet). Model performance was assessed using area under the receiver operating characteristic curve (AUC), sensitivity, and specificity, with SHAP and Grad-CAM providing interpretability. The CP rate was 44.32% (535/1,207), the LB rate was 35.82% (423/1,181; 26 lost to follow-up), and pregnancy loss occurred in 25.17% (151/600) of biochemical pregnancies. ImprovedFusion significantly outperformed both unimodal models across all endpoints. For CP, the multimodal model achieved an AUC of 0.754 (95% CI: 0.683–0.826; sensitivity: 0.889, specificity: 0.436), versus 0.689 and 0.592 for unimodal models. For LB, the AUC was 0.721 (95% CI: 0.647–0.795; sensitivity: 0.750), versus 0.665 and 0.626. For miscarriage, the AUC was 0.702 (95% CI: 0.576–0.828), versus 0.679 and 0.553. SHAP identified endometrial thickness, maternal age, trigger-day hormones, and image features as top predictors, while Grad-CAM highlighted greater attention to blastomere regions over fragmentation, suggesting cell number predicts better. Following external validation, this framework may offer a cost-effective, interpretable tool to enhance embryo selection and reproductive outcomes.

Research Square
Sixth Affiliated Hospital of Sun Yat-sen University (CN)
Good health and well-being
Assisted Reproductive Technology and Twin Pregnancy
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon