A deep learning model for automated assessment of neonates at risk of impaired breast crawl: a proof-of-concept video-based study
- 1Department of Nursing, West China Second University Hospital, Sichuan University
- 2West China School of Nursing, Sichuan University
- 3Key Laboratory of Birth Defects and Related Diseases of Women and Children (Sichuan University), Ministry of Education
Abstract
Background: Timely initiation of breastfeeding is critical for newborn health. The breast crawl, an instinctive sequence of behaviors, is fundamental to this process. However, newborns vary considerably in their ability to complete the crawl, and current assessment relies on subjective endpoint observation. This approach cannot provide an early, objective measure of crawl competency or support timely intervention for at-risk mother-infant dyads.
Objective: This study aimed to develop and validate a deep learning video analysis model for the automated, graded assessment of breast crawl competency and the identification of neonates at risk of impairment.
Methods: In this prospective observational study, we recruited 114 full-term mother-newborn dyads from a tertiary hospital. Breast crawl sessions were video recorded, and 30 neonatal behaviors were annotated using an extended Widström framework. Because the model operates on the complete recorded episode (up to first suckling), it performs retrospective characterization of crawl competency rather than prospective prediction. We developed a two-stage framework. First, a three-dimensional convolutional neural network (3D-CNN) recognized 30 neonatal behaviors from multichannel video (RGB, pose, edge, and depth) to produce an action timeline. Second, 75 behavioral descriptors derived from this timeline were combined with 65 perinatal clinical variables in a fully connected network to predict a five-level competency grade. Shapley Additive Explanations (SHAP) was used to rank feature contributions.
Results: Among 114 newborns (mean birth weight 3,185 g), the model showed moderate overall discriminative ability. It identified high-risk infants with F1-scores of 0.82 for Grade 3 (crawling without reaching the nipple) and 0.67 for Grade 4 (no crawling). Performance for the rare Grade 2 (reaching without attempting to suckle) was inconclusive because of the small sample ( n = 3). On the held-out test set ( n = 20), the model achieved a macro-average AUC of 0.77 and a micro-average AUC of 0.81. In a post-hoc binary analysis aligned with the intended clinical use (Grades 3–4 vs. 0–2), sensitivity was 1.00 (95% CI 0.69–1.00), specificity 0.70 (0.35–0.93), and accuracy 0.85 (0.62–0.97). Confidence intervals were wide throughout, reflecting the small test set. Grade 2 ( n = 3) could not be evaluated. SHAP analysis identified an earlier onset of hand-to-perioral activity and a shorter interval between maternal pethidine administration and delivery as among the most salient features associated with better and poorer crawl competency, respectively.
Conclusion: This proof-of-concept study suggests that deep learning video analysis may provide an objective adjunct for identifying neonates at risk of impaired breast crawl. However, the estimates are derived from a single center with a small test set and a non-nested feature selection procedure, and should be regarded as preliminary and hypothesis-generating. External validation with a prespecified nested analysis plan is required before any clinical application.
Related
- Early Detection of Mortality Risk in Preterm Neonates Using a Random Forest Classifier and Tree-Based Pipeline Optimization Tool Automated Machine Learning
- Code-Free Automated Machine-Learning (AutoML) Classification of Curated MRI Images Labeled for Rare Neurological Diseases: An Image-Level Proof-of-Concept Study
- Feasibility of a Deep-Learning Model for Objective Visualization-Quality Assessment During Cholangioscopy-Guided Lithotripsy: A Proof-of-Concept Study.
- Multihipposenet: multi-task deep learning for automated DDH assessment in infant hip ultrasound
- A novel morphological assessment approach for abdominal aortic aneurysms: automated segmentation and measurement using a semi-supervised deep learning algorithm
BibTeX
@article{xiang2026deep,
title = {{A deep learning model for automated assessment of neonates at risk of impaired breast crawl: a proof-of-concept video-based study}},
author = {Zhu Xiang and Yunxuan Tong and Yuqi Chen and Anqi Xiong and Xinyu Shi and Biru Luo},
journal = {Frontiers in Public Health},
year = {2026},
month = oct,
doi = {10.3389/fpubh.2026.1908206},
url = {https://www.frontiersin.org/articles/10.3389/fpubh.2026.1908206/full},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - A deep learning model for automated assessment of neonates at risk of impaired breast crawl: a proof-of-concept video-based study
AU - Zhu Xiang
AU - Yunxuan Tong
AU - Yuqi Chen
AU - Anqi Xiong
AU - Xinyu Shi
AU - Biru Luo
PY - 2026
DA - 2026/10/23
JO - Frontiers in Public Health
DO - 10.3389/fpubh.2026.1908206
UR - https://www.frontiersin.org/articles/10.3389/fpubh.2026.1908206/full
AB - Background: Timely initiation of breastfeeding is critical for newborn health. The breast crawl, an instinctive sequence of behaviors, is fundamental to this process. However, newborns vary considerably in their ability to complete the crawl, and current assessment relies on subjective endpoint observation. This approach cannot provide an early, objective measure of crawl competency or support timely intervention for at-risk mother-infant dyads. Objective: This study aimed to develop and validate a deep learning video analysis model for the automated, graded assessment of breast crawl competency and the identification of neonates at risk of impairment. Methods: In this prospective observational study, we recruited 114 full-term mother-newborn dyads from a tertiary hospital. Breast crawl sessions were video recorded, and 30 neonatal behaviors were annotated using an extended Widström framework. Because the model operates on the complete recorded episode (up to first suckling), it performs retrospective characterization of crawl competency rather than prospective prediction. We developed a two-stage framework. First, a three-dimensional convolutional neural network (3D-CNN) recognized 30 neonatal behaviors from multichannel video (RGB, pose, edge, and depth) to produce an action timeline. Second, 75 behavioral descriptors derived from this timeline were combined with 65 perinatal clinical variables in a fully connected network to predict a five-level competency grade. Shapley Additive Explanations (SHAP) was used to rank feature contributions. Results: Among 114 newborns (mean birth weight 3,185 g), the model showed moderate overall discriminative ability. It identified high-risk infants with F1-scores of 0.82 for Grade 3 (crawling without reaching the nipple) and 0.67 for Grade 4 (no crawling). Performance for the rare Grade 2 (reaching without attempting to suckle) was inconclusive because of the small sample ( n = 3). On the held-out test set ( n = 20), the model achieved a macro-average AUC of 0.77 and a micro-average AUC of 0.81. In a post-hoc binary analysis aligned with the intended clinical use (Grades 3–4 vs. 0–2), sensitivity was 1.00 (95% CI 0.69–1.00), specificity 0.70 (0.35–0.93), and accuracy 0.85 (0.62–0.97). Confidence intervals were wide throughout, reflecting the small test set. Grade 2 ( n = 3) could not be evaluated. SHAP analysis identified an earlier onset of hand-to-perioral activity and a shorter interval between maternal pethidine administration and delivery as among the most salient features associated with better and poorer crawl competency, respectively. Conclusion: This proof-of-concept study suggests that deep learning video analysis may provide an objective adjunct for identifying neonates at risk of impaired breast crawl. However, the estimates are derived from a single center with a small test set and a non-nested feature selection procedure, and should be regarded as preliminary and hypothesis-generating. External validation with a prespecified nested analysis plan is required before any clinical application.
SN - 2296-2565
ER -