AI-derived temporal workflow metrics from laparoscopic distal gastrectomy videos: model development, external validation, and exploratory clinical concordance analysis
Intelligent Translational Medicine · 2026-11 · 100004
DOI: 10.1016/j.itm.2026.100004 ↗Kurz gefasst
Fachartikel in der Zeitschrift Intelligent Translational Medicine, 2026-11. Verfasst von Ruiwen Cai und 17 weiteren Autor:innen. Die Arbeit nennt 47 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Natural Science Foundation of China, Jiangxi Provincial Natural Science Foundation.
Abstract (Originalsprache)
Background
It remains unclear whether phase durations estimated by artificial intelligence (AI) in laparoscopic distal gastrectomy (LDG) agree with manual measurements and show similar associations with clinical variables.
Methods
This retrospective three-center study included 64 surgical videos: 50 internal and 14 external. A Hiera-based model was evaluated using nine operative phases and a 12-class scheme incorporating infrapyloric dissection substeps. AI-derived and manual durations were compared in 24 matched cases using intraclass correlation coefficients (ICCs), Bland–Altman analysis, and mean absolute error. Clinical associations were explored using available matched data from the 58-case clinical cohort, with AI–manual concordance assessed in 18 cases.
Results
Internal test and external validation accuracies were 0.943 and 0.888 for the 9-phase scheme, and 0.933 and 0.873 for the 12-class scheme, respectively. External macro F1-scores were 0.863 and 0.841, respectively. Pooled duration ICCs were 0.865 and 0.824, with mean absolute errors of 5.20 and 4.82 min, respectively. Agreement varied by phase and was weak for duodenal and gastric transection. Manual durations of lymph node dissection phases were positively associated with estimated blood loss and body mass index. In 18 matched cases, AI-derived and manual clinical correlation coefficients showed a Pearson correlation of 0.835 and directional agreement of 86.3%. AI-derived durations recovered 50 of 59 manual association signals (84.7%).
Conclusions
AI-derived phase durations showed overall agreement with manual measurements and retained exploratory clinical association patterns. These findings support their potential as scalable quantitative descriptors for LDG surgical process research, while highlighting phase-specific measurement limitations.
Förderung
- National Natural Science Foundation of China
- Jiangxi Provincial Natural Science Foundation
Zitieren
APA
Cai, R., Liang, W., Wang, X., Lu, Y., Luo, X., Lei, Q., Yu, X., Fang, C., Zhong, W., Zhao, M., Li, X., Yang, C., Ming, W., Luo, S., Xue, Q., Zhong, S., Li, G., & Chen, T. (2026). AI-derived temporal workflow metrics from laparoscopic distal gastrectomy videos: model development, external validation, and exploratory clinical concordance analysis. Intelligent Translational Medicine, 100004. https://doi.org/10.1016/j.itm.2026.100004
Vancouver
Cai R, Liang W, Wang X, Lu Y, Luo X, Lei Q, et al. AI-derived temporal workflow metrics from laparoscopic distal gastrectomy videos: model development, external validation, and exploratory clinical concordance analysis. Intelligent Translational Medicine. 2026:100004. doi:10.1016/j.itm.2026.100004
BibTeX
@article{Cai2026,
author = {Cai, Ruiwen and Liang, Weiqi and Wang, Xi and Lu, Yuheng and Luo, Xiao and Lei, Qi and Yu, Xiaolong and Fang, Chuanfa and Zhong, Wu and Zhao, Mingli and Li, Xunjun and Yang, Carl and Ming, Wai-Kit and Luo, Shengwei and Xue, Qi and Zhong, Shizhen and Li, Guoxin and Chen, Tao},
title = {AI-derived temporal workflow metrics from laparoscopic distal gastrectomy videos: model development, external validation, and exploratory clinical concordance analysis},
journal = {Intelligent Translational Medicine},
year = {2026},
pages = {100004},
doi = {10.1016/j.itm.2026.100004}
}
EUVIMED