Machine-learning prediction of high implantable collamer lens vault using angle-to-angle anatomy: a patient-grouped internal validation study
- 1Tianjin Key Laboratory of Retinal Functions and Diseases, Tianjin Branch of National Clinical Research Center for Ocular Disease, Eye Institute and School of Optometry, Tianjin Medical University Eye Hospital, Tianjin, China
- 2State Key Laboratory of Ophthalmology, Zhongshan Ophthalmic Center, Sun Yat-sen University, Guangzhou, China
Abstract
Background: Machine learning has increasingly been applied to implantable collamer lens (ICL) vault prediction, but the value of anatomically informed feature representation remains uncertain. We evaluated whether angle-to-angle (ATA)-centered anatomy improves prediction of high and continuous postoperative vault compared with white-to-white (WTW)-centered anatomy.
Methods: This retrospective study included 279 eyes from 152 patients with 1-week postoperative vault measurements. Matched regularized supervised machine-learning models were constructed using either ICL size minus WTW or ICL size minus ATA; an expanded ATA model additionally included crystalline lens rise and nasal and temporal anterior chamber angles. Models were internally validated using patient-grouped nested cross-validation, with fellow eyes kept together throughout validation. High vault (>750 µm) was the primary binary outcome, continuous vault was evaluated as a secondary outcome, and low vault (<250 µm) was exploratory. Confidence intervals and between-model differences were estimated using patient-cluster bootstrap sampling.
Results: High vault occurred in 77 eyes (27.6%) from 49 patients. The ATA-centered model achieved a higher AUC for high-vault prediction than the WTW-centered model (0.805 [95% CI, 0.738–0.864] vs. 0.743 [95% CI, 0.667–0.814]), with a paired difference of 0.063 (95% CI, 0.011–0.118). For continuous vault, the ATA-centered model had an MAE of 181.3 µm (95% CI, 163.0–201.1), 17.7 µm lower than the WTW-centered model (95% CI, 4.1–31.7). Adding the expanded ATA variables produced little further improvement. The ATA-centered model also showed good discrimination for low vault (AUC, 0.852), although only 18 patients had this outcome. In a separate paired one-eye analysis, the median within-selection ATA-minus-WTW AUC difference for high vault was 0.049.
Conclusion: ATA-based models improved prediction of both high and continuous 1-week ICL vault compared with otherwise matched WTW-based models. When an ICL size has already been selected, ATA-based prediction may help identify eyes that warrant closer review of biometric measurements and the proposed sizing plan.
Related
- Development and internal validation of a machine learning prediction model for postoperative hemorrhage following modified radical mastectomy for breast cancer
- Machine learning-based prediction of myocardial injury after hip replacement surgery: a single-center study with internal validation.
- Development and Patient-Level Internal Validation of an MRI-Based Radiomics Machine Learning Model for Preoperative Prediction of Cervical Lymph Node Metastasis in Oral Squamous Cell Carcinoma: A Pilot Study.
- Can maternal characteristics improve due date prediction without ultrasound? Development and internal validation of a machine-learning prediction model in 600 000 pregnancies in Sweden
- Different vault outcomes following implantable collamer lens implantation in eyes with ciliary body cysts: a case report.
BibTeX
@article{li2026machinelearning,
title = {{Machine-learning prediction of high implantable collamer lens vault using angle-to-angle anatomy: a patient-grouped internal validation study}},
author = {Hui Li and Runzhang He and Sheng Zhou and Jing Yang},
journal = {Frontiers in Cell and Developmental Biology},
year = {2026},
month = nov,
doi = {10.3389/fcell.2026.1970640},
url = {https://doi.org/10.3389/fcell.2026.1970640},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Machine-learning prediction of high implantable collamer lens vault using angle-to-angle anatomy: a patient-grouped internal validation study
AU - Hui Li
AU - Runzhang He
AU - Sheng Zhou
AU - Jing Yang
PY - 2026
DA - 2026/11/03
JO - Frontiers in Cell and Developmental Biology
DO - 10.3389/fcell.2026.1970640
UR - https://doi.org/10.3389/fcell.2026.1970640
AB - Background: Machine learning has increasingly been applied to implantable collamer lens (ICL) vault prediction, but the value of anatomically informed feature representation remains uncertain. We evaluated whether angle-to-angle (ATA)-centered anatomy improves prediction of high and continuous postoperative vault compared with white-to-white (WTW)-centered anatomy. Methods: This retrospective study included 279 eyes from 152 patients with 1-week postoperative vault measurements. Matched regularized supervised machine-learning models were constructed using either ICL size minus WTW or ICL size minus ATA; an expanded ATA model additionally included crystalline lens rise and nasal and temporal anterior chamber angles. Models were internally validated using patient-grouped nested cross-validation, with fellow eyes kept together throughout validation. High vault (>750 µm) was the primary binary outcome, continuous vault was evaluated as a secondary outcome, and low vault (<250 µm) was exploratory. Confidence intervals and between-model differences were estimated using patient-cluster bootstrap sampling. Results: High vault occurred in 77 eyes (27.6%) from 49 patients. The ATA-centered model achieved a higher AUC for high-vault prediction than the WTW-centered model (0.805 [95% CI, 0.738–0.864] vs. 0.743 [95% CI, 0.667–0.814]), with a paired difference of 0.063 (95% CI, 0.011–0.118). For continuous vault, the ATA-centered model had an MAE of 181.3 µm (95% CI, 163.0–201.1), 17.7 µm lower than the WTW-centered model (95% CI, 4.1–31.7). Adding the expanded ATA variables produced little further improvement. The ATA-centered model also showed good discrimination for low vault (AUC, 0.852), although only 18 patients had this outcome. In a separate paired one-eye analysis, the median within-selection ATA-minus-WTW AUC difference for high vault was 0.049. Conclusion: ATA-based models improved prediction of both high and continuous 1-week ICL vault compared with otherwise matched WTW-based models. When an ICL size has already been selected, ATA-based prediction may help identify eyes that warrant closer review of biometric measurements and the proposed sizing plan.
SN - 2296-634X
ER -