This content is only available for ASCRS Members
This content from the 2020 ASCRS Virtual Annual Meeting is only available to ASCRS members. To log in, click the teal "Login" button in the upper right-hand corner of this page.
Filed Under
Refractive
machine learning
artificial intelligence
LASIK
ectasia
2020 paper presentation
Purpose
To test and compare the classic percentage tissue altered (PTA) and the residual stromal bed (RSB), and to develop a novel data-driven method to enhance the characterization of the impact of LASIK procedure on the cornea.
Methods
Preoperative clinical data from 3,278 stable post-LASIK cases and from 105 cases that developed ectasia were analyzed. Thickness at the corneal center or apex (first Purkinje) and at the thinnest point were obtained from Pentacam HR (Oculus, Wetzlar, Germany). Clinical data included age at the time of surgery, manifest refraction, the maximal ablation depth, and flap thickness. A series of data preparation methods were used to handle and transform parameters using 20 fold cross-validation, for selecting the best features in different machine-learning artificial intelligence (AI) models.
Results
The RSB and PTA calculated for the thinnest value had a higher area under the receiver operating curve (AUC) than the of the RSB and PTA calculated for the central (apex) thickness values (DeLong, p<0.05). The RSB_min had the highest AUC (0.718; sensitivity 70.5% and specificity 64.6%; cut off 344µm) than the PTA_min (0.651; sensitivity 50.48% and specificity 73.76%; cut off 37%). All machine-learning parameters had higher accuracy. The best AI model was the random forest with an AUC of 0.922 (sensitivity 65.7% and specificity 94.7%; cut off 0.10%).
Conclusion
Artificial intelligence can be used to estimate the impact of LASIK when assessing ectasia risk prior surgery. Convergence of this approach with finite element models should be tested with external validation tests. Integration with RTA and patient-derived metrics as the Pentacam Random Forest Index (PRFI) should be considered for LASIK candidacy.
To test and compare the classic percentage tissue altered (PTA) and the residual stromal bed (RSB), and to develop a novel data-driven method to enhance the characterization of the impact of LASIK procedure on the cornea.
Methods
Preoperative clinical data from 3,278 stable post-LASIK cases and from 105 cases that developed ectasia were analyzed. Thickness at the corneal center or apex (first Purkinje) and at the thinnest point were obtained from Pentacam HR (Oculus, Wetzlar, Germany). Clinical data included age at the time of surgery, manifest refraction, the maximal ablation depth, and flap thickness. A series of data preparation methods were used to handle and transform parameters using 20 fold cross-validation, for selecting the best features in different machine-learning artificial intelligence (AI) models.
Results
The RSB and PTA calculated for the thinnest value had a higher area under the receiver operating curve (AUC) than the of the RSB and PTA calculated for the central (apex) thickness values (DeLong, p<0.05). The RSB_min had the highest AUC (0.718; sensitivity 70.5% and specificity 64.6%; cut off 344µm) than the PTA_min (0.651; sensitivity 50.48% and specificity 73.76%; cut off 37%). All machine-learning parameters had higher accuracy. The best AI model was the random forest with an AUC of 0.922 (sensitivity 65.7% and specificity 94.7%; cut off 0.10%).
Conclusion
Artificial intelligence can be used to estimate the impact of LASIK when assessing ectasia risk prior surgery. Convergence of this approach with finite element models should be tested with external validation tests. Integration with RTA and patient-derived metrics as the Pentacam Random Forest Index (PRFI) should be considered for LASIK candidacy.
View More Presentations from this Session
This presentation is from the session "SPS-114 Keratorefractive Surgical Planning" from the 2020 ASCRS Virtual Annual Meeting held on May 16-17, 2020.