Machine Learning Prediction Models for Classifying Myopic Versus Hyperopic Ablation Patterns in Eyes With Previous Refractive Surgery Undergoing Cataract Surgery.
Jul 2026· Journal of refractive surgery· Vol 42 7, pp.
e662-e667
· 0 citations· 13 references
Medicine
TL;DR
Machine learning-based classification models distinguished myopic from hyperopic ablation patterns, with the Random Forest model demonstrating the highest performance using corneal topographic data.
Abstract
Purpose
To develop a machine learning model to classify myopic or hyperopic ablation patterns in eyes that had previous refractive surgery and are now having cataract surgery using corneal topographic and tomographic parameters.
Methods
This retrospective observational study was conducted at the Instituto de Oftalmología Conde de Valenciana, Mexico City, Mexico. Machine learning analyses were performed using Orange Data Mining software (Biometrics Laboratory, University of Ljubljana, Slovenia) with a 70/30 training-testing split. Predictive variables included corneal asphericity (Q-value), posterior-to-anterior curvature ratio (P/A ratio), sagittal map morphology, spherical aberration (Z04), mean keratometry, central corneal thickness, and thinnest pachymetry. Random Forest, Decision Tree, and Logistic Regression models were trained and compared. Model performance was assessed using area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, F1 score, and Matthews correlation coefficient (MCC).
Results
Seventy-two eyes with a history of refractive surgery were analyzed: 16.67% (n = 12) had hyperopic and 83.33% (n = 60) had myopic ablations. P/A ratio and Z04 showed a strong negative correlation (r = -0.793, P < .001) for classifying the ablation type. The Random Forest model achieved the highest discriminative performance (AUC = 0.933), followed by Logistic Regression (AUC = 0.881) and Decision Tree (AUC = 0.828). Decision Tree modeling identified clinically interpretable cut-off values for classification, with a P/A ratio of 82.7% and a Z04 value of -0.144 emerging as key decision nodes for differentiating myopic from hyperopic ablation patterns.
Conclusions
Machine learning-based classification models distinguished myopic from hyperopic ablation patterns, with the Random Forest model demonstrating the highest performance using corneal topographic data. These models may serve as valuable adjuncts for intraocular lens power calculation in cataract surgery in eyes that had previous refractive surgery, especially when prior surgical records are unavailable.
This study successfully developed and validated an explainable RF-based machine learning model for the prediction of postoperative ACD in highly myopic cataract patients, which supports more reliable IOL power calculation and offers a practical tool for optimizing surgical planning in highly myopic eyes.
Yuyang Yang, Hao Cui, Jiajia Gao et al.· Frontiers in Cell and Develo...· 0 citations
Purpose
The aim of this study was to use machine learning to predict 12-months progression of maximum keratometry (Kmax) within stratified repeatability limits based on a baseline topography/optical coherence tomography measurement in patients with keratoconus (KC).
Methods
This study included patients with diagnosed KC in which baseline and follow-up measurements were performed with a Placido disc topographer combined with optical coherence tomography (OCT) at a time interval of at least 7 months. Eyes were classified as "progressive" or "stable" based on the repeatability-adjusted difference in Kmax between the baseline measurement and a follow-up measurement. Nested cross-validation was used to train different algorithms, and the algorithm with the best area under the curve (AUC) value was used as final model.
Results
A total of 102 eyes of 65 patients were included. Mean age was 33.2 ± 13.0 years. Predictors with high outer fold stability included specific KC markers such as the CSIB index and posterior K values. Within the internal folds, the model achieved an AUC of 0.82, a sensitivity of 62.2%, and a specificity of 83.2% and, in the outer left-out folds, a sensitivity of 65.6%, a specificity of 84.3%, and thus an accuracy of 78.4%.
Conclusions
Progression of Kmax can be predicted with decent accuracy using a single OCT-based topographic measurement. Nested cross-validation represents a valuable tool in small datasets.
Translational Relevance
Machine learning-based prediction of Kmax progression may aid in scheduling risk-adapted follow-up appointments, potentially leading to more rapid identification of progression-suspect KC cases.
A. Schlatter, L. Pomberger, A. Honeder et al.· Translational Vision Science...· 0 citations
Purpose To evaluate keratoconus (KC) risk factors and to develop a machine-learning (ML) model for KC and myopia classification. Methods In this retrospective single-center cross-sectional study, demographic and lifestyle data from patients with KC and individuals from a preoperative refractive surgery clinic were collected from January 20, 2024, to December 1, 2024. Univariable and multivariable regression analyses were used to identify key risk factors. Additionally, random forest (RF)-recursive feature elimination (RFE), extreme gradient boosting (XGBoost)-RFE, and univariable logistic regression were applied to select factors for ML models. Seven ML models were developed for a lifestyle-based classification system, with the performance being validated through discrimination and calibration, and interpretability being improved using SHapley Additive exPlanations (SHAP). Results Analysis of 711 patients (mean [standard deviation] age, 26.6 [7.1] years; 439 males [61.7%]) revealed 275 with KC. Multivariable regression analysis identified seven risk factors for KC, including male sex, higher body-mass index (BMI), lower education level, more distant childhood residence, allergic conjunctivitis, and increased eye-rubbing intensity and frequency. After feature selection of 24 variables, the neural-network model demonstrated the highest performance (area under the receiver operating characteristic curve [AUROC] = 0.79), followed by RF (AUROC = 0.77) and XGBoost (AUROC = 0.76). SHAP analysis consistently highlighted eye-rubbing intensity, sex, BMI, and childhood residence among the top 10 factors across the top three models, which were also confirmed by univariable logistic regression. Conclusion ML models can distinguish high-risk KC groups based on clinical risk factors, facilitating risk stratification and early lifestyle interventions.
Kaiyue Du, R. Peng, Yueguo Chen et al.· PLoS ONE· 0 citations
Purpose
To determine the prevalence of abnormal ectasia screening indices among refractive surgery candidates and whether autorefraction-keratometry (ARK) can discriminate patients with abnormal Pentacam indices well enough to triage tomographic referral.
Methods
In this retrospective, cross-sectional study, 1,985 of 2,000 consecutive refractive surgery candidates with complete ARK and Pentacam data were analyzed. Screening-positivity was defined at the patient level as a Belin-Ambrósio Deviation (BAD-D) > 1.6 or Topographic Keratoconus Classification (TKC) ≥ 0.5 in either eye, and a clearly-abnormal category as BAD-D > 2.6 or TKC ≥ 1.0. Pre-specified ARK models were assessed for discrimination by the area under the receiver operating characteristic curve (AUC) with cross-validation, then translated into operating characteristics at fixed sensitivities.
Results
Screening-positivity was present in 25.0% of patients (95% CI, 23.1-26.9), whereas clearly-abnormal indices occurred in only 2.4% (1.8-3.1); positivity was driven mainly by the borderline BAD-D > 1.6 flag. ARK keratometry was significantly associated with screening-positivity (steep keratometry odds ratio, 1.48 per diopter; P < 0.001) but discriminated only modestly (AUC, 0.72; 95% CI, 0.69-0.74), with no improvement from additional ARK parameters. At a sensitivity of 95%, only 16.2% of patients could safely forgo tomography, and 100% sensitivity spared just 1.9%.
Conclusions
Abnormal ectasia screening indices were common but predominantly borderline, and anterior keratometry alone could not reliably triage which corneas require tomographic examination. They support routine tomographic evaluation where available and indicate that keratometry alone cannot replace it where it is not.
Kyo Youn Jeong, Gyuwon Ryu, Moonwon Hwang· Korean Journal of Ophthalmol...· 0 citations
Abstract Purpose This study compared traditional statistical models with machine learning algorithms for predicting surgically induced astigmatism after cataract surgery, aiming to identify the most accurate and generalizable method among linear regression, regression trees, random forests, and neural networks. Methods Retrospective analysis was performed on 321 eyes (321 patients) undergoing phacoemulsification at a tertiary center. Data were standardized and surgically induced astigmatism was modeled using three vector-based outcomes: KEQ.post (keratometric equivalent power), KAST0.post (horizontal astigmatism component), and KAST45.post (oblique component). Preoperative variables served as predictors. Data were split into training (60%) and test (40%) sets. Four predictive models were developed and evaluated on unseen data. Performance was assessed using mean squared prediction error, and variable importance analyses identified key predictors. Results Linear regression achieved the best out-of-sample performance across all outcomes (e.g. KEQ.post mean squared prediction error = 0.043). Tree-based models performed slightly worse, while neural networks showed substantial overfitting with markedly higher test errors. Preoperative astigmatism and corneal radii were the strongest predictors. Conclusion KAST0 and KAST45 were not predictably modeled, whereas changes in KEQ.post were. Multivariable linear regression provided the most accurate and reliable predictions, while more complex machine-learning models (especially neural networks) overfit the limited dataset and offered no clinical advantage. The relationship between preoperative biometrics and postoperative equivalent power appears predominantly linear, though larger datasets may enhance machine-learning performance.
Amanda Pan, K. P. Kaiser, Stefan Raidl et al.· Current Eye Research· 0 citations
INTRODUCTION
A cross-sectional comparative device study to compare the Eyerobo Vision Screener (VS), a portable handheld photorefractor, against a conventional auto-refractor for machine-learning-based prediction of cycloplegic spherical equivalent (SE) and power vector components (J0, J45) in a pediatric myopia cohort, to evaluate whether supplementary IOL Master biometry narrows the performance gap between devices, and to report screening-oriented classification metrics (sensitivity, specificity, positive and negative predictive values, area under the receiver operating characteristic curve [AUC]) at clinically meaningful referral thresholds that quantify the device-model combination's triage performance.
METHODS
Data from 1129 eyes of 574 patients aged 6-18 years were collected using three ophthalmic devices before and after pharmacological dilation. All refractive measurements were decomposed into power vectors (SE, J0, J45) following Thibos et al. Twelve clinically motivated feature scenarios were constructed from a symmetric framework comparing the Eyerobo VS and auto-refractor (each under predilation, post-dilation, and combined conditions), with and without IOL Master biometry. TabPFN, a tabular foundation model requiring no hyperparameter tuning, was evaluated alongside seven conventional machine-learning algorithms using patient-level five-fold grouped cross-validation repeated over three random seeds. Agreement was assessed using Bland-Altman analysis, Pearson correlation, and clinical threshold analysis. In addition to regression metrics, a screening-classification analysis was performed at three prespecified clinical thresholds (cycloplegic SE ≤ - 0.50 D, ≤ - 3.00 D, and ≤ - 6.00 D), reporting sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and AUC with 95% paired bootstrap confidence intervals. Formal paired statistical comparison between TabPFN and the next-best algorithm and a one-eye-per-patient sensitivity analysis are also reported.
RESULTS
TabPFN achieved the lowest mean absolute error (MAE) on all Eyerobo scenarios and was competitive with Ridge on auto-refractor scenarios; on the headline Eyerobo Pre + IOL scenario, the per-eye paired difference in MAE between TabPFN (0.425 D) and the next-best gradient-boosting learner (XGBoost, 0.542 D) was 0.117 D in favour of TabPFN, with paired Wilcoxon signed-rank p < 0.001 . For predilation SE prediction, the auto-refractor achieved an MAE of 0.295 D (87.4% within ± 0.50 D), compared with 0.657 D (61.6%) for the Eyerobo VS. Adding IOL Master biometry to the Eyerobo reduced this gap by 64%, yielding an MAE of 0.425 D (72.0%). For astigmatic power vector components, the gap narrowed further: J0 MAE was 0.106 D (auto-refractor) versus 0.143 D (Eyerobo + IOL), and J45 MAE was 0.060 D versus 0.088 D. Auto-refractor post-dilation predictions approached cycloplegic accuracy (MAE = 0.203 D, 96.0% within ± 0.50 D). Subgroup analysis showed the Eyerobo performed best for mild myopia (MAE = 0.398 D, 75.6% within ± 0.50 D) and degraded for high myopia and the small nonmyopic stratum. In the screening-classification analysis, the Eyerobo and the auto-refractor were operationally equivalent at the any-myopia threshold (Eyerobo AUC 0.964 [95% CI 0.939-0.982], sensitivity 0.946, specificity 0.894; auto-refractor AUC 0.969 [0.948-0.986], sensitivity 0.960, specificity 0.886; paired AUC difference 0.005 [95% CI - 0.022 to + 0.034 ], formally noninferior under a prespecified δ = 0.05 margin).
CONCLUSIONS
Within this single-center pediatric referral cohort, the Eyerobo VS combined with machine learning produced cycloplegic SE estimates of a precision consistent with use as a triage adjunct to identify children who should proceed to a full cycloplegic examination, with near-equivalent performance to the auto-refractor for astigmatic components. The addition of IOL Master biometry narrowed the device gap by 64%. TabPFN achieved the best performance among the evaluated algorithms within this cohort and requires no hyperparameter tuning, but the modest margin over linear and gradient-boosting baselines and the absence of external validation preclude any recommendation as the definitive algorithm of choice. The proposed approach is not a substitute for cycloplegic refraction; prospective external validation in a general pediatric screening cohort is required before clinical deployment. Video abstract. See video abstract at https://youtu.be/wy7_Fw5Dnw4 or in the online or HTML version of the manuscript. (MP4 4546 KB).
Meng-Meng Xia, Bo-Xue Yao, Yu-Mei Wang et al.· Ophthalmology and Therapy· 0 citations