An explainable machine learning framework for accurate prediction of postoperative anterior chamber depth in highly myopic cataract surgery

Retrospective Validation of an AI Based Tool for Vault Prediction After Implantable Collamer Lens Surgery for High Myopia

Purpose Our study evaluated a cohort of patients who previously underwent EVO Implantable Collamer Lens (ICL) implantation to assess the retrospective performance of the ICLGuru™ artificial intelligence (AI) sizing tool in predicting postoperative vault. Additionally, biometric factors associated with vault prediction error were identified. Methods This retrospective single-center study included 104 eyes from 56 patients who underwent EVO V4c or V5 ICL implantation between May 2023 and June 2025. Preoperative ultrasound biomicroscopy (UBM) images, from patients who had already undergone ICL surgery, were retrospectively analyzed using ICLGuru to generate predicted postoperative vaults. Actual postoperative vault was measured at 1-, 3-, or 6-month follow-up visits, with 3-month measurements preferentially included for analysis. Mixed-effects regression models were used to evaluate associations between biometric variables and vault prediction error. Results Mean absolute error was 135 ± 110 µm, indicating a tendency for ICLGuru to overestimate postoperative vault (p < 0.001). Among 16 eyes with actual postoperative overvault, ICLGuru retrospectively identified 15 cases in which a smaller lens size was predicted to achieve target vault, suggesting these outcomes may have been flagged. Out of 7 eyes with actual undervault, only two cases may have been flagged with use of ICLGuru, indicating that the tool is not as useful in preventing undervault. Mixed-effects regression demonstrated that greater iridocorneal angle and higher aRISE were significantly associated with lower vault prediction error (p < 0.05). Conclusion ICLGuru demonstrated useful postoperative vault predictions with performance comparable to existing machine learning and UBM-based nomograms, though it appeared to overestimate longer term vault in our cohort. The tool may be valuable in identifying eyes at risk for excessive vault. aRISE and iridocorneal angle were significant predictors of vault prediction error, supporting their importance in future ICL sizing models.

Predicting patient satisfaction with post-small incision lenticule extraction surgery using open-source machine learning software.

The Quest to Predict Surgically Induced Astigmatism After Cataract Surgery: Lessons for Toric IOL Planning

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.

An exploratory SMOTE-SVM approach for identifying preoperative biomechanical risk factors driving early toric intraocular lens micro-rotation in extremely imbalanced cohorts.

The predictive modeling of postoperative mechanical complications, such as the early micro-rotation of premium toric intraocular lenses (IOLs), is severely hindered by the extreme imbalance of clinical datasets. Traditional statistical methods often fail to capture complex biomechanical interactions in rare-event scenarios. This exploratory pilot study introduces a machine learning framework designed as a hypothesis-generating tool to handle extremely imbalanced ophthalmic data and identify potential preoperative biometric features associated with toric IOL micro-rotation. A prospective cohort of 35 eyes implanted with the Clareon PanOptix® Toric IOL was analyzed, quantifying true rotational stability via high-resolution photographic registration. Given the exceedingly low incidence of > 1-degree micro-rotation, a strict, leak-proof fivefold cross-validation pipeline was established. The Synthetic Minority Over-sampling Technique (SMOTE) was applied exclusively within the training folds, and an interpretable Linear Support Vector Machine (Linear SVM) was deployed to extract robust feature weights for biomechanical interpretation. Our findings highlight the "accuracy paradox" in small clinical datasets: complex ensemble models exhibited severe majority-class bias, failing to detect rare micro-rotations. Conversely, the SMOTE-enhanced Linear SVM achieved a Precision-Recall Area Under the Curve (PR-AUC) of 0.463, outperforming a random baseline by nearly a factor of three. The algorithmic feature weights successfully isolated Anterior Chamber Depth (ACD) and steep keratometry (Steep K2) as the primary geometric drivers of rotational instability, demonstrating a profound alignment with clinical ocular biomechanics. While strictly constrained by the small sample size (N = 35) and limited event rate, this preliminary pilot framework successfully bridges high-dimensional data augmentation with physical ocular biomechanics, effectively identifying minority risk features and laying the groundwork for future AI-driven surgical navigation systems.

A Risk Prediction Model for Normal-Tension Glaucoma Progression Integrating Genetic Markers and Corneal Biomechanics.

OBJECTIVE Normal-tension glaucoma (NTG) is characterized by progressive optic nerve damage despite intraocular pressure remaining consistently within the normal range. Predicting disease progression in patients with confirmed NTG remains challenging. This study aimed to develop and validate an interpretable machine learning model integrating genetic risk scores and corneal biomechanical parameters to predict progression risk in patients with NTG, identify independent predictors, and quantify the contribution of individual features to support precision clinical management. METHODS A total of 342 patients with NTG were consecutively enrolled at a tertiary hospital and randomly allocated to training set (n = 238) and validation set (n = 104) at a ratio of 7:3. Baseline characteristics and six core indicators were collected. Candidate predictors were selected through univariate analysis and least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation and the λ-1se criterion. Independent predictors were subsequently identified using multivariable logistic regression. Three machine learning models-random forest (RF), support vector machine (SVM), and logistic regression (LR)-were developed. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) analysis was performed to interpret feature contributions. RESULTS Univariate analysis revealed significant differences in all six indicators between the progression and non-progression groups (p < .05). Multivariable logistic regression further confirmed that all six indicators were independently associated with NTG progression (p < .05). The RF model demonstrated the best predictive performance, with an AUC of 0.760 (95% confidence interval (CI) : 0.678-0.842) in the training set and 0.747 (95% CI: 0.623-0.871) in the validation set. It outperformed both the SVM model (training AUC = 0.704; validation AUC = 0.694) and the LR model (training AUC = 0.742; validation AUC = 0.729). SHAP analysis ranked the features, in descending order of contribution, as mean retinal nerve fibre layer (RNFL) thickness, first applanation velocity, visual field mean deviation, relative tear GNAI1 level, polygenic risk score for NTG, and relative tear PRDX4 level. The calibration curves showed good agreement between predicted and observed probabilities, while DCA demonstrated a high clinical net benefit across a broad range of threshold probabilities. CONCLUSION A model integrating corneal biomechanical parameters, genetic risk scores, and tear molecular biomarkers was developed to predict the risk of NTG progression and demonstrated potential clinical utility. This model may provide a quantitative reference for risk stratification and personalised management in patients with NTG.

Preoperative and postoperative risk factors for persistent dry eye after corneal refractive surgery in ultra-high myopia: a retrospective cohort study.

OBJECTIVE To identify risk factors for persistent dry eye after refractive surgery in patients with ultra-high myopia and their prognostic value. METHODS Data from 260 patients were analyzed. Least absolute shrinkage and selection operator regression screened preoperative variables. Random forest, Extreme Gradient Boosting, multivariate logistic regression, mediation, and interaction analyses were used to identify predictors and mechanisms of persistent dry eye (≥ 3 months). RESULTS Persistent dry eye occurred in 33.1% of patients (86/260). Eight preoperative predictors were identified. Elevated Ocular Surface Disease Index (OSDI) (odds ratio [OR] = 1.483) and corneal fluorescein staining (CFS) (OR = 7.154) were independent risk factors; prolonged tear film breakup time (TBUT) (OR = 0.292), no systemic history (OR = 0.042), and normal meibomian gland dysfunction (MGD) (OR = 0.042) were protective. Age and Schirmer test were marginally significant. Medication adherence fully mediated the effect of age on dry eye duration, while ocular inflammation partly mediated the effect of photorefractive keratectomy surgery. A significant synergistic interaction was found between postoperative inflammation and medication adherence (P < 0.05). CONCLUSION Persistent dry eye after refractive surgery for ultra-high myopia is influenced by multiple factors. Preoperative OSDI, CFS, TBUT, systemic history, MGD, daily screen time, medication adherence, and inflammation play important roles. Individualized perioperative management targeting these factors may improve prognosis.