Contribution of Corneal Biomechanics to Machine Learning-Based Refractive Procedure Selection: A Retrospective Observational Study

Contribution of Corneal Biomechanics to Machine Learning-Based Refractive Procedure Selection: A Retrospective Observational Study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Contribution of Corneal Biomechanics to Machine Learning-Based Refractive Procedure Selection: A Retrospective Observational Study Yinhao Li, Gang Li, Chuanyun Xu, Jiahui Wang, Xin Wang, Yanli Peng This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9362570/v1 This work is licensed under a CC BY 4.0 License Status: Under Revision Version 1 posted 14 You are reading this latest preprint version Abstract Background: Corneal biomechanical parameters are widely used in preoperative screening for refractive surgery, primarily to identify contraindications and ensure surgical safety. However, their potential role in guiding procedure selection has not been fully explored. This study aimed to evaluate the contribution of corneal biomechanical parameters beyond conventional refractive and tomographic features to refractive procedure selection. Methods: We conducted a retrospective observational study of 395 patients (763 eyes) who underwent refractive surgery at Chongqing Aier Eye Hospital between October 2023 and November 2024. Preoperative data included 48 features comprising biomechanical indices, tomographic parameters, and demographic characteristics. Multiple machine learning models have been developed to predict surgical procedure selection, and their performance has been evaluated in terms of accuracy, recall, and macro-F1 score. Ablation experiments were performed to assess the contributions of corneal biomechanical parameters. Model interpretability was analyzed via Shapley additive explanations (SHAPs). Statistical analyses were conducted via the Shapiro–Wilk test, Kruskal–Wallis test, and Mann–Whitney U test with Bonferroni correction. Results: The two-stage XGBoost-based model demonstrated the best performance, achieving a validation accuracy of 86.08% and an accuracy of 82.67% on the independent test set. The inclusion of corneal biomechanical parameters improved model performance, resulting in a 4% increase in accuracy on the independent test set. SHAP analysis revealed that refractive parameters contributed most prominently to model predictions, whereas corneal biomechanical features exhibited additional contributions across multiple procedures. Conclusions: An interpretable machine learning framework can support refractive procedure selection across multiple surgical options. Corneal biomechanical parameters provide complementary, nonredundant information beyond conventional features and may enhance AI-assisted clinical decision-making. Refractive surgery Corneal biomechanics Machine learning Surgical decision support Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Background Refractive surgery is widely used to correct refractive errors and restore visual function, with current procedures broadly classified into corneal refractive surgery and phakic intraocular lens implantation, most commonly the implantable collamer lens (ICL) 1 . Corneal refractive procedures, including Keratorefractive lenticule extraction (KLEx), Laser-Assisted In Situ Keratomileusis (LASIK), Laser-Assisted Subepithelial Keratectomy (LASEK), Photorefractive Keratectomy (PRK), and Transepithelial PRK (Trans-PRK), reshape the anterior cornea and are generally safe and predictable, whereas ICL implantation preserves corneal tissue and is often preferred in patients with high myopia, thin corneas, or suspected ectatic risk 2 . Despite favorable safety profiles, inappropriate procedure selection may still result in vision-threatening complications, such as post-operative corneal ectasia following corneal refractive surgery or vault-related anterior segment complications after ICL implantation 3 , 4 . Therefore, comprehensive preoperative evaluation incorporating corneal thickness, tomographic patterns, biomechanical characteristics, anterior chamber anatomy, endothelial status, and patient-specific factors is essential for minimizing surgical risk and optimizing refractive outcomes 5 . With the rapid development of artificial intelligence (AI) and machine learning (ML), these techniques have shown substantial promise in preoperative screening, identifying contraindications, and predicting postoperative outcomes in refractive surgery 6 – 12 . Yoo et al. 6 developed multimodal ML classifiers incorporating demographic data, Pentacam tomography, and questionnaire information, achieving excellent AUCs (0.983/0.972) for identifying contraindications. Their subsequent four-class models (LASEK, LASIK, KLEx, and contraindication) achieved accuracies of 81.0% and 78.9%, respectively 8 . However, both studies lacked corneal biomechanical inputs, which limited the model's ability to capture biomechanically relevant information that may contribute to individualized risk assessment. Li et al. 11 reported 87.75% agreement with surgeons using long-term historical data, but only indirect biomechanical indicators (SRI, SAI, CSI) were available. Overall, existing ML models rely primarily on refractive and tomographic features and lack comprehensive incorporation of corneal biomechanical parameters. In recent years, corneal biomechanics has become an increasingly important dimension in refractive surgery safety assessment, gaining greater prominence in preoperative screening. The Corvis ST, which is based on dynamic Scheimpflug imaging, quantifies corneal deformation under an air‒puff load, and its material-related parameters and composite risk indices have been progressively used to identify subclinical keratoconus and biomechanically vulnerable corneas 13 – 19 . Moreover, multiple studies have demonstrated that KLEx, LASIK, and surface ablation procedures result in systematic differences in postoperative biomechanical weakening due to variations in ablation depth, flap creation, and tissue-removal patterns 20 – 26 . From the perspective of residual biomechanical stability, these disparities suggest that different procedures may correspond to different acceptable preoperative risk thresholds: for corneas in a biomechanically borderline state, certain procedures may warrant greater caution or even avoidance, whereas relatively "conservative" procedures may still remain within a safe range. Accordingly, incorporating biomechanical information into procedure selection may provide more comprehensive insights for surgical planning beyond its traditional role in contraindication screening. However, its role in procedure selection remains underexplored. To address these limitations, we developed a two-stage surgical recommendation model that integrates Corvis ST biomechanical parameters, Pentacam tomographic features, and demographic information to jointly evaluate ICL, KLEx, LASIK, and surface ablation procedures (including PRK, Trans-PRK, and LASEK, hereafter referred to as SURFACE). This unified, biomechanics-informed framework learns clinically relevant decision logic and provides interpretability through SHAP-based analyses. This study aims to evaluate the performance of this framework and to quantify the contribution of corneal biomechanical parameters to refractive procedure selection beyond conventional refractive and tomographic features. Methods Patients This single-center retrospective study included patients who underwent refractive surgery at the Refractive Surgery Department of Chongqing Aier Eye Hospital between October 2023 and November 2024. Only patients whose preoperative records were complete were eligible. The type of refractive procedure was determined jointly by experienced refractive surgeons on the basis of clinical indications and patient preference. All the data were anonymized prior to analysis to remove personal identifiers. The study was approved by the Institutional Review Board of Aier Eye Hospital (Approval No. 202401007) and adhered to the tenets of the Declaration of Helsinki. The study involved four commonly performed refractive procedures: ICL and three types of corneal laser surgery, namely, the KLEx, LASIK, and SURFACE procedures, where KLEx specifically denotes small-incision lenticule extraction (SMILE). The inclusion and exclusion criteria were established with reference to published guidelines and expert consensus statements, previous literature 27 , 28 , and the clinical experience of senior refractive surgeons. Eligible patients were 18 to 45 years of age; had completed comprehensive ophthalmic examinations and preoperative assessments; and underwent ICL, SMILE, LASIK, or SURFACE with complete clinical documentation. Eyes with keratoconus or other ectatic corneal disease, visually significant cataracts or glaucoma, active ocular infection or inflammation, severe adnexal abnormalities, retinal pathology affecting visual function, systemic or metabolic disease likely to affect ocular status, or psychiatric conditions impairing cooperation were excluded. Preoperative examinations and surgery All enrolled patients underwent standardized preoperative screening. Examinations included Pentacam HR (OCULUS) for corneal tomography and anterior segment evaluation; Corvis ST (OCULUS) for corneal biomechanical assessment; ultrasound biomicroscopy (UBM; SW-3200L) for ciliary body and anterior lens surface evaluation; and anterior segment OCT (AS-OCT; Carl Zeiss Meditec) for corneal thickness and anterior chamber depth measurements. UBM and AS-OCT were used solely for clinical safety assessment, and their quantitative parameters were not included as model features. Routine examinations included visual acuity, refraction, intraocular pressure, and systemic evaluation. Surgical decision-making included corneal thickness, corneal shape, residual stromal bed thickness, anterior chamber depth, and refractive stability, in accordance with current safety standards for refractive surgery. Data collection A total of 395 patients were included in this study. The preoperative parameters consisted of 48 features encompassing corneal biomechanical properties, Pentacam tomographic data, and demographic characteristics. A full feature list is provided in Table S1, an aggregated overview of feature categories is presented in Table 1 , and definitions of biomechanical parameters are provided in Table S2. This dataset supports multidimensional clinical feature analysis and was used for subsequent model development and validation. Table 1 Categorization of preoperative variables included in the refractive surgery decision-making model. Type Feature Names Amount Corneal Biomechanical Properties bIOP, PRFI, SSI, SP-A1, IR, ARTh, DA-Ratio, CBI, BAD-D, TBI 10 Pentacam Corneal Topography Data Front K1, Front K2, Front K mean, Front K1 Axis, Back K1, Back K2, Back K mean, Back Axis K1, TCT, CV, ACV, ACA, ACD, Front K max, Q Value, Anterior Surface Elevation at the Corneal Thinnest Point, Posterior Surface Elevation at the Corneal Thinnest Point, Df, Db, Dp, Dt, Da, Pupil Center X, Pupil Center Y, Angle Kappa, CD, IS Value, Average Pachymetric Progression Rate 28 Demographics and Baseline Characteristics Gender, Age, Ni-BUT, SPH, CYL, Axis, SE, UDVA, CDVA, Dark Pupil Diameter 10 Model Design To compare different strategies for refractive surgery classification via real-world clinical data, three machine learning frameworks have been developed (Fig. 1 ). Model A performs direct four-class classification of ICL, SMILE, LASIK, and SURFACE. Model B uses a two-stage hierarchical structure: Stage 1 distinguishes ICL from corneal refractive surgery; Stage 2 classifies SMILE, LASIK, and SURFACE. Model C extends this hierarchy by adding a third stage: Stage 1 separates ICL from corneal procedures; Stage 2 differentiates SURFACE from stromal procedures; and Stage 3 performs binary classification between SMILE and LASIK. Training Setup Data preprocessing began with the removal of missing values attributable to measurement or acquisition errors. Clinically plausible missing entries were imputed according to standard ophthalmic assessment criteria. The outliers were screened via Z-scores, and samples with any feature exceeding |Z| > 5 were excluded. After filtering, 742 eyes remained for model development. All numeric features were standardized via Z-score normalization, and categorical variables were encoded via one-hot encoding. Redundant or low-variance features were removed via Pearson correlation analysis (threshold = 0.95) and variance filtering (threshold = 0.015), resulting in 38 features for Models A and B. In Model C, the third-stage classifier excluded the PRFI due to its limited discriminatory value, retaining 37 features. Six supervised learning algorithms were evaluated: decision tree (DT), random forest (RF), logistic regression (LR), support vector machine (SVM), multilayer perceptron (MLP), and eXtreme gradient boosting (XGBoost). The data were split at the patient level at a 9:1 train–test ratio. Fivefold cross-validation was performed on the training set via patient-based grouping to prevent data leakage between the eyes of the same patient. To address class imbalance, SMOTE oversampling was applied within each training fold, whereas the validation and testing sets remained unchanged. The hyperparameters were tuned via a grid search. All the experiments were implemented in Python 3.10 via scikit-learn 1.2 and XGBoost 1.7 on an NVIDIA RTX 3070 platform. Evaluation Metrics Model performance was assessed via accuracy, recall, and macro-F1, which capture overall predictive correctness, sensitivity to minority classes, and balanced multiclass performance, respectively. The definitions were as follows: $$\:\text{Accuracy}\text{=}\frac{\text{TP}\text{+}\text{TN}}{\text{TP}\text{+}\text{TN}\text{+}\text{FP}\text{+}\text{FN}}\text{,}$$ $$\:\text{Recall}\text{=}\frac{\text{TP}}{\text{TP}\text{+}\text{FN}}\text{,}$$ $$\:\text{Precision}\text{=}\frac{\text{TP}}{\text{TP}\text{+}\text{FN}}\text{,}$$ $$\:\text{F}\text{1=}\frac{\text{2}\text{×}\text{Precision}\text{×}\text{Recall}}{\text{Precision}\text{+}\text{Recall}}\text{,}$$ where \(\:\text{TP}\) = true positives, \(\:\text{TN}\) = true negatives, \(\:\text{FP}\) = false positives, and \(\:\text{FN}\text{}\) = false negatives. Macro-F1 is particularly informative under imbalanced class distributions. For validation in multistage models, a prediction was considered correct only if the sample passed all decision stages along its true clinical pathway. Therefore, overall validation accuracy was computed via weighted multiplicative aggregation: where \(\:\text{n}\) represents the number of surgical categories, \(\:{\text{p}}_{\text{i}}\) represents the proportion of class \(\:\text{i}\) in the true dataset, \(\:{\text{a}}_{\text{i}\text{j}}\) represents the validation accuracy of class \(\:\text{}\text{i}\) at decision stage \(\:\text{j}\) , and \(\:{\text{m}}_{\text{i}}\text{}\) represents the number of stages required by class \(\:\text{i}\) . In this study, SMILE and LASIK share the same decision pathway; thus, \(\:\text{i}\) =1, 2, and 3 correspond to the ICL, stromal (SMILE & LASIK), and SURFACE procedures, respectively. Statistical analysis The Shapiro–Wilk test was used to assess the normality of all variables. Normally distributed variables are reported as the means ± standard deviations, whereas nonnormally distributed variables are summarized as medians with interquartile ranges. Because several variables violated normality assumptions, all intergroup comparisons were performed via the Kruskal–Wallis test, ensuring methodological consistency and robustness. To further examine the contribution of corneal biomechanics to surgical decision-making, ten clinically established tomographic and biomechanical parameters were analyzed across surgical groups. For parameters demonstrating overall statistical significance, pairwise group comparisons were conducted via the Mann–Whitney U test, with Bonferroni correction applied to adjust for multiple comparisons. Results A total of 763 eyes of 395 patients were included in this study, comprising bilateral cases with identical procedures, bilateral cases with different procedures, and unilateral cases (27 single-eye records). Baseline characteristics Normality testing revealed that several variables were not normally distributed; therefore, all intergroup comparisons were conducted via the Kruskal–Wallis test, with detailed results provided in Table S3. To examine distributional differences across surgical groups, ten widely used tomographic and biomechanical parameters were compared (Table 2 ). Most biomechanical parameters, including the PRFI, SSI, SP-A1, IR, ARTh, DA-Ratio, CBI, BAD-D, and TBI, were significantly different among the four surgical procedures (p < 0.05). Parameters that demonstrated overall significance were subsequently subjected to Bonferroni-adjusted Mann–Whitney U tests, and pairwise comparison results are reported in Fig. 2 . The pairwise results revealed that across surgical types, all dynamic biomechanical parameters except bIOP demonstrated significant variability. Table 2 Distribution and statistical comparison of corneal biomechanical metrics across the four refractive surgery groups. Biomechanical indices derived from Corvis ST are summarized for the ICL, SMILE, SURFACE, and LASIK procedures. S.W. represents the Shapiro–Wilk test, which assesses variable distribution, and K.W. represents the Kruskal–Wallis test, which is used to determine intergroup differences. Statistical significance was defined as P < 0.05. Feature Name P(S.W.) ICL SMILE SURFACE LASIK P (K.W.) bIOP 0.1979 17.30 ± 1.54 17.19 ± 1.49 17.27 ± 1.53 17.05 ± 1.67 0.43 PRFI 0.0000 0.23 (0.12, 0.36) 0.10 (0.05, 0.19) 0.24 (0.14, 0.36) 0.13 (0.06, 0.22) < 0.01 SSI 0.0000 0.82 (0.74, 0.92) 0.92 (0.81, 1.02) 0.88 (0.80, 0.94) 0.83 (0.76, 0.90) < 0.01 SP-A1 0.2431 100.85 ± 15.25 110.64 ± 12.83 95.63 ± 15.48 105.65 ± 14.13 < 0.01 IR 0.2581 9.08 ± 1.00 8.45 ± 0.99 9.03 ± 1.13 8.82 ± 1.01 < 0.01 ARTh 0.0000 370.75 (338.43, 411.47) 421.00 (380.75, 476.65) 370.10 (323.95, 401.70) 407.95 (365.18, 467.22) < 0.01 DA-Ratio 0.0019 4.50 (4.20, 4.70) 4.20 (4.00, 4.50) 4.50 (4.20, 4.80) 4.40 (4.10, 4.60) < 0.01 CBI 0.0000 0.24 (0.07, 0.46) 0.04 (0.01, 0.15) 0.42 (0.13, 0.65) 0.05 (0.01, 0.20) < 0.01 BAD-D 0.0983 1.81 ± 0.51 1.33 ± 0.53 1.78 ± 0.43 1.42 ± 0.49 < 0.01 TBI 0.0000 0.45 (0.34, 0.59) 0.33 (0.09, 0.42) 0.46 (0.38, 0.59) 0.31 (0.16, 0.43) < 0.01 Comparison of Model Frameworks Three model frameworks (A, B, and C) and six machine learning algorithms were compared. The primary results are summarized in Table 3 , with complete metrics provided in Table S4. Model B achieved the highest overall accuracy and was selected as the optimal architecture for subsequent analyses. In Model B, Stage 1 and Stage 2 reached validation accuracies of 93.95% and 86.51%, respectively, resulting in an overall validation accuracy of 86.08% and a testing accuracy of 82.67%. Table 3 Comparison of the best-performing models and their independent generalization performance. Model Best Algorithm Cross-Validation (Accuracy%) Independent Test Set(%) Stage 1 Stage 2 Stage 3 \(\:{\text{ACC}}_{\text{Total}}\) \(\:{\text{ACC}}_{\text{Test}}\) A XGBoost 76.86 / / 76.86 76 B XGBoost 93.95 86.51 / 86.08 82.67 C XGBoost 93.95 96.17 84.19 82.83 80 Performance of the Optimal Model The corresponding ROC curves and AUC values for Model B during cross-validation are shown in Fig. 3 , which demonstrate that XGBoost achieved substantially higher AUCs than the competing algorithms did in both stages. To evaluate the clinical applicability of the final model, its performance was tested on an independent dataset. XGBoost-based Model B achieved an 82.67% agreement rate with surgeons' actual procedural decisions, demonstrating strong predictive reliability and generalizability. The confusion matrix for the test set (Fig. 4 ) further illustrates stable classification performance across categories. Ablation Study on Biomechanical Features To quantify the contribution of corneal biomechanical parameters, ablation experiments were performed by comparing models trained with and without biomechanical features across five machine learning algorithms and the two-stage classification task (Table 4 ). In Stage 1, the accuracy of the biomechanical features improved by approximately 0.5–1.5 percentage points, with the largest improvement observed in the DT model. In Stage 2, the benefits were more pronounced, with overall performance gains of 1.09–4.60 percentage points across algorithms. In addition, an approximately 4% improvement was also observed in the independent test set, suggesting that the effect of biomechanical features can be generalized beyond internal validation. Interpretability Results To assess the clinical validity of the model predictions, SHAP analysis was conducted for both stages of Model B. The SHAP summary plot for Stage 1 is presented in Fig. 5 . A nonlinear SHAP contribution pattern of CBI was observed, and its corresponding dependency plot is provided in Fig. 6 . The overall SHAP summary plot for Stage 2 is shown in Fig. 7 . For a more granular view of feature contributions to each surgical category, one-vs-others SHAP summary plots for SMILE, SURFACE, and LASIK are displayed in Fig. 8 . Table 4 Ablation results showing the impact of corneal biomechanical features. Cross-validation and independent testing demonstrated consistent performance gains when biomechanical parameters were included, particularly in the second-stage classification. Stage Algorithm Accuracy Advance Without biomechanical Within biomechanical First-stage cross validation LR 78.53 79.03 0.5 DT 81.45 82.96 1.51 SVM 78.33 78.33 0 MLP 90.52 91.03 0.51 XGBoost 92.84 93.95 1.11 Second-stage cross validation LR 75.35 79.95 4.6 DT 71.32 72.71 1.39 SVM 74.26 75.50 1.24 MLP 79.22 80.31 1.09 XGBoost 83.26 86.51 3.25 Independent testset 78.67 82.67 4 Discussion This study demonstrated that corneal biomechanics provide additional and nonredundant information in cases where conventional refractive and tomographic screening offers limited discriminatory power. Unlike prior studies that focused mainly on laser procedures or used a limited feature set, our framework simultaneously covers ICL, KLEx, LASIK, and SURFACE and follows a clinically consistent pathway that determines whether corneal tissue should be altered and then specifies how it should be altered. In ablation experiments, biomechanical features provided consistent performance gains in both stages, indicating that indices reflecting material stiffness, dynamic deformation, and ectasia-related risk add information beyond conventional refractive and tomographic measures. Within this framework, different model architectures were evaluated to balance decision granularity and overall predictive stability. Although Model C introduces a third stage to address the overlap between SMILE and LASIK, its overall performance remains inferior to that of Model B. This difference likely reflects several factors: progressive sample splitting reduces the effective training size for Stage 3 (limited to stromal procedures), the chained design propagates and amplifies upstream errors, and the deeper hierarchy may have increased sensitivity to noise and class imbalance under the current sample size and distribution. Given its superior overall performance and robustness, Model B was selected for further analysis. Interpretability analyses were subsequently performed to examine how biomechanical parameters contributed to model behavior across different procedures. In Stage 1, the SHAP results indicated that refractive status remained the primary driver of separating the ICL from corneal ablation (Fig. 5 ), with the SPH ranking highest; a lower SPH contributed more strongly to ICL predictions, which is consistent with the tendency to avoid corneal tissue removal in eyes with thin corneas or high myopia 29 . As indications for ICLs expand beyond high myopia to moderate myopia and presbyopia 30 , 31 , increasing evidence and consensus emphasize that corneal morphology and biomechanical safety are becoming important considerations beyond refractive error alone 32 , and ICLs are often favored when the cornea is thin or biomechanically suspicious 33 . Our SHAP patterns were consistent with this practice: higher BAD-D, PRFI, and TBI were associated with greater contributions toward ICLs, whereas higher SSI supported laser procedures, linking greater material stability with corneal ablation candidacy. In addition, CBI demonstrated a nonlinear, context-dependent pattern (Fig. 6 ): the model favored laser procedures at lower CBI values (approximately − 1.0–0.4); in the intermediate range (approximately − 0.4–0.9), accompanied by higher BAD-D and DA-Ratio, the model favored ICL; and at very high CBI values (> 0.9), SHAP contributions plateaued, suggesting a diminishing marginal discriminatory value of CBI under high-risk backgrounds and a greater reliance on multiparameter patterns. The Stage 2 SHAP results described how the model distributed predictions across laser subtypes. CYL, SPH, and Dt remained among the most influential features (Fig. 7 ), indicating that refractive status continued to shape laser selection. Given evidence that different procedures produce different degrees of biomechanical weakening 23 , 24 , 26 , the required preoperative biomechanical reserve may differ by procedure. Consistently, the SSI, PRFI, and ARTh ranked among the top contributors in Stage 2, supporting the use of biomechanical information beyond refractive parameters. Specifically (Fig. 8 ), higher SSIs and IRs contributed to SMILE, whereas higher PRFIs reduced the probability of SMILE, which is consistent with SMILE being more frequently assigned to eyes with greater stiffness, more uniform deformation responses, and lower risk signals in this cohort. For LASIK, higher CBI and TBI reduced its SHAP contributions, whereas LASIK occasionally increased under borderline biomechanical profiles (e.g., lower SSI or higher DA-Ratio). This apparent allocation pattern, where SMILE aligns with a greater biomechanical reserve, does not necessarily contradict the theoretical advantages of SMILE in preserving anterior stromal lamellae under comparable geometric conditions 22 , 34 , 35 . Longitudinal studies suggest that postoperative biomechanical changes are strongly influenced by geometric factors such as residual stromal bed thickness and percent tissue altered; under specific parameter combinations, a thicker cap and deeper lenticule plane may yield a thinner residual stromal bed and greater biomechanical reduction 24 . Because cap/flap/RSB data were not available in this study, this interpretation should be considered mechanism-based. In contrast, SURFACE more consistently reflected a "conservative" pathway: lower ARTh or IR favored SURFACE, which was consistent with avoiding flap creation and preserving anterior stromal continuity; prior work also suggested that surface ablation may preserve relatively more biomechanical reserve under certain conditions 22 , 26 . Collectively, biomechanical and risk-related indices not only improved discrimination among laser procedures but also shaped allocation boundaries consistent with clinical risk-control principles. Several limitations of this study should be acknowledged when interpretability analyses that elucidate model behavior. This was a single-center retrospective study, and model outputs may reflect local practice patterns. Postoperative ectasia or long-term biomechanical stability was not used as a clinical safety endpoint, and safety-related inferences should therefore be interpreted cautiously. In addition, the procedure categories were imbalanced, and the model inputs were limited to structured preoperative variables without incorporating raw imaging data or deformation waveforms. Finally, prospective validation in real-world clinical workflows was not performed. Collectively, these factors indicate that the present findings should be interpreted as model-level evidence of information gain rather than definitive clinical guidance. Conclusion This study demonstrated that a multistage AI model integrating multidimensional preoperative parameters improved the predictability of refractive surgery selection. The inclusion of corneal biomechanical indices resulted in consistent performance gains across multiple machine learning models and enabled interpretability analyses that helped contextualize how biomechanical features contributed to model predictions, supporting the nonredundant value of biomechanics beyond refractive and tomographic parameters. Although the present results do not establish a definitive role for biomechanics in defining procedural safety boundaries or resolving ambiguous surgical indications, they support its potential value as a complementary component within AI-assisted refractive surgery planning. Further multicenter prospective studies are warranted to validate the clinical applicability and generalizability of these findings. Abbreviations ICL Implantable Collamer Lens KLEx Keratorefractive lenticule extraction LASIK Laser-assisted in situ keratomileusis SHAP Shapley additive explanations LASEK Laser-assisted subepithelial keratectomy PRK Photorefractive Keratectomy Trans-PRK Transepithelial PRK SURFACE Surface ablation (including PRK, Trans-PRK, LASEK) SMILE Small Incision Lenticule Extraction DT Decision Tree RF Random Forest LR Logistic Regression SVM Support Vector Machine MLP Multilayer Perceptron XGBoost eXtreme Gradient Boosting SMOTE Synthetic Minority Oversampling Technique AUC Area Under the Curve SRI Surface Regularity Index SAI Surface Asymmetry Index CSI Corneal Symmetry Index RSB Residual Stromal Bed Declarations Ethics approval and consent to participate This study was approved by the Chongqing Aier Eye Hospital Medical Ethics Review Committee, Chinese Academy of Medical Sciences. All methods were carried out in accordance with relevant guidelines and regulations in the Declaration of Helsinki. All the subjects and/or their legal guardians signed an informed consent form. Consent for publication Not applicable. Availability of data and materials The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. Competing interests The authors declare that they have no competing interests. Funding This work was supported by the Chongqing Municipal Health Commission (Grant No. 2023MSXM122). Authors' contributions Study concept and design (YHL, GL, JHW); data collection (JHW, YHL, YLP, GL, CYX, XW); analysis and interpretation of data (YHL, JHW); writing of the manuscript (YHL, GL, CYX, JHW, YLP); critical revision of the manuscript (YHL, GL, YLP); statistical expertise (JHW, XW); administrative, technical, or material support (YLP, GL, CYX); supervision (YLP). All the authors read and approved the final manuscript. Acknowledgement Not applicable. References Shuang S, Dan W, Yewei Y, Fuying Q, Huilan X, Xiaobo X. Correction of presbyopia. Journal of Central South University Medical Sciences . 2022;47(10):1454. Wang Q, Fan L, Zhou Q. The best choice for low and moderate myopia patients incapable for corneal refractive surgery: implantation of a posterior chamber phakic intraocular lens. International Ophthalmology . 2023;43(2):575-581. Swaminathan U, Daigavane S. Comparative analysis of visual outcomes and complications in intraocular collamer lens, small-incision lenticule extraction, and laser-assisted In situ keratomileusis surgeries: a comprehensive review. Cureus . 2024;16(4) Chen Q, Yan H, Chen L, Wang L, Zheng Q, Ren Y. Adverse events associated with Implantable Collamer Lens: insights from the FDA MAUDE database. Front Med-Lausanne . 2025;12:1613060. Di Y, Li Y, Luo Y. Prediction of implantable collamer lens vault based on preoperative biometric factors and lens parameters. J Refract Surg . 2023;39(5):332-339. Yoo TK, Ryu IH, Lee G, et al. Adopting machine learning to automatically identify candidate patients for corneal refractive surgery. Npj Digit Med . 2019;2(1):59. Ting DSW, Pasquale LR, Peng L, et al. Artificial intelligence and deep learning in ophthalmology. British Journal of Ophthalmology . 2019;103(2):167-175. Yoo TK, Ryu IH, Choi H, et al. Explainable machine learning approach as a tool to understand factors used to select the refractive surgery technique on the expert level. Translational vision science & technology . 2020;9(2):8-8. Cui T, Wang Y, Ji S, et al. Applying machine learning techniques in nomogram prediction and analysis for SMILE treatment. Am J Ophthalmol . 2020;210:71-77. Wang K, Xu C, Li G, Zhang Y, Zheng Y, Sun C. Combining convolutional neural networks and self-attention for fundus diseases identification. Sci Rep-Uk . 2023;13(1):76. Li J, Dai Y, Mu Z, et al. Choice of refractive surgery types for myopia assisted by machine learning based on doctors’ surgical selection data. Bmc Med Inform Decis . 2024;24(1):41. Li L, Xiang Y, Chen X, et al. Machine learning model for predicting corneal stiffness and identifying keratoconus based on ocular structures. Intelligent Medicine . 2025;5(01):66-72. Ambrósio Jr R, Correia FF, Lopes B, et al. Corneal biomechanics in ectatic diseases: refractive surgery implications. The open ophthalmology journal . 2017;11:176. Steinberg J, Siebert M, Katz T, et al. Tomographic and biomechanical Scheimpflug imaging for keratoconus characterization: a validation of current indices. J Refract Surg . 2018;34(12):840-847. Moshirfar M, Motlagh MN, Murri MS, Momeni-Moghaddam H, Ronquillo YC, Hoopes PC. Advances in biomechanical parameters for screening of refractive surgery candidates: a review of the literature, part III. Medical Hypothesis, Discovery and Innovation in Ophthalmology . 2019;8(3):219. Yang K, Xu L, Fan Q, Zhao* D, Ren* S. Repeatability and comparison of new Corvis ST parameters in normal and keratoconus eyes. Sci Rep-Uk . 2019;9(1):15379. Liu Y, Zhang Y, Chen Y. Application of a scheimpflug-based biomechanical analyser and tomography in the early detection of subclinical keratoconus in chinese patients. BMC ophthalmology . 2021;21(1):339. Esporcatte LP, Salomão MQ, Junior NS, et al. Corneal biomechanics for corneal ectasia: Update. Saudi Journal of Ophthalmology . 2022;36(1):17-24. Borderie V, Beauruel J, Cuyaubère R, Georgeon C, Memmi B, Sandali O. Comprehensive assessment of Corvis ST biomechanical indices in normal and keratoconus corneas with reference to corneal enantiomorphism. J Clin Med . 2023;12(2):690. Damgaard IB, Reffat M, Hjortdal J. Review of corneal biomechanical properties following LASIK and SMILE for myopia and myopic astigmatism. The Open Ophthalmology Journal . 2018;12:164. Guo H, Hosseini-Moghaddam SM, Hodge W. Corneal biomechanical properties after SMILE versus FLEX, LASIK, LASEK, or PRK: a systematic review and meta-analysis. BMC ophthalmology . 2019;19(1):167. Xin Y, Lopes BT, Wang J, et al. Biomechanical effects of tPRK, FS-LASIK, and SMILE on the cornea. Front Bioeng Biotech . 2022;10:834270. Gao W, Zhao X, Wang Y. Change in the corneal material mechanical property for small incision lenticule extraction surgery. Front Bioeng Biotech . 2023;11:1034961. Hashemi H, Roberts CJ, Elsheikh A, Mehravaran S, Panahi P, Asgari S. Corneal biomechanics after SMILE, femtosecond-assisted LASIK, and photorefractive keratectomy: a matched comparison study. Translational vision science & technology . 2023;12(3):12-12. Qu Z, Li X, Yuan Y, et al. In vivo corneal biomechanical response to three different laser corneal refractive surgeries. J Refract Surg . 2024;40(5):e344-e352. Joshi S, Bari A, Shakkarwal C, et al. The visual outcomes and corneal biomechanical properties after PRK and SMILE in low to moderate myopia. Indian Journal of Ophthalmology . 2025;73(1):128-133. Shortt AJ, Allan BD, Evans JR. Laser‐assisted in‐situ keratomileusis (LASIK) versus photorefractive keratectomy (PRK) for myopia. Cochrane Database of systematic reviews . 2013;(1) Jacobs DS, Lee JK, Shen TT, et al. Refractive surgery preferred practice pattern®. Ophthalmology . 2023;130(3):P61-P135. Wang X, Zhou X. Update on treating high myopia with implantable collamer lenses. Asia-Pacific journal of ophthalmology . 2016;5(6):445-449. Kamiya K, Takahashi M, Takahashi N, Shoji N, Shimizu K. Monovision by implantation of posterior chamber phakic intraocular lens with a central hole (hole ICL) for early presbyopia. Sci Rep-Uk . 2017;7(1):11302. Li F, Ma Y, Qi W, Pazo EE, Yang R, Zhao S. Characteristics of biological parameters and implantable collamer lens (ICL) size selection in moderate, high, and super-high myopia eyes. BMC ophthalmology . 2025;25(1):103. HERZIG S, MD, FRCSC, DABO. AN EVOLUTION OF INDICATIONS FOR THEEVO VISIAN ICL FAMILY OF LENSES. STAAR Surgical Company . 2022; Akrobetu D, Nikpoor N. Choosing the Best Option for Refractive Surgery. CRSToday . 2023; Reinstein DZ, Archer TJ, Randleman JB. Mathematical model to compare the relative tensile strength of the cornea after PRK, LASIK, and small incision lenticule extraction. J Refract Surg . 2013;29(7):454-460. Roy AS, Dupps Jr WJ, Roberts CJ. Comparison of biomechanical effects of small-incision lenticule extraction and laser in situ keratomileusis: finite-element analysis. Journal of Cataract & Refractive Surgery . 2014;40(6):971-980. Additional Declarations No competing interests reported. Supplementary Files Supplementalmaterial.docx Cite Share Download PDF Status: Under Revision Version 1 posted Editorial decision: Revision requested 18 May, 2026 Reviews received at journal 15 May, 2026 Reviews received at journal 11 May, 2026 Reviews received at journal 09 May, 2026 Reviewers agreed at journal 29 Apr, 2026 Reviewers agreed at journal 28 Apr, 2026 Reviewers agreed at journal 28 Apr, 2026 Reviewers agreed at journal 23 Apr, 2026 Reviewers agreed at journal 18 Apr, 2026 Reviewers invited by journal 18 Apr, 2026 Editor invited by journal 14 Apr, 2026 Editor assigned by journal 13 Apr, 2026 Submission checks completed at journal 13 Apr, 2026 First submitted to journal 08 Apr, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9362570","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":629627020,"identity":"e9d6993e-ab01-4b6d-bebb-e34f5d89b9fe","order_by":0,"name":"Yinhao Li","email":"","orcid":"","institution":"Chongqing University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Yinhao","middleName":"","lastName":"Li","suffix":""},{"id":629627021,"identity":"4748ac9c-ffbb-45ee-b0cf-1d8a5001b677","order_by":1,"name":"Gang Li","email":"","orcid":"","institution":"Chongqing University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Gang","middleName":"","lastName":"Li","suffix":""},{"id":629627022,"identity":"877d1e93-f104-4b27-8080-47488ffca6e4","order_by":2,"name":"Chuanyun Xu","email":"","orcid":"","institution":"Chongqing Normal University","correspondingAuthor":false,"prefix":"","firstName":"Chuanyun","middleName":"","lastName":"Xu","suffix":""},{"id":629627023,"identity":"1735458a-db77-4f9b-a661-a91ce74c1192","order_by":3,"name":"Jiahui Wang","email":"","orcid":"","institution":"Chongqing Aier Eye Hospital","correspondingAuthor":false,"prefix":"","firstName":"Jiahui","middleName":"","lastName":"Wang","suffix":""},{"id":629627024,"identity":"97c9c798-10bf-484c-91fd-da8a15f6f019","order_by":4,"name":"Xin Wang","email":"","orcid":"","institution":"Chongqing University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Xin","middleName":"","lastName":"Wang","suffix":""},{"id":629627025,"identity":"e146b2ee-9b93-41ac-a6c5-7ec0a858f3b9","order_by":5,"name":"Yanli Peng","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAArUlEQVRIiWNgGAWjYNCCAhswdYAELQZppGs5TIJic/bew695DM7nGdw+Y3iAoeYOYS2WPefSrHkMbhcbnMsxOMBw7BkRTrqRY2YM1JK44QyPwQHGBiJcaHD/DUjLOVK03OAxfgxUTIIWy54cM8Y5BsnFkmfYCg4kHCNCizn7GeMPbyrs8vjOMG/+8KGGGIcxMLBJ8TAwJIB5CYQ1gLUwf/xBpOJRMApGwSgYoQAAaJA8/c0cYF0AAAAASUVORK5CYII=","orcid":"","institution":"Chongqing Aier Eye Hospital","correspondingAuthor":true,"prefix":"","firstName":"Yanli","middleName":"","lastName":"Peng","suffix":""}],"badges":[],"createdAt":"2026-04-09 03:24:03","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9362570/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9362570/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":107916263,"identity":"d2bdab67-9699-4b22-b891-5b4c1775bdaf","added_by":"auto","created_at":"2026-04-27 14:13:30","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":187592,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic illustration of the three refractive surgery classification models.\u003c/p\u003e\n\u003cp\u003eModel A performs direct four-class classification of ICL, SMILE, LASIK, and SURFACE. Model B uses a two-stage hierarchical structure: Stage 1 distinguishes ICL from corneal refractive surgery; Stage 2 classifies SMILE, LASIK, and SURFACE. Model C extends this hierarchy by adding a third stage: Stage 1 separates ICL from corneal procedures; Stage 2 differentiates SURFACE from stromal procedures; and Stage 3 performs binary classification between SMILE and LASIK.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9362570/v1/e319db756af0f85de67dfbd1.png"},{"id":107916272,"identity":"cdc3f05f-dc61-4169-9e46-6b04b7a6e987","added_by":"auto","created_at":"2026-04-27 14:13:31","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":244561,"visible":true,"origin":"","legend":"\u003cp\u003eBonferroni-adjusted post hoc pairwise comparisons across the four refractive surgery types for each biomechanical metric. The heatmap intensity reflects the magnitude of intergroup differences. Statistical significance is marked by asterisks (*** P_adj\u0026lt;0.001, ** P_adj\u0026lt;0.01, * P_adj\u0026lt;0.05).\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-9362570/v1/6aa876c172f1a1923de6f09b.png"},{"id":107916265,"identity":"73367137-eaf7-41db-8181-13fb3b0f3715","added_by":"auto","created_at":"2026-04-27 14:13:30","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":234415,"visible":true,"origin":"","legend":"\u003cp\u003eROC curves of the six classification algorithms used in the two stages of Model B. (A) ROC curves for Stage 1 of Model B. (B) Macro-ROC curves for Stage 2 of Model B. Classes 0, 1, and 2 represent SMILE, SURFACE, and LASIK,respectively.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-9362570/v1/5cf26e1f466fc6181557bffb.png"},{"id":108007459,"identity":"0b11d2bb-8615-439c-a976-7fb05ec55f8b","added_by":"auto","created_at":"2026-04-28 13:00:09","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":69351,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion matrix illustrating the prediction performance of the model on the independent test set. Class indices 0–3 denote ICL, SMILE, SURFACE, and LASIK. The color scale indicates the sample count in each cell, with deeper shades reflecting more accurate predictions.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-9362570/v1/8af074adf4cd5d962860076d.png"},{"id":108006918,"identity":"911a52f9-cf2d-4bf3-849a-98cab8fee949","added_by":"auto","created_at":"2026-04-28 12:57:54","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":140154,"visible":true,"origin":"","legend":"\u003cp\u003eSHAP summary plot illustrating feature contributions in Stage 1 of Model B. The laser procedure was labeled the positive class,and the ICL was labeled the negative class. The color gradient reflects the magnitude of each feature value, and SHAP values represent the direction and extent of their impact on model prediction.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-9362570/v1/0c43fda71a2558cf52863aeb.png"},{"id":107916267,"identity":"a89de555-b5e5-4dce-b763-1d30ed454eae","added_by":"auto","created_at":"2026-04-27 14:13:30","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":286794,"visible":true,"origin":"","legend":"\u003cp\u003eSHAP dependency plots illustrating the interaction between CBI and key biomechanical parameters in the Stage 2 classifier. The contribution of the Corvis Biomechanical Index (CBI) to model prediction varies with (A) SP-A1, (B) thinnest corneal thickness (TCT), (C) deformation‒amplitude ratio (DA‒Ratio), and (D) BAD-D.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-9362570/v1/c7d93989e231b267909e2dd9.png"},{"id":108007242,"identity":"8be25d5a-3cd9-49f3-83c0-573d72fc747c","added_by":"auto","created_at":"2026-04-28 12:59:07","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":55034,"visible":true,"origin":"","legend":"\u003cp\u003eGlobal feature importance for the second-stage classifiers (SMILE, SURFACE, and LASIK). The mean absolute SHAP values quantify each feature's contribution to model predictions.\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-9362570/v1/a6a3368de3fc5bdcf5dae747.png"},{"id":107916269,"identity":"133f5e64-833f-439f-9774-ee2e5ec1767a","added_by":"auto","created_at":"2026-04-27 14:13:30","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":208637,"visible":true,"origin":"","legend":"\u003cp\u003eSHAP summary plots for the second-stage one-vs-rest classifiers distinguishing SMILE, SURFACE, and LASIK. The x-axis shows SHAP values (direction and magnitude of feature contribution), and the y-axis lists features. Each dot represents one sample; the color indicates the feature value (blue = low, red = high). Higher SHAP values increase the probability of the positive class, whereas lower values favor the negative classes. (A) SMILE vs others. (B) SURFACE vs Others. (C) LASIK vs others.\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-9362570/v1/669e96ba7ed2a90cce2abeef.png"},{"id":108490979,"identity":"4041c147-e903-41de-8e13-824b3e287d0e","added_by":"auto","created_at":"2026-05-05 09:50:41","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1583448,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9362570/v1/c84c85e3-0ecb-4df5-906d-329a75ef0e76.pdf"},{"id":108007265,"identity":"87ae5310-f691-4121-a738-9da74116cca1","added_by":"auto","created_at":"2026-04-28 12:59:12","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":35193,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementalmaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-9362570/v1/a56462bb06b103f0c6910a7f.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Contribution of Corneal Biomechanics to Machine Learning-Based Refractive Procedure Selection: A Retrospective Observational Study","fulltext":[{"header":"Background","content":"\u003cp\u003eRefractive surgery is widely used to correct refractive errors and restore visual function, with current procedures broadly classified into corneal refractive surgery and phakic intraocular lens implantation, most commonly the implantable collamer lens (ICL) \u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. Corneal refractive procedures, including Keratorefractive lenticule extraction (KLEx), Laser-Assisted In Situ Keratomileusis (LASIK), Laser-Assisted Subepithelial Keratectomy (LASEK), Photorefractive Keratectomy (PRK), and Transepithelial PRK (Trans-PRK), reshape the anterior cornea and are generally safe and predictable, whereas ICL implantation preserves corneal tissue and is often preferred in patients with high myopia, thin corneas, or suspected ectatic risk \u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. Despite favorable safety profiles, inappropriate procedure selection may still result in vision-threatening complications, such as post-operative corneal ectasia following corneal refractive surgery or vault-related anterior segment complications after ICL implantation \u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e. Therefore, comprehensive preoperative evaluation incorporating corneal thickness, tomographic patterns, biomechanical characteristics, anterior chamber anatomy, endothelial status, and patient-specific factors is essential for minimizing surgical risk and optimizing refractive outcomes \u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eWith the rapid development of artificial intelligence (AI) and machine learning (ML), these techniques have shown substantial promise in preoperative screening, identifying contraindications, and predicting postoperative outcomes in refractive surgery \u003csup\u003e\u003cspan additionalcitationids=\"CR7 CR8 CR9 CR10 CR11\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. Yoo et al. \u003csup\u003e6\u003c/sup\u003e developed multimodal ML classifiers incorporating demographic data, Pentacam tomography, and questionnaire information, achieving excellent AUCs (0.983/0.972) for identifying contraindications. Their subsequent four-class models (LASEK, LASIK, KLEx, and contraindication) achieved accuracies of 81.0% and 78.9%, respectively \u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. However, both studies lacked corneal biomechanical inputs, which limited the model's ability to capture biomechanically relevant information that may contribute to individualized risk assessment. Li et al. \u003csup\u003e11\u003c/sup\u003e reported 87.75% agreement with surgeons using long-term historical data, but only indirect biomechanical indicators (SRI, SAI, CSI) were available. Overall, existing ML models rely primarily on refractive and tomographic features and lack comprehensive incorporation of corneal biomechanical parameters.\u003c/p\u003e \u003cp\u003eIn recent years, corneal biomechanics has become an increasingly important dimension in refractive surgery safety assessment, gaining greater prominence in preoperative screening. The Corvis ST, which is based on dynamic Scheimpflug imaging, quantifies corneal deformation under an air‒puff load, and its material-related parameters and composite risk indices have been progressively used to identify subclinical keratoconus and biomechanically vulnerable corneas \u003csup\u003e\u003cspan additionalcitationids=\"CR14 CR15 CR16 CR17 CR18\" citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e. Moreover, multiple studies have demonstrated that KLEx, LASIK, and surface ablation procedures result in systematic differences in postoperative biomechanical weakening due to variations in ablation depth, flap creation, and tissue-removal patterns \u003csup\u003e\u003cspan additionalcitationids=\"CR21 CR22 CR23 CR24 CR25\" citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e. From the perspective of residual biomechanical stability, these disparities suggest that different procedures may correspond to different acceptable preoperative risk thresholds: for corneas in a biomechanically borderline state, certain procedures may warrant greater caution or even avoidance, whereas relatively \"conservative\" procedures may still remain within a safe range. Accordingly, incorporating biomechanical information into procedure selection may provide more comprehensive insights for surgical planning beyond its traditional role in contraindication screening. However, its role in procedure selection remains underexplored.\u003c/p\u003e \u003cp\u003eTo address these limitations, we developed a two-stage surgical recommendation model that integrates Corvis ST biomechanical parameters, Pentacam tomographic features, and demographic information to jointly evaluate ICL, KLEx, LASIK, and surface ablation procedures (including PRK, Trans-PRK, and LASEK, hereafter referred to as SURFACE). This unified, biomechanics-informed framework learns clinically relevant decision logic and provides interpretability through SHAP-based analyses. This study aims to evaluate the performance of this framework and to quantify the contribution of corneal biomechanical parameters to refractive procedure selection beyond conventional refractive and tomographic features.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\n \u003ch2\u003ePatients\u003c/h2\u003e\n \u003cp\u003eThis single-center retrospective study included patients who underwent refractive surgery at the Refractive Surgery Department of Chongqing Aier Eye Hospital between October 2023 and November 2024. Only patients whose preoperative records were complete were eligible. The type of refractive procedure was determined jointly by experienced refractive surgeons on the basis of clinical indications and patient preference. All the data were anonymized prior to analysis to remove personal identifiers. The study was approved by the Institutional Review Board of Aier Eye Hospital (Approval No. 202401007) and adhered to the tenets of the Declaration of Helsinki.\u003c/p\u003e\n \u003cp\u003eThe study involved four commonly performed refractive procedures: ICL and three types of corneal laser surgery, namely, the KLEx, LASIK, and SURFACE procedures, where KLEx specifically denotes small-incision lenticule extraction (SMILE). The inclusion and exclusion criteria were established with reference to published guidelines and expert consensus statements, previous literature \u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e,\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e, and the clinical experience of senior refractive surgeons. Eligible patients were 18 to 45 years of age; had completed comprehensive ophthalmic examinations and preoperative assessments; and underwent ICL, SMILE, LASIK, or SURFACE with complete clinical documentation. Eyes with keratoconus or other ectatic corneal disease, visually significant cataracts or glaucoma, active ocular infection or inflammation, severe adnexal abnormalities, retinal pathology affecting visual function, systemic or metabolic disease likely to affect ocular status, or psychiatric conditions impairing cooperation were excluded.\u003c/p\u003e\n\u003c/div\u003e\n\u003ch3\u003ePreoperative examinations and surgery\u003c/h3\u003e\n\u003cp\u003eAll enrolled patients underwent standardized preoperative screening. Examinations included Pentacam HR (OCULUS) for corneal tomography and anterior segment evaluation; Corvis ST (OCULUS) for corneal biomechanical assessment; ultrasound biomicroscopy (UBM; SW-3200L) for ciliary body and anterior lens surface evaluation; and anterior segment OCT (AS-OCT; Carl Zeiss Meditec) for corneal thickness and anterior chamber depth measurements. UBM and AS-OCT were used solely for clinical safety assessment, and their quantitative parameters were not included as model features. Routine examinations included visual acuity, refraction, intraocular pressure, and systemic evaluation. Surgical decision-making included corneal thickness, corneal shape, residual stromal bed thickness, anterior chamber depth, and refractive stability, in accordance with current safety standards for refractive surgery.\u003c/p\u003e\n\u003ch3\u003eData collection\u003c/h3\u003e\n\u003cp\u003eA total of 395 patients were included in this study. The preoperative parameters consisted of 48 features encompassing corneal biomechanical properties, Pentacam tomographic data, and demographic characteristics. A full feature list is provided in Table S1, an aggregated overview of feature categories is presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, and definitions of biomechanical parameters are provided in Table S2. This dataset supports multidimensional clinical feature analysis and was used for subsequent model development and validation.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eCategorization of preoperative variables included in the refractive surgery decision-making model.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" colname=\"c1\"\u003e\n \u003cp\u003eType\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colname=\"c2\"\u003e\n \u003cp\u003eFeature Names\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003eAmount\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c1\"\u003e\n \u003cp\u003eCorneal Biomechanical Properties\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c2\"\u003e\n \u003cp\u003ebIOP, PRFI, SSI, SP-A1, IR, ARTh, DA-Ratio, CBI, BAD-D, TBI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\n \u003cp\u003e10\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c1\"\u003e\n \u003cp\u003ePentacam Corneal Topography Data\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c2\"\u003e\n \u003cp\u003eFront K1, Front K2, Front K mean, Front K1 Axis, Back K1, Back K2, Back K mean, Back Axis K1, TCT, CV, ACV, ACA, ACD, Front K max, Q Value, Anterior Surface Elevation at the Corneal Thinnest Point, Posterior Surface Elevation at the Corneal Thinnest Point, Df, Db, Dp, Dt, Da, Pupil Center X, Pupil Center Y, Angle Kappa, CD, IS Value, Average Pachymetric Progression Rate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\n \u003cp\u003e28\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c1\"\u003e\n \u003cp\u003eDemographics and Baseline Characteristics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c2\"\u003e\n \u003cp\u003eGender, Age, Ni-BUT, SPH, CYL, Axis, SE, UDVA, CDVA, Dark Pupil Diameter\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\n \u003cp\u003e10\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003ch3\u003eModel Design\u003c/h3\u003e\n\u003cp\u003eTo compare different strategies for refractive surgery classification via real-world clinical data, three machine learning frameworks have been developed (Fig. \u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e\n\u003cp\u003eModel A performs direct four-class classification of ICL, SMILE, LASIK, and SURFACE. Model B uses a two-stage hierarchical structure: Stage 1 distinguishes ICL from corneal refractive surgery; Stage 2 classifies SMILE, LASIK, and SURFACE. Model C extends this hierarchy by adding a third stage: Stage 1 separates ICL from corneal procedures; Stage 2 differentiates SURFACE from stromal procedures; and Stage 3 performs binary classification between SMILE and LASIK.\u003c/p\u003e\n\u003ch3\u003eTraining Setup\u003c/h3\u003e\n\u003cp\u003eData preprocessing began with the removal of missing values attributable to measurement or acquisition errors. Clinically plausible missing entries were imputed according to standard ophthalmic assessment criteria. The outliers were screened via Z-scores, and samples with any feature exceeding |Z| \u0026gt; 5 were excluded. After filtering, 742 eyes remained for model development. All numeric features were standardized via Z-score normalization, and categorical variables were encoded via one-hot encoding. Redundant or low-variance features were removed via Pearson correlation analysis (threshold\u0026thinsp;=\u0026thinsp;0.95) and variance filtering (threshold\u0026thinsp;=\u0026thinsp;0.015), resulting in 38 features for Models A and B. In Model C, the third-stage classifier excluded the PRFI due to its limited discriminatory value, retaining 37 features.\u003c/p\u003e\n\u003cp\u003eSix supervised learning algorithms were evaluated: decision tree (DT), random forest (RF), logistic regression (LR), support vector machine (SVM), multilayer perceptron (MLP), and eXtreme gradient boosting (XGBoost). The data were split at the patient level at a 9:1 train\u0026ndash;test ratio. Fivefold cross-validation was performed on the training set via patient-based grouping to prevent data leakage between the eyes of the same patient. To address class imbalance, SMOTE oversampling was applied within each training fold, whereas the validation and testing sets remained unchanged. The hyperparameters were tuned via a grid search. All the experiments were implemented in Python 3.10 via scikit-learn 1.2 and XGBoost 1.7 on an NVIDIA RTX 3070 platform.\u003c/p\u003e\n\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\n \u003ch2\u003eEvaluation Metrics\u003c/h2\u003e\n \u003cp\u003eModel performance was assessed via accuracy, recall, and macro-F1, which capture overall predictive correctness, sensitivity to minority classes, and balanced multiclass performance, respectively. The definitions were as follows:\u003c/p\u003e\n \u003cdiv id=\"Equa\" class=\"Equation\"\u003e\n \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e$$\\:\\text{Accuracy}\\text{=}\\frac{\\text{TP}\\text{+}\\text{TN}}{\\text{TP}\\text{+}\\text{TN}\\text{+}\\text{FP}\\text{+}\\text{FN}}\\text{,}$$\u003c/div\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Equb\" class=\"Equation\"\u003e\n \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e$$\\:\\text{Recall}\\text{=}\\frac{\\text{TP}}{\\text{TP}\\text{+}\\text{FN}}\\text{,}$$\u003c/div\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Equc\" class=\"Equation\"\u003e\n \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e$$\\:\\text{Precision}\\text{=}\\frac{\\text{TP}}{\\text{TP}\\text{+}\\text{FN}}\\text{,}$$\u003c/div\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Equd\" class=\"Equation\"\u003e\n \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e$$\\:\\text{F}\\text{1=}\\frac{\\text{2}\\text{\u0026times;}\\text{Precision}\\text{\u0026times;}\\text{Recall}}{\\text{Precision}\\text{+}\\text{Recall}}\\text{,}$$\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{TP}\\)\u003c/span\u003e\u003c/span\u003e = true positives, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{TN}\\)\u003c/span\u003e\u003c/span\u003e = true negatives, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{FP}\\)\u003c/span\u003e\u003c/span\u003e = false positives, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{FN}\\text{}\\)\u003c/span\u003e\u003c/span\u003e= false negatives. Macro-F1 is particularly informative under imbalanced class distributions.\u003c/p\u003e\n \u003cp\u003eFor validation in multistage models, a prediction was considered correct only if the sample passed all decision stages along its true clinical pathway. Therefore, overall validation accuracy was computed via weighted multiplicative aggregation:\u003c/p\u003e\n \u003cdiv id=\"Eque\" class=\"Equation\"\u003e\n \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Eque\" name=\"EquationSource\"\u003e\u003cimg src=\"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAV0AAABwCAYAAAC9zaPrAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAAFiUAABYlAUlSJPAAABWRSURBVHhe7d3Pb9P2/wfwZ773jaXZiaGPUNwdph0isQATlEqbtDqCaeLAqMd2QBqi2JM4TKzdWqod+LVkbJMmQVq0SRxgSRGTpgkKKVKRGi8aLCAibdqB2poQ2ylWYPsD3t/D+vbsd9I2bWzHSV4PyaLxOy6pY7/yzuv9K8IYYyCEEBKI/xN3EEII8Q8FXUIICRAFXUIICRAFXUIICRAFXUIICRAFXUIICRAFXUIICRAFXUIICRAFXUIICRAFXUIICRAFXUIICRAFXRIaExMTmJ2dFXcTD1iWhVQqBcuyxCISMAq6JBTy+Tyy2SxeeOEFsYh4IBaL4bnnnkMqlRKLSMAo6JK2M00TmqbhzJkzSCQSYjHxyLlz52AYBiYmJsQiEqAITe1I2m3r1q2IxWK4efOmWEQ8Njs7iz179qBYLGJgYEAsJgGgmi5pq+npaZTLZZw/f14sahtd1xGJRNDf3+96zL+am6bpetxJdu/eDVmWcfDgQbGIBISCLmkb0zTxySefYHx8HPF4XCxum4GBARSLRQwNDUHXdZw8eRK5XA6FQgGmaWJoaAjFYhGLi4vioR3h/PnzMAwDmUxGLCIBoKBL2ubs2bMAgA8//FAsartSqYTNmzfj5MmTuHnzJhYWFiDLMs6ePYvFxUU8fvzYrglzqVSqI2q/8Xgcqqris88+g2maYjHxGyM9J5fLMQBMkiR7HwAGgBmG4XquXwzDYABYOp0Wi0JBlmXX+ZEkiUmSxIrFol2ey+UcR3QWfv5VVRWLiM8o6PYowzDsG45/9gYZSFRVZQBYtVoVi0IBgH0uGgUo8QMqnU6H+kOkEf4eBPVBS/5F6YUedfXqVQwODqK/vx+8A8vi4iI2bdokPtVzpmkim81CVVXEYjGxuO3y+TxkWYaiKACAu3fvAku5UCw1rEmS5MpDj46OQpIk7Nixw94XdseOHQMcaR4SDAq6PWp+fh4LCwuYm5sDlgIJlhqR/MZv8pGREbEoFBYWFvD666+7Hquqaj/+7rvvcPjwYWiaZudE+b9BnD+vxONxyLKMbDZLI9UCREG3RxUKBWzevNmurZVKJQwNDYlP85xlWchms0gmk6EdCDE3N4d9+/bZj7PZLAYHB+3HpmlibGwMg4OD9vm7evUqDh8+bD+nU/CuY99++61YRPwi5htI9ysWi0yWZdc+WZbtRiI/TU1NufKl3cLZyNZpotGoq9GQ+Itquj3ou+++c3WON00ThUIBGzduhKZprud67fPPPwcAvPHGG2JRx+rE1IKToigwDIMmGwoIBd0eNDc3h23bttmP//rrLwDA0NCQryPDKpUKDMPA8PBwKBvQ1uvu3bswDAORSEQs6ghvvvkmAODatWtiEfEBzb1AAjMxMYHTp08jl8vZPQNIOPT19aFWq6FarXbVB2IYUU2XBGZmZgYAXLVsEg68EfXWrVtiEfEYBV0SCJ5aSCaToZpngfxr7969AIAffvhBLCIeo6BLAnHlyhXAUaMi4cK/ffB+28Q/FHRJIPjNvHPnTrGIhEA8HockSajVavZAGeIPCrrEd6ZpolwuAwC2b98uFpOQ4N9Cbty4IRYRD1HQ7QJ8ku0gt7Xgcxckk8m2t4ynUqm6v6UdWxjxEYKUYvAXBd0u8Mwzz4i7QmVhYQFYWpan3S5fvoxcLodoNCoWBSKdTsMwDHF3KLz66qsAgHK5THMx+IiCbhdIJBJIp9OufclkEtVqFUvTd7a8TU1NuX7/WvCaUxjmWojFYlAUBbdv33btT6fTdX9zq1uxWHT9H7lcDqOjo6HtveF8f+7cueMqI96hoNslRkdHIcuy/bhcLmNyctL1nFaMjIzUBfZmWJZl1+xefvllsbhtEomE63wFoRMGhPBz8uuvv4pFxCMUdLvI5cuXXV+bs9ks8vm86zmteP/998Vdq/r999/tn8M2N0GQNc6gA/x68XMyPz8vFhGPUNDtIrFYDD/++KNrn6ZpqFQqrn3rFYvF1hw8SqUSAECSJLGo7TZv3izu6nn8nPDGT+I9CrpdZmBgwJV/rdVqOHTokGcNI87JvZtx//59AKhbxJGEE1/5olar0aKVPqGg24VGRkYwPDxsP/Yyvzs6Ogq2hjmSeP/ctQZr0h4bN260f+azzxFvUdDtUufOnXN9pfc6v9ss3oi2YcMGsYiEkDPPzVNDxFsUdLtULBbD999/79rnZX63Gc7/K0w9F8jKkskkAODp06diEfEABd0ulkgkfM3vruaff/6xfw77AA7yHz5q8N69e2IR8QAF3S43MjLiWsnWy/zuapxfT8MwMII0h6cYFhcXxSLiAQq6PeDEiRP2V0a0Mb9LOgPvNhbW4cqdjoJuD4jFYvjmm29cAyeCyO/y7mLOgE9Ir/M96FqW1fTwR8uykM/noSgK+vr6EIlE0NfXB03ToOv6qr+r1eO7WSKRcC06GUR+98mTJ4AjR0g6A++ri6UZ7Ii3fA+6k5OT9s23kunpabz44os4fvw49u7di3K5DMYYHj58iEQigV27duH5559HX1+feCjgwfG9QFGUtuV3CSH/8jXo5vN5ZLNZcbeLZVlIpVI4cuQIFEXB4uIiFEWxk/mxWAwjIyN2K/zg4KCnx/eaIPO7fCjpc889JxYR0rN8C7qVSgXHjx8HVhmNlEqlUCgUoKqq6+uvaN++fQCAl156ybW/1eN7TZD53VqtBgDYsmWLWEQ6xG+//SbuIi3yJehaloWxsTHXUNRGJiYmUC6XIUkSTpw4IRa7xGIxRKNRV9ejVo/vVYlEApcuXbIfB5HfJZ3DORscDZDwni9Bd3JyEgcPHrSHfv7vf/8TnwLTNHH69GkAwNdff91UYwtfsRQeHN/rdu/ejfHxcftxGPO7pmkin89D0zQoigLLspDJZOxG0omJCfGQrqPrOhRFQSQSQX9/PyqVCjRNQyQS8S0tRPzledDlF4KiKPjjjz8AAJs2bRKeBZw9exZYmvJv9+7dYnFDN2/etH9u9fiwENfOWm3LZDLir1i3U6dO1eV3Z2dnXc9pF13XIUkS3nnnHWSzWUiShMnJSezbtw8PHz5EMpnE6dOnu7p1fWJiArt27cKWLVtQrVZx584dXLlyxV6J44033hAPIR3A06BbqVRw8eJFO7e60tRwPDivloJYTqvHh4W4xMtq2+joqPgrWjIzM+PK74ZlxYCBgQEwxuza+NzcHE6cOIF4PI5YLIa3334b6OJJWTRNw+nTp+0lfmKxGGKxGDZs2ADDMEKxyCdZH8+CLs/jOhuzeI5QXDGgUqnYjSw7d+50lTWj1ePJf+LxOM6cOQMAiEajdoNjWMzMzAAAvvrqq54JMplMBtlsFul0etl+5YcOHRJ3kQ7hWdCdnJzE0aNHXVPD8blURc6JUJ599llXWTNaPb5dTNO00wT8a7GYPlht8zK9gKUPxk8++QQAcP78+UCXsFlNpVKBYRiQZbnug5unrhq1F3Qy0zQxNjYGSZIafqvhfzelFjqXJ0GX98fds2ePK0AgpMu0oIlg58dKB/F4HIwxSJJkBxExfbDa1uhGbMW7776LWq2G4eHhZWtV7XLlyhUAwNGjR8UiO6/ZbY2jvK3i5MmTYhGwdK8lk8m6D8dIJIJUKrXqPtJ+LQddnscVgwNffrpR8HLOTv/333+7yprR6vG8kYa/VlVVXY9zuRwOHz4sHlbHNM011zx1XW94TtphenoahUIBkiTh3LlzYnHb8Xyz2FA6OzsLwzCgqmpd8Ol0vK2iUU12enoatVoNQ0NDYhEYY3UNxY32kfZrKeg2yuNyKwXDeDxuL3B47do1sbiOruuuVvVWjy+VSnZNCUu1JueFvGnTpqZym5qmucapN6NUKq04WCQolUoFR44cAQBcvHgxdPlS0zRhGEbD2venn36KaDSKY8eOiUUdTdd11Go1yLJc935UKhU7DbR//35XGeks6w66fPjt8ePHG9Y2eCv4cgGGj1ZbbRhqpVLBjRs36mo7rRw/Ojpqv2Z+czuHB2/cuBHxeByZTKZhuoHvLxQK2LVrFzRNs8ucedt8Po/+/n5X+fz8fFMB3U+WZdkNMel0ui5fGga3bt0CALz55puu/ZlMBuVyOXT5Zz9ZloWff/4ZWErXJRIJu2cQvxad1ye//pzXnR9mZ2fR19eHrVu30sCatWDrcP36dSZJEgPADMMQi5lhGCyZTDIATJZlVq1WxacwxhjL5XIsGo0yAExVVfbgwQO77MGDB0xVVTY+Pu46xqnV4xljLJ1Os0anQZIkJsuy/ViWZaaqqv04nU67ytnS6wHAisUiY4wxVVVdj9nS7203/rqSyaRY5CkADABLp9Ni0ar49TM1NcUYY6xarbLx8XEGgOVyOfHpNkmSmj7H/L1f72tcTbFYtH+/eK00Uq1W7euZXzO5XI6l02n24MED+zqfmppyXevFYtF1bbKl63Wl87SSarXa1HmRZdl+nvj/k+XVR5tV8GDr3JyB1/lGOLfl3jzDMNj4+Lh9kwFgkiQxVVVdwWo5rR4vy3LdDaGqat2Nm06nXfvEi9owDAYhwPIg7Hzc7ovz+vXrDACLRqMNPzC9tNp7vxx+LiVJsoNQNBplw8PDrmDTqrAFXSZUJIaHh+3riQfdZDJZdw5UVW14La73/XW+7pUCN7++k8lk038fWUfQ7TbiDccvWPFiU1XVdWGJF7VYE2YNasPizRE0wzDsG/r69etisef4B/Rab0h+M/t9rsIYdNdDrCDkcrm6fWvhfN3NVFyKxSIbHh4Wd5NlrDun2w14X1lnYxifjlBswMlms3Z+mvd+cOaFC4UCDhw44DrmwoULrpz23NxcW7s4aZqGWq0GVVXrcuR+WG8vjR9++AFYpgV/Ofl83s5l9pJGvWEWFhYa9nDwS6lUwt69e8XdZBk9HXRLpZKrzywAPHr0qK5vMe8WxhvASqWS66Lmgdr5ezKZDAzDsAM6b/iIx+NtmS8gk8nY3cNWm5GtnSzLwszMTMMW/JUoimIPpOgljXrDzM3NtTRv9OPHj+2fV1vFuVKp4OnTp3WVFLK8ng66Y2NjdTWCHTt2wDAMV+vw2NgYisWiXbOdn5/HgQMHkMlkXAGUH6NpGi5cuGDv13UdV69exeHDh1fsaeEXXdcxNjYGAPj+++/XFMxawc9Xs6vKWpZlz3T2yiuviMWrunr1al0A6nbz8/PYsWMHNE2DaZp2b5xWvlE9evTI/nmlqVDz+TwKhQJOnTolFpGViPmGXtCosc+Zn3Xm+hrlxniu0pkHdDYw8hZnZy6P9xgQ875+q1arDV9vEJzncTU8jytuYqPRSmRZXlPjUTfkdJ3XHPMgn8vW+L6RtaOz2uWGh4d9velXMjU1Zd+8y3Ub9IphGGsONt0QdEWyLLf8t/Brxu8uhb2qp9ML3S6fz9tDaS9fviwW++7ll1+2f/79999dZV67e/duXaqoFxUKhZYH39Aqzv6ioNul+AoDAHDp0qW23EDOOTL8trCwgMHBwbo8ey+IRCIwTROapnkyHwVvGO61/HhQKOh2qUOHDqFWq2F8fNzT7mGWZTU9e5Xz5vd7svG5uTm88847QIP5m3uBJEnYvHlzw3lQ1orPVc2X2yLeoqDbhfiCnclkEh9++KFY3JI7d+4Aa+hdwLtw+b3A4eLiIpgPU192Aj47nhd/u3NVaGd6iHiHgm6X0XXdXrDzm2++8Tyt8NNPPwFrqAXx2u69e/fEIhJCf/75p/3zSy+95Coj3qCg20Usy8Jbb70FAJiamlqxj+V68eVzmsVfA88ThglfhYH8h88OKEmS5x/Y5F8UdLvIBx98YM/HOjIyIha3hDfUGIYBrKGmy7+i1mq10E3/51w41Y/0h3NkV6FQcJWF1f379wHAtUo08RYF3S4xPT1t10ILhYI9D4FXmyRJyGaz9v/XbL7P2ajld7exteDzZXAzMzOefyhcvHjR9Tgsy9uvhH+obtmyRSwiHqGg2wWcq0CEEa81+d2DoRmWZSGfz9f16TUMA6lUCrOzsy0H39nZWSiKUle7fe+995DJZFw17LDhi8mudUUUsgbiaAnSefgIoiC3Zqb84/jk42GY/q/REHBxa2X0mHOU20pbGDlH0BH/RNi/47cJ8c3s7Cz27NkDSZKanvyGBI9P7iTLMi1o6SNKLxDfbd++HVj6Ch/mr9a9jjei0Ug0f1HQJb6LxWJ2XjeMXcfIv/gK2b02J3HQKOiSQLz99tvA0hwJJHwqlQpqtRqi0agv/bvJfyjodjlN0xBZmhBlrVKpFCIeLX/Da0+8NhVWuq4jEonYq4WshZfnK2h8iXdaAcJ/1JBG6pim6VqyyKtLpL+/H4ZhwDCMlmfCChO/zleQFEXBzMwMrl+/7ukESaQe1XRJnXg8DsYY0um0WNSS4eFhAMCtW7fEoo7m1/kKCl+XLhqNUsANAAXdLmWapj2aLCzzy+7fvx9YmognbHhaIRKJtGUdu3biM8dRaiEYFHS7FK998dWOM5lM3dBeceOTnvslkUggmUyiXC6vK8fsp4GBATstsG3bNteS7sttzcwp3AmuXbsGADhw4IBYRHxAQbeL6bqO/v5+AMDo6Kg97+pymxcTYK/m0KFDQEhTDPl8HpIkIR6PQ1GUuvMjbt0wgMCyLGSzWfvDmfiPgm4XK5VKOHjwoLi7rfbt24doNBrKFMPCwkLdnAzdjn/4ffTRR2IR8QkF3S524cIFbNu2DVga4il+PRY3v9MLWBoooSgKyuWya5WCMJibm8Pg4CCwVOsVz4+4dUN64YsvvkA0Gm15MUvSPAq6XYrnTHnXrLCkFwDg2LFjwNJ0lGFhmiYMw7A/pHohvVCpVFAul6EoCk1YHiAKul2KL0mu6/q6Ovr7KR6PQ1VVZLPZlqdR9Mrdu3ftfG6v4B96/EOQBEScdox0B1VVWSvTFIpTEeZyOfEpLTEMgwFg6XRaLGoLVVWZqqri7qb5fb68xs9/K38zWR8akUbaRtM05PN5PHz4sO1fbyORCIrFYs+04Ifp3PcaSi+QtuFfa7/88kuxKFC6rvdUlynTNJHNZvHxxx9TwG0DqumStpqensaRI0cCn49B13UcPHgQi4uL6O/vx8mTJ3tmRFYqlcLi4iJNKN8mFHRJ223duhWxWCzQ3gC6rmPXrl0A0FNpBb6KRy/9zWFDQZe0XaVSwWuvvYYzZ854vnQ8+Y9lWXjxxRehqipOnTolFpOAUNAloZDP56FpGm7fvk2TaPtEURQYhoFffvlFLCIBooY0EgqKokBVVfz5559iEfGAZVl48uRJoCkc0hjVdAkhJEBU0yWEkABR0CWEkABR0CWEkABR0CWEkABR0CWEkABR0CWEkABR0CWEkABR0CWEkABR0CWEkABR0CWEkAD9P1FPmUi6As7+AAAAAElFTkSuQmCC\"\u003e\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{n}\\)\u003c/span\u003e\u003c/span\u003e represents the number of surgical categories, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{p}}_{\\text{i}}\\)\u003c/span\u003e\u003c/span\u003e represents the proportion of class \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{i}\\)\u003c/span\u003e\u003c/span\u003e in the true dataset, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{a}}_{\\text{i}\\text{j}}\\)\u003c/span\u003e\u003c/span\u003e represents the validation accuracy of class\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{}\\text{i}\\)\u003c/span\u003e\u003c/span\u003e at decision stage \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{j}\\)\u003c/span\u003e\u003c/span\u003e, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{m}}_{\\text{i}}\\text{}\\)\u003c/span\u003e\u003c/span\u003e represents the number of stages required by class \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{i}\\)\u003c/span\u003e\u003c/span\u003e. In this study, SMILE and LASIK share the same decision pathway; thus, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{i}\\)\u003c/span\u003e\u003c/span\u003e=1, 2, and 3 correspond to the ICL, stromal (SMILE \u0026amp; LASIK), and SURFACE procedures, respectively.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\n \u003ch2\u003eStatistical analysis\u003c/h2\u003e\n \u003cp\u003eThe Shapiro\u0026ndash;Wilk test was used to assess the normality of all variables. Normally distributed variables are reported as the means\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviations, whereas nonnormally distributed variables are summarized as medians with interquartile ranges. Because several variables violated normality assumptions, all intergroup comparisons were performed via the Kruskal\u0026ndash;Wallis test, ensuring methodological consistency and robustness.\u003c/p\u003e\n \u003cp\u003eTo further examine the contribution of corneal biomechanics to surgical decision-making, ten clinically established tomographic and biomechanical parameters were analyzed across surgical groups. For parameters demonstrating overall statistical significance, pairwise group comparisons were conducted via the Mann\u0026ndash;Whitney U test, with Bonferroni correction applied to adjust for multiple comparisons.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003eA total of 763 eyes of 395 patients were included in this study, comprising bilateral cases with identical procedures, bilateral cases with different procedures, and unilateral cases (27 single-eye records).\u003c/p\u003e\n\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\n \u003ch2\u003eBaseline characteristics\u003c/h2\u003e\n \u003cp\u003eNormality testing revealed that several variables were not normally distributed; therefore, all intergroup comparisons were conducted via the Kruskal\u0026ndash;Wallis test, with detailed results provided in Table S3. To examine distributional differences across surgical groups, ten widely used tomographic and biomechanical parameters were compared (Table \u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e\n \u003cp\u003eMost biomechanical parameters, including the PRFI, SSI, SP-A1, IR, ARTh, DA-Ratio, CBI, BAD-D, and TBI, were significantly different among the four surgical procedures (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05). Parameters that demonstrated overall significance were subsequently subjected to Bonferroni-adjusted Mann\u0026ndash;Whitney U tests, and pairwise comparison results are reported in Fig. \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\n \u003cp\u003eThe pairwise results revealed that across surgical types, all dynamic biomechanical parameters except bIOP demonstrated significant variability.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eDistribution and statistical comparison of corneal biomechanical metrics across the four refractive surgery groups. Biomechanical indices derived from Corvis ST are summarized for the ICL, SMILE, SURFACE, and LASIK procedures. S.W. represents the Shapiro\u0026ndash;Wilk test, which assesses variable distribution, and K.W. represents the Kruskal\u0026ndash;Wallis test, which is used to determine intergroup differences. Statistical significance was defined as P\u0026thinsp;\u0026lt;\u0026thinsp;0.05.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"7\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" colname=\"c1\"\u003e\n \u003cp\u003eFeature Name\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colname=\"c2\"\u003e\n \u003cp\u003eP(S.W.)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003eICL\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003eSMILE\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003eSURFACE\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colname=\"c6\"\u003e\n \u003cp\u003eLASIK\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colname=\"c7\"\u003e\n \u003cp\u003eP (K.W.)\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c1\"\u003e\n \u003cp\u003ebIOP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\n \u003cp\u003e0.1979\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e17.30\u0026thinsp;\u0026plusmn;\u0026thinsp;1.54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e17.19\u0026thinsp;\u0026plusmn;\u0026thinsp;1.49\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e17.27\u0026thinsp;\u0026plusmn;\u0026thinsp;1.53\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c6\"\u003e\n \u003cp\u003e17.05\u0026thinsp;\u0026plusmn;\u0026thinsp;1.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\n \u003cp\u003e0.43\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c1\"\u003e\n \u003cp\u003ePRFI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\n \u003cp\u003e0.0000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e0.23 (0.12, 0.36)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e0.10 (0.05, 0.19)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e0.24 (0.14, 0.36)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c6\"\u003e\n \u003cp\u003e0.13 (0.06, 0.22)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026lt;\u0026thinsp;0.01\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c1\"\u003e\n \u003cp\u003eSSI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\n \u003cp\u003e0.0000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e0.82 (0.74, 0.92)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e0.92 (0.81, 1.02)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e0.88 (0.80, 0.94)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c6\"\u003e\n \u003cp\u003e0.83 (0.76, 0.90)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026lt;\u0026thinsp;0.01\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c1\"\u003e\n \u003cp\u003eSP-A1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\n \u003cp\u003e0.2431\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e100.85\u0026thinsp;\u0026plusmn;\u0026thinsp;15.25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e110.64\u0026thinsp;\u0026plusmn;\u0026thinsp;12.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e95.63\u0026thinsp;\u0026plusmn;\u0026thinsp;15.48\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c6\"\u003e\n \u003cp\u003e105.65\u0026thinsp;\u0026plusmn;\u0026thinsp;14.13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026lt;\u0026thinsp;0.01\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c1\"\u003e\n \u003cp\u003eIR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\n \u003cp\u003e0.2581\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e9.08\u0026thinsp;\u0026plusmn;\u0026thinsp;1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e8.45\u0026thinsp;\u0026plusmn;\u0026thinsp;0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e9.03\u0026thinsp;\u0026plusmn;\u0026thinsp;1.13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c6\"\u003e\n \u003cp\u003e8.82\u0026thinsp;\u0026plusmn;\u0026thinsp;1.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026lt;\u0026thinsp;0.01\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c1\"\u003e\n \u003cp\u003eARTh\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\n \u003cp\u003e0.0000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e370.75 (338.43, 411.47)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e421.00 (380.75, 476.65)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e370.10 (323.95, 401.70)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c6\"\u003e\n \u003cp\u003e407.95 (365.18, 467.22)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026lt;\u0026thinsp;0.01\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c1\"\u003e\n \u003cp\u003eDA-Ratio\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\n \u003cp\u003e0.0019\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e4.50 (4.20, 4.70)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e4.20 (4.00, 4.50)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e4.50 (4.20, 4.80)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c6\"\u003e\n \u003cp\u003e4.40 (4.10, 4.60)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026lt;\u0026thinsp;0.01\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c1\"\u003e\n \u003cp\u003eCBI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\n \u003cp\u003e0.0000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e0.24 (0.07, 0.46)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e0.04 (0.01, 0.15)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e0.42 (0.13, 0.65)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c6\"\u003e\n \u003cp\u003e0.05 (0.01, 0.20)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026lt;\u0026thinsp;0.01\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c1\"\u003e\n \u003cp\u003eBAD-D\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\n \u003cp\u003e0.0983\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e1.81\u0026thinsp;\u0026plusmn;\u0026thinsp;0.51\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e1.33\u0026thinsp;\u0026plusmn;\u0026thinsp;0.53\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e1.78\u0026thinsp;\u0026plusmn;\u0026thinsp;0.43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c6\"\u003e\n \u003cp\u003e1.42\u0026thinsp;\u0026plusmn;\u0026thinsp;0.49\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026lt;\u0026thinsp;0.01\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c1\"\u003e\n \u003cp\u003eTBI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\n \u003cp\u003e0.0000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e0.45 (0.34, 0.59)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e0.33 (0.09, 0.42)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e0.46 (0.38, 0.59)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c6\"\u003e\n \u003cp\u003e0.31 (0.16, 0.43)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026lt;\u0026thinsp;0.01\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\n \u003ch2\u003eComparison of Model Frameworks\u003c/h2\u003e\n \u003cp\u003eThree model frameworks (A, B, and C) and six machine learning algorithms were compared. The primary results are summarized in Table \u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, with complete metrics provided in Table S4. Model B achieved the highest overall accuracy and was selected as the optimal architecture for subsequent analyses. In Model B, Stage 1 and Stage 2 reached validation accuracies of 93.95% and 86.51%, respectively, resulting in an overall validation accuracy of 86.08% and a testing accuracy of 82.67%.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eComparison of the best-performing models and their independent generalization performance.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"7\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e\n \u003cp\u003eBest Algorithm\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"4\" nameend=\"c6\" namest=\"c3\"\u003e\n \u003cp\u003eCross-Validation (Accuracy%)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colname=\"c7\"\u003e\n \u003cp\u003eIndependent Test Set(%)\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e\u003cstrong\u003eStage 1\u003c/strong\u003e\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e\u003cstrong\u003eStage 2\u003c/strong\u003e\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e\u003cstrong\u003eStage 3\u003c/strong\u003e\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colname=\"c6\"\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{ACC}}_{\\text{Total}}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colname=\"c7\"\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{ACC}}_{\\text{Test}}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c1\"\u003e\n \u003cp\u003eA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c2\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\n \u003cp\u003e76.86\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e/\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e/\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\n \u003cp\u003e76.86\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c7\"\u003e\n \u003cp\u003e76\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c1\"\u003e\n \u003cp\u003eB\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c2\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\n \u003cp\u003e93.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e86.51\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e/\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\n \u003cp\u003e\u003cstrong\u003e86.08\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c7\"\u003e\n \u003cp\u003e\u003cstrong\u003e82.67\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c1\"\u003e\n \u003cp\u003eC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c2\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\n \u003cp\u003e93.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e96.17\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e84.19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\n \u003cp\u003e82.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c7\"\u003e\n \u003cp\u003e80\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\n \u003ch2\u003ePerformance of the Optimal Model\u003c/h2\u003e\n \u003cp\u003eThe corresponding ROC curves and AUC values for Model B during cross-validation are shown in Fig. \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, which demonstrate that XGBoost achieved substantially higher AUCs than the competing algorithms did in both stages.\u003c/p\u003e\n \u003cp\u003eTo evaluate the clinical applicability of the final model, its performance was tested on an independent dataset. XGBoost-based Model B achieved an 82.67% agreement rate with surgeons\u0026apos; actual procedural decisions, demonstrating strong predictive reliability and generalizability. The confusion matrix for the test set (Fig. \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e) further illustrates stable classification performance across categories.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\n \u003ch2\u003eAblation Study on Biomechanical Features\u003c/h2\u003e\n \u003cp\u003eTo quantify the contribution of corneal biomechanical parameters, ablation experiments were performed by comparing models trained with and without biomechanical features across five machine learning algorithms and the two-stage classification task (Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e\n \u003cp\u003eIn Stage 1, the accuracy of the biomechanical features improved by approximately 0.5\u0026ndash;1.5 percentage points, with the largest improvement observed in the DT model. In Stage 2, the benefits were more pronounced, with overall performance gains of 1.09\u0026ndash;4.60 percentage points across algorithms. In addition, an approximately 4% improvement was also observed in the independent test set, suggesting that the effect of biomechanical features can be generalized beyond internal validation.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\n \u003ch2\u003eInterpretability Results\u003c/h2\u003e\n \u003cp\u003eTo assess the clinical validity of the model predictions, SHAP analysis was conducted for both stages of Model B. The SHAP summary plot for Stage 1 is presented in Fig. \u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e. A nonlinear SHAP contribution pattern of CBI was observed, and its corresponding dependency plot is provided in Fig. \u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e. The overall SHAP summary plot for Stage 2 is shown in Fig. \u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e. For a more granular view of feature contributions to each surgical category, one-vs-others SHAP summary plots for SMILE, SURFACE, and LASIK are displayed in Fig. \u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable float=\"Yes\" id=\"Tab7\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eAblation results showing the impact of corneal biomechanical features. Cross-validation and independent testing demonstrated consistent performance gains when biomechanical parameters were included, particularly in the second-stage classification.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"5\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\n \u003cp\u003eStage\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e\n \u003cp\u003eAlgorithm\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\" nameend=\"c4\" namest=\"c3\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colname=\"c5\" morerows=\"1\" rowspan=\"2\"\u003e\n \u003cp\u003eAdvance\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003eWithout biomechanical\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003eWithin biomechanical\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c1\" morerows=\"4\" rowspan=\"5\"\u003e\n \u003cp\u003eFirst-stage cross validation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c2\"\u003e\n \u003cp\u003eLR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e78.53\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e79.03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e0.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c2\"\u003e\n \u003cp\u003eDT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e81.45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e82.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e1.51\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c2\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e78.33\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e78.33\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c2\"\u003e\n \u003cp\u003eMLP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e90.52\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e91.03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e0.51\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c2\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e92.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e93.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e1.11\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c1\" morerows=\"4\" rowspan=\"5\"\u003e\n \u003cp\u003eSecond-stage cross validation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c2\"\u003e\n \u003cp\u003eLR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e75.35\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e79.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e4.6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c2\"\u003e\n \u003cp\u003eDT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e71.32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e72.71\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e1.39\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c2\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e74.26\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e75.50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e1.24\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c2\"\u003e\n \u003cp\u003eMLP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e79.22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e80.31\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e1.09\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colname=\"c2\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e83.26\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e86.51\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e3.25\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\n \u003cp\u003eIndependent testset\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c3\"\u003e\n \u003cp\u003e78.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c4\"\u003e\n \u003cp\u003e82.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colname=\"c5\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003cbr\u003e\u003c/div\u003e\n \u003c/div\u003e\n\u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study demonstrated that corneal biomechanics provide additional and nonredundant information in cases where conventional refractive and tomographic screening offers limited discriminatory power. Unlike prior studies that focused mainly on laser procedures or used a limited feature set, our framework simultaneously covers ICL, KLEx, LASIK, and SURFACE and follows a clinically consistent pathway that determines whether corneal tissue should be altered and then specifies how it should be altered. In ablation experiments, biomechanical features provided consistent performance gains in both stages, indicating that indices reflecting material stiffness, dynamic deformation, and ectasia-related risk add information beyond conventional refractive and tomographic measures.\u003c/p\u003e \u003cp\u003eWithin this framework, different model architectures were evaluated to balance decision granularity and overall predictive stability. Although Model C introduces a third stage to address the overlap between SMILE and LASIK, its overall performance remains inferior to that of Model B. This difference likely reflects several factors: progressive sample splitting reduces the effective training size for Stage 3 (limited to stromal procedures), the chained design propagates and amplifies upstream errors, and the deeper hierarchy may have increased sensitivity to noise and class imbalance under the current sample size and distribution. Given its superior overall performance and robustness, Model B was selected for further analysis.\u003c/p\u003e \u003cp\u003eInterpretability analyses were subsequently performed to examine how biomechanical parameters contributed to model behavior across different procedures. In Stage 1, the SHAP results indicated that refractive status remained the primary driver of separating the ICL from corneal ablation (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e), with the SPH ranking highest; a lower SPH contributed more strongly to ICL predictions, which is consistent with the tendency to avoid corneal tissue removal in eyes with thin corneas or high myopia \u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e. As indications for ICLs expand beyond high myopia to moderate myopia and presbyopia \u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e,\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e, increasing evidence and consensus emphasize that corneal morphology and biomechanical safety are becoming important considerations beyond refractive error alone \u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e, and ICLs are often favored when the cornea is thin or biomechanically suspicious \u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e. Our SHAP patterns were consistent with this practice: higher BAD-D, PRFI, and TBI were associated with greater contributions toward ICLs, whereas higher SSI supported laser procedures, linking greater material stability with corneal ablation candidacy. In addition, CBI demonstrated a nonlinear, context-dependent pattern (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e): the model favored laser procedures at lower CBI values (approximately\u0026thinsp;\u0026minus;\u0026thinsp;1.0\u0026ndash;0.4); in the intermediate range (approximately\u0026thinsp;\u0026minus;\u0026thinsp;0.4\u0026ndash;0.9), accompanied by higher BAD-D and DA-Ratio, the model favored ICL; and at very high CBI values (\u0026gt;\u0026thinsp;0.9), SHAP contributions plateaued, suggesting a diminishing marginal discriminatory value of CBI under high-risk backgrounds and a greater reliance on multiparameter patterns.\u003c/p\u003e \u003cp\u003eThe Stage 2 SHAP results described how the model distributed predictions across laser subtypes. CYL, SPH, and Dt remained among the most influential features (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e), indicating that refractive status continued to shape laser selection. Given evidence that different procedures produce different degrees of biomechanical weakening \u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e,\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e,\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e, the required preoperative biomechanical reserve may differ by procedure. Consistently, the SSI, PRFI, and ARTh ranked among the top contributors in Stage 2, supporting the use of biomechanical information beyond refractive parameters. Specifically (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e), higher SSIs and IRs contributed to SMILE, whereas higher PRFIs reduced the probability of SMILE, which is consistent with SMILE being more frequently assigned to eyes with greater stiffness, more uniform deformation responses, and lower risk signals in this cohort. For LASIK, higher CBI and TBI reduced its SHAP contributions, whereas LASIK occasionally increased under borderline biomechanical profiles (e.g., lower SSI or higher DA-Ratio). This apparent allocation pattern, where SMILE aligns with a greater biomechanical reserve, does not necessarily contradict the theoretical advantages of SMILE in preserving anterior stromal lamellae under comparable geometric conditions \u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e,\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e,\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e. Longitudinal studies suggest that postoperative biomechanical changes are strongly influenced by geometric factors such as residual stromal bed thickness and percent tissue altered; under specific parameter combinations, a thicker cap and deeper lenticule plane may yield a thinner residual stromal bed and greater biomechanical reduction \u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e. Because cap/flap/RSB data were not available in this study, this interpretation should be considered mechanism-based. In contrast, SURFACE more consistently reflected a \"conservative\" pathway: lower ARTh or IR favored SURFACE, which was consistent with avoiding flap creation and preserving anterior stromal continuity; prior work also suggested that surface ablation may preserve relatively more biomechanical reserve under certain conditions \u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e,\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e. Collectively, biomechanical and risk-related indices not only improved discrimination among laser procedures but also shaped allocation boundaries consistent with clinical risk-control principles.\u003c/p\u003e \u003cp\u003eSeveral limitations of this study should be acknowledged when interpretability analyses that elucidate model behavior. This was a single-center retrospective study, and model outputs may reflect local practice patterns. Postoperative ectasia or long-term biomechanical stability was not used as a clinical safety endpoint, and safety-related inferences should therefore be interpreted cautiously. In addition, the procedure categories were imbalanced, and the model inputs were limited to structured preoperative variables without incorporating raw imaging data or deformation waveforms. Finally, prospective validation in real-world clinical workflows was not performed. Collectively, these factors indicate that the present findings should be interpreted as model-level evidence of information gain rather than definitive clinical guidance.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis study demonstrated that a multistage AI model integrating multidimensional preoperative parameters improved the predictability of refractive surgery selection. The inclusion of corneal biomechanical indices resulted in consistent performance gains across multiple machine learning models and enabled interpretability analyses that helped contextualize how biomechanical features contributed to model predictions, supporting the nonredundant value of biomechanics beyond refractive and tomographic parameters. Although the present results do not establish a definitive role for biomechanics in defining procedural safety boundaries or resolving ambiguous surgical indications, they support its potential value as a complementary component within AI-assisted refractive surgery planning. Further multicenter prospective studies are warranted to validate the clinical applicability and generalizability of these findings.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eICL\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eImplantable Collamer Lens\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eKLEx\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eKeratorefractive lenticule extraction\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLASIK\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eLaser-assisted in situ keratomileusis\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSHAP\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eShapley additive explanations\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLASEK\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eLaser-assisted subepithelial keratectomy\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003ePRK\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003ePhotorefractive Keratectomy\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eTrans-PRK\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eTransepithelial PRK\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSURFACE\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSurface ablation (including PRK, Trans-PRK, LASEK)\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSMILE\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSmall Incision Lenticule Extraction\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eDT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eDecision Tree\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eRF\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLR\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSVM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSupport Vector Machine\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMLP\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMultilayer Perceptron\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eXGBoost\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eeXtreme Gradient Boosting\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSMOTE\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSynthetic Minority Oversampling Technique\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eAUC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eArea Under the Curve\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSRI\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSurface Regularity Index\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSAI\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSurface Asymmetry Index\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCSI\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eCorneal Symmetry Index\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eRSB\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eResidual Stromal Bed\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was approved by the Chongqing Aier Eye Hospital Medical Ethics Review Committee, Chinese Academy of Medical Sciences. All methods were carried out in accordance with relevant guidelines and regulations in the Declaration of Helsinki. All the subjects and/or their legal guardians signed an informed consent form.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the Chongqing Municipal Health Commission (Grant No. 2023MSXM122).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors' contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eStudy concept and design (YHL, GL, JHW); data collection (JHW, YHL, YLP, GL, CYX, XW); analysis and interpretation of data (YHL, JHW); writing of the manuscript (YHL, GL, CYX, JHW, YLP); critical revision of the manuscript (YHL, GL, YLP); statistical expertise (JHW, XW); administrative, technical, or material support (YLP, GL, CYX); supervision (YLP). All the authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eShuang S, Dan W, Yewei Y, Fuying Q, Huilan X, Xiaobo X. Correction of presbyopia. \u003cem\u003eJournal of Central South University Medical Sciences\u003c/em\u003e. 2022;47(10):1454.\u003c/li\u003e\n \u003cli\u003eWang Q, Fan L, Zhou Q. The best choice for low and moderate myopia patients incapable for corneal refractive surgery: implantation of a posterior chamber phakic intraocular lens. \u003cem\u003eInternational Ophthalmology\u003c/em\u003e. 2023;43(2):575-581.\u003c/li\u003e\n \u003cli\u003eSwaminathan U, Daigavane S. Comparative analysis of visual outcomes and complications in intraocular collamer lens, small-incision lenticule extraction, and laser-assisted In situ keratomileusis surgeries: a comprehensive review. \u003cem\u003eCureus\u003c/em\u003e. 2024;16(4)\u003c/li\u003e\n \u003cli\u003eChen Q, Yan H, Chen L, Wang L, Zheng Q, Ren Y. Adverse events associated with Implantable Collamer Lens: insights from the FDA MAUDE database. \u003cem\u003eFront Med-Lausanne\u003c/em\u003e. 2025;12:1613060.\u003c/li\u003e\n \u003cli\u003eDi Y, Li Y, Luo Y. Prediction of implantable collamer lens vault based on preoperative biometric factors and lens parameters. \u003cem\u003eJ Refract Surg\u003c/em\u003e. 2023;39(5):332-339.\u003c/li\u003e\n \u003cli\u003eYoo TK, Ryu IH, Lee G, et al. Adopting machine learning to automatically identify candidate patients for corneal refractive surgery. \u003cem\u003eNpj Digit Med\u003c/em\u003e. 2019;2(1):59.\u003c/li\u003e\n \u003cli\u003eTing DSW, Pasquale LR, Peng L, et al. Artificial intelligence and deep learning in ophthalmology. \u003cem\u003eBritish Journal of Ophthalmology\u003c/em\u003e. 2019;103(2):167-175.\u003c/li\u003e\n \u003cli\u003eYoo TK, Ryu IH, Choi H, et al. Explainable machine learning approach as a tool to understand factors used to select the refractive surgery technique on the expert level. \u003cem\u003eTranslational vision science \u0026amp; technology\u003c/em\u003e. 2020;9(2):8-8.\u003c/li\u003e\n \u003cli\u003eCui T, Wang Y, Ji S, et al. Applying machine learning techniques in nomogram prediction and analysis for SMILE treatment. \u003cem\u003eAm J Ophthalmol\u003c/em\u003e. 2020;210:71-77.\u003c/li\u003e\n \u003cli\u003eWang K, Xu C, Li G, Zhang Y, Zheng Y, Sun C. Combining convolutional neural networks and self-attention for fundus diseases identification. \u003cem\u003eSci Rep-Uk\u003c/em\u003e. 2023;13(1):76.\u003c/li\u003e\n \u003cli\u003eLi J, Dai Y, Mu Z, et al. Choice of refractive surgery types for myopia assisted by machine learning based on doctors\u0026rsquo; surgical selection data. \u003cem\u003eBmc Med Inform Decis\u003c/em\u003e. 2024;24(1):41.\u003c/li\u003e\n \u003cli\u003eLi L, Xiang Y, Chen X, et al. Machine learning model for predicting corneal stiffness and identifying keratoconus based on ocular structures. \u003cem\u003eIntelligent Medicine\u003c/em\u003e. 2025;5(01):66-72.\u003c/li\u003e\n \u003cli\u003eAmbr\u0026oacute;sio Jr R, Correia FF, Lopes B, et al. Corneal biomechanics in ectatic diseases: refractive surgery implications. \u003cem\u003eThe open ophthalmology journal\u003c/em\u003e. 2017;11:176.\u003c/li\u003e\n \u003cli\u003eSteinberg J, Siebert M, Katz T, et al. Tomographic and biomechanical Scheimpflug imaging for keratoconus characterization: a validation of current indices. \u003cem\u003eJ Refract Surg\u003c/em\u003e. 2018;34(12):840-847.\u003c/li\u003e\n \u003cli\u003eMoshirfar M, Motlagh MN, Murri MS, Momeni-Moghaddam H, Ronquillo YC, Hoopes PC. Advances in biomechanical parameters for screening of refractive surgery candidates: a review of the literature, part III. \u003cem\u003eMedical Hypothesis, Discovery and Innovation in Ophthalmology\u003c/em\u003e. 2019;8(3):219.\u003c/li\u003e\n \u003cli\u003eYang K, Xu L, Fan Q, Zhao* D, Ren* S. Repeatability and comparison of new Corvis ST parameters in normal and keratoconus eyes. \u003cem\u003eSci Rep-Uk\u003c/em\u003e. 2019;9(1):15379.\u003c/li\u003e\n \u003cli\u003eLiu Y, Zhang Y, Chen Y. Application of a scheimpflug-based biomechanical analyser and tomography in the early detection of subclinical keratoconus in chinese patients. \u003cem\u003eBMC ophthalmology\u003c/em\u003e. 2021;21(1):339.\u003c/li\u003e\n \u003cli\u003eEsporcatte LP, Salom\u0026atilde;o MQ, Junior NS, et al. Corneal biomechanics for corneal ectasia: Update. \u003cem\u003eSaudi Journal of Ophthalmology\u003c/em\u003e. 2022;36(1):17-24.\u003c/li\u003e\n \u003cli\u003eBorderie V, Beauruel J, Cuyaub\u0026egrave;re R, Georgeon C, Memmi B, Sandali O. Comprehensive assessment of Corvis ST biomechanical indices in normal and keratoconus corneas with reference to corneal enantiomorphism. \u003cem\u003eJ Clin Med\u003c/em\u003e. 2023;12(2):690.\u003c/li\u003e\n \u003cli\u003eDamgaard IB, Reffat M, Hjortdal J. Review of corneal biomechanical properties following LASIK and SMILE for myopia and myopic astigmatism. \u003cem\u003eThe Open Ophthalmology Journal\u003c/em\u003e. 2018;12:164.\u003c/li\u003e\n \u003cli\u003eGuo H, Hosseini-Moghaddam SM, Hodge W. Corneal biomechanical properties after SMILE versus FLEX, LASIK, LASEK, or PRK: a systematic review and meta-analysis. \u003cem\u003eBMC ophthalmology\u003c/em\u003e. 2019;19(1):167.\u003c/li\u003e\n \u003cli\u003eXin Y, Lopes BT, Wang J, et al. Biomechanical effects of tPRK, FS-LASIK, and SMILE on the cornea. \u003cem\u003eFront Bioeng Biotech\u003c/em\u003e. 2022;10:834270.\u003c/li\u003e\n \u003cli\u003eGao W, Zhao X, Wang Y. Change in the corneal material mechanical property for small incision lenticule extraction surgery. \u003cem\u003eFront Bioeng Biotech\u003c/em\u003e. 2023;11:1034961.\u003c/li\u003e\n \u003cli\u003eHashemi H, Roberts CJ, Elsheikh A, Mehravaran S, Panahi P, Asgari S. Corneal biomechanics after SMILE, femtosecond-assisted LASIK, and photorefractive keratectomy: a matched comparison study. \u003cem\u003eTranslational vision science \u0026amp; technology\u003c/em\u003e. 2023;12(3):12-12.\u003c/li\u003e\n \u003cli\u003eQu Z, Li X, Yuan Y, et al. In vivo corneal biomechanical response to three different laser corneal refractive surgeries. \u003cem\u003eJ Refract Surg\u003c/em\u003e. 2024;40(5):e344-e352.\u003c/li\u003e\n \u003cli\u003eJoshi S, Bari A, Shakkarwal C, et al. The visual outcomes and corneal biomechanical properties after PRK and SMILE in low to moderate myopia. \u003cem\u003eIndian Journal of Ophthalmology\u003c/em\u003e. 2025;73(1):128-133.\u003c/li\u003e\n \u003cli\u003eShortt AJ, Allan BD, Evans JR. Laser‐assisted in‐situ keratomileusis (LASIK) versus photorefractive keratectomy (PRK) for myopia. \u003cem\u003eCochrane Database of systematic reviews\u003c/em\u003e. 2013;(1)\u003c/li\u003e\n \u003cli\u003eJacobs DS, Lee JK, Shen TT, et al. Refractive surgery preferred practice pattern\u0026reg;. \u003cem\u003eOphthalmology\u003c/em\u003e. 2023;130(3):P61-P135.\u003c/li\u003e\n \u003cli\u003eWang X, Zhou X. Update on treating high myopia with implantable collamer lenses. \u003cem\u003eAsia-Pacific journal of ophthalmology\u003c/em\u003e. 2016;5(6):445-449.\u003c/li\u003e\n \u003cli\u003eKamiya K, Takahashi M, Takahashi N, Shoji N, Shimizu K. Monovision by implantation of posterior chamber phakic intraocular lens with a central hole (hole ICL) for early presbyopia. \u003cem\u003eSci Rep-Uk\u003c/em\u003e. 2017;7(1):11302.\u003c/li\u003e\n \u003cli\u003eLi F, Ma Y, Qi W, Pazo EE, Yang R, Zhao S. Characteristics of biological parameters and implantable collamer lens (ICL) size selection in moderate, high, and super-high myopia eyes. \u003cem\u003eBMC ophthalmology\u003c/em\u003e. 2025;25(1):103.\u003c/li\u003e\n \u003cli\u003eHERZIG S, MD, FRCSC, DABO. AN EVOLUTION OF INDICATIONS FOR THEEVO VISIAN ICL FAMILY OF LENSES. \u003cem\u003eSTAAR Surgical Company\u003c/em\u003e. 2022;\u003c/li\u003e\n \u003cli\u003eAkrobetu D, Nikpoor N. Choosing the Best Option for Refractive Surgery. \u003cem\u003eCRSToday\u003c/em\u003e. 2023;\u003c/li\u003e\n \u003cli\u003eReinstein DZ, Archer TJ, Randleman JB. Mathematical model to compare the relative tensile strength of the cornea after PRK, LASIK, and small incision lenticule extraction. \u003cem\u003eJ Refract Surg\u003c/em\u003e. 2013;29(7):454-460.\u003c/li\u003e\n \u003cli\u003eRoy AS, Dupps Jr WJ, Roberts CJ. Comparison of biomechanical effects of small-incision lenticule extraction and laser in situ keratomileusis: finite-element analysis. \u003cem\u003eJournal of Cataract \u0026amp; Refractive Surgery\u003c/em\u003e. 2014;40(6):971-980.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-ophthalmology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"boph","sideBox":"Learn more about [BMC Ophthalmology](http://bmcophthalmol.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/boph","title":"BMC Ophthalmology","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Refractive surgery, Corneal biomechanics, Machine learning, Surgical decision support","lastPublishedDoi":"10.21203/rs.3.rs-9362570/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9362570/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground: \u003c/strong\u003eCorneal biomechanical parameters are widely used in preoperative screening for refractive surgery, primarily to identify contraindications and ensure surgical safety. However, their potential role in guiding procedure selection has not been fully explored. This study aimed to evaluate the contribution of corneal biomechanical parameters beyond conventional refractive and tomographic features to refractive procedure selection.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods: \u003c/strong\u003eWe conducted a retrospective observational study of 395 patients (763 eyes) who underwent refractive surgery at Chongqing Aier Eye Hospital between October 2023 and November 2024. Preoperative data included 48 features comprising biomechanical indices, tomographic parameters, and demographic characteristics. Multiple machine learning models have been developed to predict surgical procedure selection, and their performance has been evaluated in terms of accuracy, recall, and macro-F1 score. Ablation experiments were performed to assess the contributions of corneal biomechanical parameters. Model interpretability was analyzed via Shapley additive explanations (SHAPs). Statistical analyses were conducted via the Shapiro–Wilk test, Kruskal–Wallis test, and Mann–Whitney U test with Bonferroni correction.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults:\u003c/strong\u003e The two-stage XGBoost-based model demonstrated the best performance, achieving a validation accuracy of 86.08% and an accuracy of 82.67% on the independent test set. The inclusion of corneal biomechanical parameters improved model performance, resulting in a 4% increase in accuracy on the independent test set. SHAP analysis revealed that refractive parameters contributed most prominently to model predictions, whereas corneal biomechanical features exhibited additional contributions across multiple procedures.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions:\u003c/strong\u003e An interpretable machine learning framework can support refractive procedure selection across multiple surgical options. Corneal biomechanical parameters provide complementary, nonredundant information beyond conventional features and may enhance AI-assisted clinical decision-making.\u003c/p\u003e","manuscriptTitle":"Contribution of Corneal Biomechanics to Machine Learning-Based Refractive Procedure Selection: A Retrospective Observational Study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-27 14:13:25","doi":"10.21203/rs.3.rs-9362570/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-05-18T07:44:03+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-15T07:15:39+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-12T01:34:58+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-09T20:10:16+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"27744040358238619794481355193178491366","date":"2026-04-29T10:42:29+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"326401381807696078384344670679389174640","date":"2026-04-28T09:40:56+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"273705935933720364700266380888737791890","date":"2026-04-28T09:38:23+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"48181641812040229029609374057414897440","date":"2026-04-24T02:27:18+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"149324638008734158913276424971106101744","date":"2026-04-19T01:57:42+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-04-19T01:20:38+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-04-14T07:05:50+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-13T07:14:39+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-13T07:14:11+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Ophthalmology","date":"2026-04-09T03:07:08+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-ophthalmology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"boph","sideBox":"Learn more about [BMC Ophthalmology](http://bmcophthalmol.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/boph","title":"BMC Ophthalmology","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a794b9dd-ae78-4b8f-9495-ee0a78ce78da","owner":[],"postedDate":"April 27th, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Revision requested","date":"2026-05-18T07:44:03+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-15T07:15:39+00:00","index":61,"fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-12T01:34:58+00:00","index":60,"fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-09T20:10:16+00:00","index":59,"fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"in-revision","subjectAreas":[],"tags":[],"updatedAt":"2026-05-18T07:54:49+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-27 14:13:25","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9362570","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9362570","identity":"rs-9362570","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}