Incremental contribution of Corvis ST dynamic biomechanical parameters to interpretable machine learning prediction of clinician-selected refractive procedures: a retrospective observational study

Abstract Background This study aimed to evaluate whether structured preoperative variables can reproduce clinician-selected refractive procedure patterns and to assess the incremental contribution of Corvis ST dynamic biomechanical parameters beyond conventional refractive, tomographic, pachymetric, and risk-related features. 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. The outcome label was the procedure actually selected and performed by clinicians, including ICL, SMILE, LASIK, and SURFACE. Forty-eight structured preoperative features were analyzed, including demographic, refractive, visual, ocular-surface, tomographic, pachymetric, anterior-segment, Pentacam-derived deviation/risk, and Corvis ST-derived variables. Data were split at the patient level into a training cohort and an internal held-out test cohort. Multiple machine learning models were developed using patient-grouped cross-validation, and the final model was evaluated using accuracy, balanced accuracy, macro-F1 score, class-wise metrics, calibration analysis, patient-level clustered bootstrap resampling, and one-eye-per-patient sensitivity analysis. Structured feature-set ablation and SHAP analysis were performed to evaluate feature-domain contributions and model behavior. Results The two-stage XGBoost-based model, which first separated ICL from corneal laser procedures and then classified laser procedures into SMILE, SURFACE, and LASIK, achieved the most favorable overall performance. Its estimated total validation accuracy was 83.36% ± 2.59%. On the internal held-out test cohort, the model achieved an end-to-end accuracy of 78.43%, balanced accuracy of 79.17%, macro-F1 score of 77.96%, and macro AUC of 93.77%. Ablation analysis showed that most predictive gain was derived from tomographic and pachymetric variables, whereas pure Corvis ST dynamic biomechanical parameters provided modest complementary information beyond the full non-pure-Corvis feature set. SHAP analysis indicated that refractive parameters were the dominant contributors, while tomographic, pachymetric, risk-related, and Corvis ST-derived variables contributed to procedure-specific model behavior. Conclusions Structured preoperative variables can partially reproduce single-center clinician-selected refractive procedure patterns. Corvis ST-derived dynamic biomechanical parameters suggested modest complementary predictive information beyond conventional refractive, tomographic, pachymetric, and risk-related features. These findings support internally validated prediction of clinician-selected procedure patterns but do not establish optimal surgical recommendation or external generalizability.