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

OBJECTIVE

To predict patient satisfaction 6 months after small-incision lenticule extraction (SMILE) surgery from preoperative clinical data demonstrating the methodological utility of open-source orange data mining software (Orange) feature selection and machine learning (ML) for practice-based evidence.

DESIGN

A retrospective study.

PARTICIPANTS

Seventy-eight patients who had undergone bilateral uncomplicated SMILE surgery at a private ophthalmology clinic for treatment of their myopic refractive error with complete records available for analysis.

METHODS

Open-source data mining software Orange was applied to the data. Ten preoperative features (predictors) were investigated. The target (predicted variable) was patient expectations met or exceeded 6 months postoperatively. To simplify downstream clinical interpretation, continuous metrics were discretized using a median-split protocol. Fast correlation-based filter (FCBF) feature selection and Naïve Bayes supervised ML were applied. Optimistic and realistic ML performance were expressed as the area under the receiver operating characteristic (AUC) and Matthews' correlation coefficient (MCC). A nomogram generated log-odds ratios for estimating how each feature influenced the probability of exceeded expectations.

MAIN OUTCOME MEASURES: Primary outcome was ML performance assessed using AUC and MCC.

RESULTS

Only 4 clinically relevant features (mean sphere equivalent, sex, quality of vision [QoV] night, corneal staining) were selected using FCBF. Comparable ML performance arose using these 4 features (optimistic: AUC = 0.792, MCC = 0.441; realistic 10-fold cross-validation: AUC = 0.749, MCC = 0.335) or all 10 (optimistic: AUC = 0.778, MCC = 0.447; realistic: AUC = 0.679, MCC = 0.329). Similarity between optimistic and realistic performance indicated sample sufficiency and model stability. The probability of a patient's expectations being exceeded increased by 13.68% for lower magnitude MSE (≥ -3.875 D), 15.15% for males, 10.17% for QoV night scores ≥9, and 8.73% when corneal staining was present, with Orange restricting log-odds measurements to point estimates.

CONCLUSIONS

Open-source ML software was applied for preoperative prediction of patient satisfaction following SMILE surgery. By using feature selection to ensure feature independence, ML enables clinicians to identify the most relevant preoperative factors, improving surgical counselling and patient expectation management. Reducing redundant clinical tests could optimize preoperative screening, making no-code visual ML decision support tools valuable and accessible in real-world clinical settings.