Interpretable machine learning model based on blood parameters for screening high myopia.

OBJECTIVE

To develop an interpretable machine learning (ML) model using routine blood parameters for high myopia (HM) screening as a convenient and cost-effective alternative to traditional methods.

METHODS

This cross-sectional study enrolled 313 participants (158 HM and 155 non-HM). Blood parameters were comprehensively analysed, and features were selected via univariate analysis and Lasso regression. Eight ML algorithms were trained and validated using the selected features with bootstrap resampling to develop an HM diagnostic model. Model performance was evaluated by area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1 score, calibration curve, and decision curve analysis (DCA). In addition, parameter optimisation and model validation were conducted using fivefold cross-validation. Feature importance was analysed using Shapley additive explanation (SHAP).

RESULTS

Eight key indicators were identified through feature selection, including direct bilirubin (DBIL), total bilirubin (TBIL), albumin (ALB), alkaline phosphatase (ALP), age, glucose (GLU), creatinine (CREA), and uric acid (UA). Extreme Gradient Boosting (XGBoost) was found to be the optimal model. In the fivefold cross-validation, the AUC values of the training set, validation set, and test set were 0.954, 0.822, and 0.898, respectively. SHAP analysis was performed to determine the contribution of the eight variables to the model and their relationship with the occurrence of HM. The model demonstrated good calibration and clinical utility, as evidenced by DCA.

CONCLUSION

This study validates the feasibility of a blood-based ML model for screening HM, providing an accessible and interpretable tool for early detection in resource-limited settings.