Abstract:AIM: To develop and validate a clinician-friendly logistic regression prediction model for self-reported visual impairment (VI) in middle-aged and older adults (≥45y) with diabetes. METHODS: Leveraging data from the China Health and Retirement Longitudinal Study (CHARLS), a model for VI among adults aged ≥45y with diabetes were developed. Feature selection involved LASSO regression and subsequent multivariable logistic regression. Eight machine learning algorithms were explored and compared for predictive performance. Logistic regression for its consistent performance, interpretability, and clinical usability was finally selected. A nomogram and interactive web-based tool were constructed to facilitate application. RESULTS: Totally 1918 participants (45.83% males) in CHARLS 2011 with aged ≥45y were analyzed in the training cohort and 1553 in CHARLS 2015 were in validation cohort. Among all participants in the training cohort, 39.6% reported VI. Seven variables were found to be independently associated with VI. The optimal model, logistic regression model, achieved area under the curve (AUC) of 0.702 and 0.706 in the training and validation cohorts, respectively. The model’s potential for clinical application was supported by calibration and decision-curve analyses; the resulting nomogram and web calculator provided individualized risk prediction. CONCLUSION: We developed a clinically interpretable logistic regression model to predict the risk of VI in adults aged ≥45y with diabetes. The accompanying nomogram and web tool may assist with early identification and targeted vision care.
