A Robust and Trustable Approach to Incorporate Medical Domain Knowledge in Machine Learning Models for Diabetic Retinopathy Screening Using Routine Lab Results
文献摘要
Many AI models for disease prediction remain difficult to integrate into clinical practice because model complexity can limit transparent interpretation. This study develops a domain-informed, low-dimensional machine-learning framework for diabetic retinopathy (DR) screening using routinely collected electronic health record and laboratory data. Medical-domain information was extracted from PubMed using a Term Frequency-Relative Citation Ratio (TF-RCR) procedure, and the selected feature set was constrained to remain cognitively manageable. The final model uses eight variables - age, glucose, nephropathy, neuropathy, HbA1c, hemoglobin, albumin, and creatinine - and its predictions are examined using post-hoc SHapley Additive exPlanations (SHAP) feature attribution. The data contained a 97,786-patient cohort, including 16,865 DR cases (17.25%). After a 64%/16%/20% training/validation/test split, SMOTE was applied only to the training subset, while validation and test sets retained the observed prevalence. On the untouched test set (n = 19,558), the eight-feature XGBoost model achieved 95.35% accuracy, 91.17% sensitivity, 96.22% specificity, an F1-score of 87.12%, AUROC of 97.24%, and AUPRC of 94.55%. Same-cohort comparisons with logistic regression, a 24-feature XGBoost model, a multilayer perceptron (MLP), and an FT-Transformer showed that the eight-feature model retained competitive discrimination while using a substantially smaller feature set. The framework provides a reproducible approach for combining literature-derived knowledge with empirical data while keeping the final model concise and its feature contributions available for post-hoc interpretation.