Construction of an auxiliary diagnostic model for Alzheimer’s disease based on routine peripheral blood parameters

This study aimed to construct clinical auxiliary diagnostic models for Alzheimer’s disease (AD) using multiple machine learning algorithms based on routine peripheral blood parameters and to identify the optimal model. This retrospective study included 84 patients diagnosed with AD and 250 healthy controls from physical examination populations between January and December 2025. Data on age, sex, and various blood test results were collected. The dataset was divided into a training set (n = 233) and a validation set (n = 101) at a 7:3 ratio. Variables were first subjected to differential analysis in the training set, with variables meeting P < 0.05 selected. Subsequently, feature selection was performed using Boruta algorithm combined with LASSO regression. Eight machine learning algorithms were employed to construct models using the selected variables: Random Forest (RF), Gradient Boosting Machine (GBM), XGBoost, AdaBoost, K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Logistic Regression, and Neural Network (NN), with hyperparameter tuning performed. DeLong test was used to compare AUC differences between models, Brier score was employed to assess model calibration, and decision curve analysis (DCA) was conducted to evaluate clinical utility. Finally, the final model was implemented as a freely accessible web-based calculator (https://ad-model.netlify.app) to facilitate clinical application. Boruta algorithm and LASSO regression jointly identified HDL, HGB, BASO, NLR, and SIRI as candidate features; NLR was subsequently removed during backward stepwise regression, yielding a final Logistic Regression model comprising HDL, HGB, BASO, and SIRI. Among the eight machine learning models, Random Forest achieved the highest validation set AUC (0.897). DeLong test showed no statistically significant differences in AUC among models on the validation set ( P > 0.05). Considering model robustness, Logistic Regression was ultimately selected as the final model. This model achieved an AUC of 0.896 in the training set and 0.882 in the validation set, with Brier scores of 0.101 and 0.102 in the training and validation sets, respectively. DCA analysis demonstrated that using the machine learning model yielded positive net benefit within the threshold probability range of 2%-98%. The clinical auxiliary diagnostic model constructed based on peripheral blood HDL, HGB, BASO, and SIRI demonstrates good discrimination, calibration, and clinical applicability, providing an objective and reliable decision support tool for auxiliary diagnosis of AD.