Multiclass neurodegenerative disease classification framework based on explainable ensemble learning using cerebrospinal fluid proteomic data

Early and accurate stratification of Alzheimer Disease (AD) is a key requirement in the field of neuroscience. This article presents a machine learning framework for multiclass AD classification across three classes: Control, Asymptomatic Alzheimer’s Disease (ASYMAD), and Alzheimer’s Disease using Cerebrospinal Fluid (CSF) proteomics data. A publicly available dataset comprising 393 CSF samples with 62 synaptic peptide features was utilized in this study. We used six base classifiers—Elastic Net, Random Forest, XGBoost, Linear SVM, LightGBM and CatBoost—trained and evaluated with nested stratified 5 fold cross validation with inner loop randomized hyperparameter search. Based on their complementary linear and nonlinear inductive biases, an ensemble consisting of Elastic Net, CatBoost, and Linear SVM as base learners with a logistic regression meta learner was designed, and was benchmarked against an equal weight soft voting ensemble. The stacked ensemble achieved the highest point estimate for accuracy of \(0.832 \pm 0.052\) and the highest macro one-versus-rest ROC-AUC of \(0.952 \pm 0.012\) , with a macro F1 of \(0.826 \pm 0.059\) ; however, its improvement over the best single model , Elastic Net at macro \(\hbox {F1} = 0.827\) , and over the simpler soft voting ensemble at macro \(\hbox {F1} = 0.828\) was small and not statistically significant after a Friedman omnibus test and Bonferroni corrected pairwise Wilcoxon tests. Model calibration was additionally quantified using reliability diagrams, the multiclass Brier score, and Expected Calibration Error. Through eXplainable SHAP analysis consensus candidate peptide features were identified that are consistently selected across Elastic Net, Random Forest, and XGBoost, including ELDVLQGR, AQALEQAK, and ELQAAQAR, offering biologically interpretable signals that warrant independent validation before clinical use.