Early Detection of Alzheimer's Disease Using Explainable Machine Learning on Clinical Biomarkers: A Multi-Class Classification Study Using the Alzheimer's Disease Neuroimaging Initiative (ADNI) Dataset
An explainable XGBoost model used routine clinical measures to classify normal cognition, MCI, and Alzheimer's disease with high accuracy on ADNI data.
Intelligence analysis by GPT-5.4 Mini
The paper reports a three-class Alzheimer's detection model built on eight common clinical features from the ADNI dataset. Using XGBoost, SMOTE, Optuna tuning, and SHAP explanations, it claims near-perfect separation of normal cognition, mild cognitive impairment, and Alzheimer's disease.
The paper tries to teach a computer to spot memory problems using regular doctor test results, like scorecards and simple questions. It says the computer became very good at telling apart healthy people, people with early memory trouble, and people with Alzheimer's, like sorting colored marbles into three jars.
Analysis
What the paper did
The study targets a practical screening problem: distinguishing normal cognition (NC), mild cognitive impairment (MCI), and Alzheimer's disease (AD) from routine clinical information. It uses eight features from the Alzheimer's Disease Neuroimaging Initiative (ADNI): MMSE, CDR Global, CDR Sum of Boxes (CDR-SB), MoCA, FAQ, age, sex, and education.
Method and evaluation
The model is an XGBoost classifier tuned with Optuna over 50 trials. To address class imbalance, the author applied SMOTE. Performance was measured with macro AUC-ROC, macro F1, balanced accuracy, and Cohen's kappa, including 1,000-iteration bootstrap confidence intervals.
Results
The dataset included 1,641 baseline subjects: 608 NC, 767 MCI, and 266 AD. On five-fold cross-validation, the paper reports a mean macro AUC of 0.983, accuracy of 0.944, and macro F1 of 0.929. On a held-out test set of 247 subjects, macro AUC was 0.982, accuracy 0.943, balanced accuracy 0.932, macro F1 0.927, and Cohen's kappa 0.909.
Explainability
SHAP analysis is used to show which inputs drive the model. The paper says CDR Global was the main predictor for NC versus MCI, while CDR-SB and MMSE were most important for AD classification. That feature pattern is presented as supporting clinical plausibility, not just predictive performance.
What it claims next
The conclusion frames the model as an explainable system trained on routine clinical assessments that can detect three classes of cognitive status with near-perfect performance. The future direction is to add speech biomarkers for multimodal detection.
Key points
- The paper builds a three-class classifier for NC, MCI, and AD using eight routine clinical features from ADNI.
- It uses XGBoost with Optuna tuning, SMOTE for imbalance, and SHAP for explainability.
- The reported test-set performance is very high, with macro AUC of 0.982 and accuracy of 0.943.
- CDR Global, CDR-SB, and MMSE are highlighted as the most important predictors.
- The author says the next step is adding speech biomarkers for multimodal detection.
If the findings generalize, clinics could use low-cost, routine cognitive measures to flag Alzheimer's earlier and more consistently. The SHAP results may also make the system easier for doctors to trust because the model's decisions line up with familiar clinical markers.
The strong results come from a single research dataset, so performance may drop on different hospitals, populations, or real-world screening settings. The study also relies on features that are already fairly diagnostic, which could make the model less novel or less useful outside the ADNI context.
