Explainable ML nails three-class Alzheimer's detection

An explainable AI model detects Alzheimer's stages from eight routine tests.

In a new study, researchers show that an explainable machine learning model can separate normal cognition, mild cognitive impairment, and Alzheimer's disease with near-perfect accuracy using everyday clinical assessments. The team trained an XGBoost classifier on eight features drawn from the ADNI dataset: MMSE, CDR Global, CDR Sum of Boxes, MoCA, FAQ, age, sex, and education. They optimized hyperparameters with Optuna over 50 trials and addressed class imbalance with SMOTE. The study evaluated performance with macro AUC, macro F1, and other metrics, offering a rare blend of accuracy and interpretability in a clinical context. The Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset comprised 1,641 baseline subjects, including 608 with normal cognition, 767 with mild cognitive impairment, and 266 with Alzheimer's disease.

Benchmarks indicate strong, robust performance. In five-fold cross validation, the model achieved a mean macro AUC of 0.983 with a small standard deviation of 0.007, accuracy of 0.944, and macro F1 of 0.929. On the held-out test set of 247 individuals, macro AUC held steady at 0.982 (95 percent confidence interval 0.965 to 0.995), with an accuracy of 0.943, balanced accuracy of 0.932, macro F1 of 0.927, and Cohen’s kappa of 0.909. The authors report that SHAP values provide feature-level explainability, revealing clinically plausible patterns behind each class. Specifically, CDR Global is the dominant predictor for NC and MCI, while CDR-SB and MMSE jointly drive AD classification.

The study is notable not just for accuracy but for its emphasis on interpretability in a clinical setting. The team reports that SHAP analysis yields class-specific feature importance that aligns with conventional clinical intuition, which matters when a clinician must trust a software-assisted decision. By anchoring decisions to eight routine assessments, the model demonstrates how current practice data can serve as a strong signal when processed through modern, transparent ML methods. The authors conclude that their framework achieves near-perfect three-class detection and anticipate extending it with speech biomarkers to enable multimodal detection in the future.

From an engineering standpoint, several takeaways matter for product teams. First, the use of SMOTE to address class imbalance, combined with a compact feature set, points to a practical path for building scalable, hospital-friendly tools without a data collection burden beyond standard intake assessments. Second, the SHAP-based explanations are not ornamental; they translate model logic into clinically meaningful cues, helping practitioners validate and operate the model with confidence rather than as a black box. Third, while the results are striking, the data come from the ADNI cohort, a research registry with multi-site imaging and clinical protocols; deployment in diverse real-world clinics will require external validation to guard against population or measurement biases. Fourth, the plan to add speech biomarkers signals a move toward multimodal, more resilient detection, but it also raises integration challenges around data collection, privacy, and latency in routine visits.

Looking ahead, the paper’s benchmarks set a high bar for clinically grounded AI in cognitive disease. The next wave will test whether this level of interpretability and performance translates to routine care, and if multimodal inputs can push accuracy even higher without sacrificing trust or workflow efficiency. In the meantime, the ADNI-based results offer a clear blueprint: strong performance can coexist with transparent decision-making when practitioners and engineers align on features, explanations, and validation paths.