Explainable AI for Early Alzheimer’s Risk Prediction
Details:-
Zenesys × IIT Delhi — An advanced, explainable ML framework for early Alzheimer’s risk detection using routine clinical data,
designed for India’s healthcare context. Explore the Live Demo View Full Case Study
Model Performance at a Glance:
XGBoost outperformed Logistic Regression, Random Forest, and SVM across all seven evaluation metrics with Wilcoxonvalidated statistical significance
A Research Partnership Built for India’s Reality:
IIT Delhi, one of India’s premier academic and research institutions, partnered with Zenesys to build an explainable ML framework for early Alzheimer’s risk detection. Zenesys served as the AI/ML engineering partner, designing a scalable, transparent solution using non-invasive, routinely available data.
Why Early Alzheimer’s Goes Undetected in India:
Diagnostic Barriers
Neuroimaging and CSF biomarkers are costly and invasive not viable for population-scale screening. Standard cognitive tests like MMSE lack sensitivity for early-stage detection.
Research Gaps
Many ML studies rely on single holdout validation, risking data leakage and optimistic bias. Black-box models reduce clinical trust and hinder adoption.
The Framing Question
How can an India-focused AI framework use routine clinical, functional, and behavioral data to reliably flag Alzheimer’s risk — while remaining transparent, statistically robust, and clinically aligned?
Data & Feature Strategy: No Scans, No Invasive Tests:
The framework uses only routinely collected, non-invasive data — conceptually suitable for PHC and district-level settings across India.
1. Cognitive Markers
MMSE scores and memory complaints — top SHAP predictors of Alzheimer’s risk.
2. Functional Status
Activities of Daily Living (ADL) and functional assessment highest mean SHAP values.
3. Behavioral & Vitals
Sleep quality, physical activity, diet, cholesterol, blood pressure, BMI, age, and cardiovascular history.
4. No Invasive Diagnostics
No imaging. No CSF biomarkers. Stratified sampling preserves class proportions with no artificial feature engineering.
Modeling & Evaluation Framework:
Four classifiers — Logistic Regression, Random Forest, SVM, and XGBoost were benchmarked usingstratified crossvalidation. XGBoost emerged as the top performer across all metrics.
Nested cross-validation with randomized hyperparameter search eliminated overfitting and data leakage, achieving an outerfold AUC of 0.9496.
Explainable AI with SHAP: Transparent by Design
Zenesys integrated SHAP to make predictions transparent at both global and individual levels — aligning model reasoning with clinical criteria.
Statistical Rigor: Validated Beyond Standard Practice
Repeated stratified cross-validation (5 folds × 10 repeats = 50 evaluations) provided robust statistical power. Wilcoxon signed-rank tests confirmed XGBoost’s superiority across all baselines.
The Alzheimer’s Risk Predictor: Live on Hugging Face:
This interactive demo showcases real-world applicability as a clinical
decision support concept. It delivers instant risk assessment using noninvasive parameters — accessible at PHC and district hospital level.
Research Contributions & Future Directions:
This interactive demo showcases real-world applicability as a clinical decision support concept. It delivers instant risk assessment using non invasive parameters — accessible at PHC and district hospital level.
