A Multidomain Model for Dementia Classification using Harmonized LASI and LASI-DAD Data

Technical Summary: A Multidomain Model for Dementia Classification using Harmonized LASI and LASI-DAD Data

Problem Statement
Dementia classification in heterogeneous populations, particularly in low- and middle-income settings like India, is complicated by the influence of education, language, socioeconomic status, and health on cognitive test performance. Relying on fixed cognitive thresholds or isolated predictor sets often leads to inconsistent performance and systematic misclassification across diverse older adult populations. Furthermore, dementia-related decline involves multiple biological and functional domains (vascular, metabolic, functional) that are rarely captured by task-based cognitive assessments alone. The challenge is to develop a predictive model that integrates these multidomain factors to improve discrimination while remaining interpretable, specifically within the context of the Longitudinal Ageing Study in India (LASI) and its diagnostic sub-study (LASI-DAD), where diagnostic uncertainty and data heterogeneity are significant.

Methodology
The study utilized harmonized data from LASI and LASI-DAD, resulting in an analytic sample of 3,186 participants (excluding those with Mild Cognitive Impairment to define a binary boundary between dementia and cognitively unimpaired status).

• Outcome Definition: The binary dementia outcome was derived from consensus-based Clinical Dementia Rating (CDR) assessments. To address diagnostic uncertainty inherent in large-scale datasets, the authors averaged 20 multiply imputed versions of the binary dementia indicator provided by the LASI-DAD protocol. A threshold of 0.5 was applied to this averaged value to assign the final training label.
• Predictors: A multidomain set of 22 predictors was retained, spanning four domains:
  1. Cognitive Performance: MMSE orientation, immediate/delayed word recall, story recall, clock drawing, animal fluency, and composite scores for memory and executive function.
  2. Informant-Reported Decline: The Informant Questionnaire on Cognitive Decline in the Elderly (IQCODE).
  3. Cardiometabolic Biomarkers: HbA1c, lipid profiles (total, HDL, LDL, triglycerides), hsCRP, homocysteine, Vitamin B12, Vitamin D, and albumin.
  4. Sociodemographics: Age, sex, and educational attainment.
     Note: The Hindi Mental State Examination (HMSE) was excluded to avoid measurement inequivalence due to regional linguistic variations in India.
• Data Preprocessing: Missing values were imputed using k-nearest neighbours (KNN, k=5). Categorical variables were ordinally encoded, and continuous predictors were standardized. All preprocessing steps were fitted exclusively within the training pipeline to prevent data leakage.
• Model Development: The dataset was split 70:30 (stratified). Class imbalance (approx. 13.6% dementia cases) was addressed using Synthetic Minority Oversampling Technique (SMOTE) applied only within training folds. Five supervised learning algorithms were evaluated: Logistic Regression, Random Forest, Gradient Boosting, XGBoost, and Support Vector Machines (SVM).
• Validation Strategy: Nested cross-validation (3-fold inner loop for hyperparameter tuning, 5-fold outer loop for performance estimation) was conducted within the training set. The final model was evaluated on the held-out test set.
• Interpretability: Permutation feature importance and SHapley Additive exPlanations (SHAP) were used to analyze feature contributions.

Key Results

• Model Performance: Logistic Regression emerged as the optimal model, achieving a ROC-AUC of 0.932 and an Average Precision of 0.668 on the held-out test set. This outperformed ensemble methods (Random Forest: 0.909, Gradient Boosting: 0.912, XGBoost: 0.897, SVM: 0.904).
• Multidomain Value: A "cognition-only" comparator model (restricted to task-based cognitive measures) achieved a ROC-AUC of 0.908. The full multidomain model demonstrated incremental discriminatory value, confirming that non-cognitive features add structure to the classification.
• Classification Metrics: At an optimal probability threshold of 0.70, the model achieved a sensitivity of 0.771, specificity of 0.905, and a Negative Predictive Value (NPV) of 0.985. The Positive Predictive Value (PPV) was 0.325, reflecting the challenges of positive prediction in imbalanced datasets.
• Feature Importance: Permutation importance and SHAP analyses identified informant-reported decline (IQCODE) and orientation as the strongest predictors. Cardiometabolic variables provided smaller but consistent incremental contributions.
• Risk Stratification: The model generated a graded risk distribution where dementia prevalence increased progressively across strata, reaching approximately 50% in the highest risk category.

Significance and Claims
The paper claims that dementia classification in a socially and clinically heterogeneous Indian cohort can be improved by integrating cognitive, informant, cardiometabolic, and sociodemographic information within a single interpretable model.

• Interpretability vs. Complexity: The study demonstrates that a transparent logistic regression model can match or exceed the performance of complex ensemble methods while offering greater interpretability, which is crucial for clinical reasoning.
• Role of Informants: The findings highlight that informant-reported measures (IQCODE) carry a stronger predictive signal than individual cognitive tasks in this population. This is attributed to the informant's ability to capture functional decline relative to an individual's baseline, rather than performance against population norms that may be biased by educational disadvantage.
• Multidomain Necessity: The results support the view that dementia is a multidomain syndrome where vascular and metabolic factors, though not dominant individually, provide necessary structural context to the core cognitive signal.
• Operational Utility: The high NPV suggests the model is effective for ruling out dementia in resource-limited settings, while the risk gradient allows for prioritizing high-risk individuals for specialist evaluation.

Limitations and Future Directions
The authors explicitly state that the model is internally validated and requires external validation and calibration before broader application. Key limitations include:

• The cross-sectional nature of the data prevents establishing causal direction or modeling progression.
• The outcome relies on imputed data, introducing residual uncertainty near the classification boundary.
• The exclusion of survey weights limits the generalizability of prevalence estimates, though the authors argue this is acceptable for predictive modeling.
• Formal calibration analysis (e.g., Brier score) was not performed, meaning the model is currently best understood as a ranking tool rather than a source of reliable individual probability estimates.

The paper concludes that while the approach successfully accommodates complexity and remains interpretable, future work must prioritize external validation in independent cohorts, formal calibration, and the incorporation of longitudinal data to model risk trajectories.