Alzheimer's Classification — Master's Thesis Research

Problem

Early cognitive decline is easy to miss in routine care. The research question: can a small set of clinical and neuroimaging biomarkers reliably separate demented from non-demented patients in a cross-sectional snapshot, and which algorithms balance accuracy with recall for disease detection?

This was not a production deployment — it was thesis-grade methodology: explicit preprocessing choices, imputation strategies compared side by side, and nine models tuned through grid search with held-out test evaluation.

Dataset & features

Source: OASIS Longitudinal Dataset — 150 subjects at first clinical visit, 8 predictive variables.

Category Features
Clinical Gender, age, years of education (EDUC), socioeconomic status (SES)
Cognitive MMSE (Mini-Mental State Examination)
Neuroimaging Estimated intracranial volume (eTIV), normalized whole brain volume (nWBV), atlas scaling factor (ASF)

MMSE and brain volume metrics showed the strongest separation between groups — MMSE ranges clustered around 25–30 for non-demented vs 17–30 for demented cohorts.

Methodology

  1. Selection: First-visit rows only for cross-sectional analysis.
  2. Encoding: Binary gender; dementia label standardized to binary target.
  3. Missing values: Strategy A — drop rows with missing SES (8 rows). Strategy B — EDUC-stratified median imputation.
  4. Scaling: MinMax normalization on training folds.
  5. Split: 75% train/validation (5-fold CV), 25% held-out test; random_state=0 for reproducibility.

Models evaluated

Nine distinct approaches with grid-search hyperparameters:

  • Logistic Regression (with and without imputation)
  • SVM — RBF, linear, polynomial, sigmoid kernels
  • Decision Tree (max depth search)
  • Random Forest (estimators, features, depth)
  • AdaBoost (estimators, learning rate)

Metrics: Accuracy, recall (dementia sensitivity), ROC-AUC, confusion matrices. Feature importance from tree-based models; Graphviz export of optimal decision tree structure.

Results & insights

  • Class imbalance required emphasizing recall, not accuracy alone — missing dementia is costlier than a false alarm in screening context.
  • MMSE dominated feature importance across tree ensembles; neuroimaging ratios added signal but cognitive score carried most discriminative power.
  • SVM and ensemble methods (Random Forest, AdaBoost) competed on AUC; logistic regression anchored interpretability.
  • Imputation strategy materially shifted performance — documenting both paths was essential for thesis rigor.

Lessons

  1. Preprocessing is the experiment. Imputation vs complete-case analysis is a scientific choice, not a footnote.
  2. Recall is the clinical metric. Optimize for the error you cannot afford.
  3. Interpretability has value. Decision trees and feature importance charts support clinician conversation even when ensembles win on AUC.
  4. Reproducibility is non-negotiable. Fixed random seeds and explicit train/test walls keep thesis results defensible.

Source

Notebook, methodology, and dataset notes: github.com/ax5hay/AlzheimersDiagnosis