Bridging the Generalization Gap in AI-Based Neurodegenerative Disease Detection: A Systematic Review with Reproducibility and Validation Insights

Abstract

Artificial intelligence shows strong potential for early detection of neurodegenerative diseases (NDDs) using medical imaging and multimodal data. However, many models suffer from limited generalizability on independent external cohorts. This systematic review and meta-analysis (PRISMA 2020) of 152 deep-learning studies ($2020-2025$) quantifies the performance gap: a consistent drop of $\mathbf{8} \boldsymbol{-} \mathbf{1 5}$ percentage points in AUC occurs between internal and external validation. Patientselection bias and reliance on curated cohorts (ADNI, PPMI) are the main contributors (QUADAS-2). Multimodal fusion achieves the highest pooled AUC (0.96) but remains fragile in real-world settings. The study calls for prioritizing reproducibility, external validation and clinical robustness over architectural novelty.