Artificial Intelligence for Alzheimer’s Disease Detection Across Neuroimaging and Speech: A Reproducible Cross-Modal Secondary Analysis

Artificial-intelligence (AI) systems for Alzheimer’s disease (AD) increasingly use neuroimaging or spontaneous speech, but reported performance is difficult to interpret across studies because cohorts, endpoints, languages, preprocessing, and validation protocols differ. We applied a systematic coding framework to a predefined 51-reference evidence corpus and performed a conservative secondary quantitative synthesis. Twenty-six references met prespecified criteria for primary empirical AI studies: 7 neuroimaging and 19 speech/language studies. Fourteen studies contributed 15 model-level benchmark entries with extractable headline accuracy. Neuroimaging entries ranged from 66.25–99.95% (median 97.74%); speech/language entries ranged from 78.70–92.08% (median 83.81%). These values are descriptive and are not pooled diagnostic-effect estimates. Only 5/26 primary studies (19.2%) explicitly evaluated external, cross-domain, or cross-language generalization, 5/26 (19.2%) used explicit multimodal input, and 4/26 (15.4%) emphasized explainability or uncertainty. Multilingual or cross-lingual evaluation appeared in 3/19 speech/language studies (15.8%). The evidence therefore supports a distinction between high within-study accuracy and clinically credible transportability. Independent validation, patient-disjoint evaluation, calibrated uncertainty, representative cohorts, and linguistically diverse testing should be treated as core design requirements rather than optional extensions. The study-level coding table and benchmark extract are provided for reproducibility and updating.