Interpretable Alzheimer’s Disease Detection Via Multi-Scale Fusion of Disentangled Speech Features

Spontaneous speech has emerged as a promising biomarker for the non-invasive detection of Alzheimer’s disease (AD). Existing approaches rely on theory-based acoustic features, which are interpretable but may incompletely capture AD-related speech patterns, and on high-dimensional deep learning representations, which are expressive less interpretable. To address these limitations, this study proposes an interpretable AD detection framework based on multi-scale fusion of disentangled speech representations. A neural audio codec is employed to decompose the speech signal into three interpretable attributes: content, prosody, and timbre. These representations are integrated with sentence-level and global linguistic embeddings through a Graph Attention Network (GAT), enabling the modeling of complex interdependencies across multiple scales. Experimental results demonstrate that our proposed method achieves performance comparable to state-of-the-art models on cross-cultural and cross-linguistic datasets, attaining accuracies of 89.6%, 85.9%, and 95.4% on the ADReSS, ADReSSo, and a Chinese dataset, respectively. Feature analysis further indicates that timbre provides a discriminative signal comparable to linguistic features, suggesting its potential as a biomarker. The results highlight that fusing disentangled, multi-scale speech representations can improve both the performance and interpretability of automated AD detection systems.