Alzheimer's Disease Detection via Large Pre-Trained Speech and Language Models: Taxonomy, Progress, and Open Challenges
Abstract
Alzheimer's Disease (AD) detection from spoken language has advanced considerably with the rise of large pre-trained speech and language models, including wav2vec 2.0, BERT, generative LLMs, and emerging audio-LLMs. These models often report higher performance on shared AD detection benchmarks than handcrafted-feature pipelines, while reducing the need for manual feature design. Yet their advantages remain difficult to interpret: pre-trained representations are largely a black box, benchmark gains are not always statistically robust, and performance can vary substantially across languages, elicitation protocols, and clinical task definitions.
This survey proposes a taxonomy of how large pre-trained speech and language models have been used for AD detection and progression-related tasks. We define Large Pre-trained Models (LPMs) as Transformer-based models trained with self-supervised or weakly supervised objectives, distinguishing them from earlier non-Transformer pre-trained representations. We then organize the literature into six usage strategies: direct application, fine-tuning, transfer learning, multitask learning, multilingual modeling, and generative-LLM-based assessment. Across these categories, we synthesize evidence from shared benchmarks and recent studies to answer four questions: 1) what taxonomy captures how LPMs are used for AD detection; 2) what progress and advantages do LPMs offer for AD detection; 3) what challenges and limitations do LPMs exhibit; and 4) what future opportunities and open tasks they present.
Related
- TREM2 Signaling and Alzheimer's Disease: Experimental Modulators, Progress, and Challenges.
- A neurocognitive speech taxonomy for voice biomarkers of Alzheimer's disease
- A Taxonomy of Mental Health and Technology Needs for Alzheimer’s and Dementia Caregivers
- The application progress of artificial intelligence in the diagnosis of Alzheimer’s disease
- The application value of artificial intelligence in medical imaging: foundations, progress, and challenges.
BibTeX
@misc{li2026alzheimers,
title = {{Alzheimer's Disease Detection via Large Pre-Trained Speech and Language Models: Taxonomy, Progress, and Open Challenges}},
author = {Yuan-Chao Li and Yilin Pan and Yin-Long Liu and Liu He and Qi Sun and Siyu Zhang and Wei-Qiang Zhang and Saturnino Luz and Zhenhua Ling and Jiahong Yuan},
howpublished = {PsyArXiv},
year = {2026},
month = sep,
doi = {10.31234/osf.io/s57tz_v2},
url = {https://osf.io/preprints/psyarxiv/s57tz_v2/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Alzheimer's Disease Detection via Large Pre-Trained Speech and Language Models: Taxonomy, Progress, and Open Challenges
AU - Yuan-Chao Li
AU - Yilin Pan
AU - Yin-Long Liu
AU - Liu He
AU - Qi Sun
AU - Siyu Zhang
AU - Wei-Qiang Zhang
AU - Saturnino Luz
AU - Zhenhua Ling
AU - Jiahong Yuan
PY - 2026
DA - 2026/09/03
PB - PsyArXiv
DO - 10.31234/osf.io/s57tz_v2
UR - https://osf.io/preprints/psyarxiv/s57tz_v2/
AB - Alzheimer's Disease (AD) detection from spoken language has advanced considerably with the rise of large pre-trained speech and language models, including wav2vec 2.0, BERT, generative LLMs, and emerging audio-LLMs. These models often report higher performance on shared AD detection benchmarks than handcrafted-feature pipelines, while reducing the need for manual feature design. Yet their advantages remain difficult to interpret: pre-trained representations are largely a black box, benchmark gains are not always statistically robust, and performance can vary substantially across languages, elicitation protocols, and clinical task definitions. This survey proposes a taxonomy of how large pre-trained speech and language models have been used for AD detection and progression-related tasks. We define Large Pre-trained Models (LPMs) as Transformer-based models trained with self-supervised or weakly supervised objectives, distinguishing them from earlier non-Transformer pre-trained representations. We then organize the literature into six usage strategies: direct application, fine-tuning, transfer learning, multitask learning, multilingual modeling, and generative-LLM-based assessment. Across these categories, we synthesize evidence from shared benchmarks and recent studies to answer four questions: 1) what taxonomy captures how LPMs are used for AD detection; 2) what progress and advantages do LPMs offer for AD detection; 3) what challenges and limitations do LPMs exhibit; and 4) what future opportunities and open tasks they present.
ER -