ADCeleb: A Longitudinal Speech Dataset from Public Figures for Early Detection of Alzheimer's Disease

Kunxiao Gao, Anna Favaro, Najim Dehak, Laureano Moro Velazquez

TL;DR

This paper introduces ADCeleb, a novel longitudinal speech dataset of public figures diagnosed with Alzheimer's Disease (AD), collected up to 10 years before diagnosis. They demonstrate that multimodal speech and text analysis can accurately detect early signs of AD, with performance improving as the diagnosis approaches.

What Problem Does This Paper Address?

Early detection of Alzheimer's Disease (AD) is crucial for timely intervention and management, but it remains a significant challenge. Traditional diagnostic methods can be invasive or expensive, and current speech-based approaches often suffer from small, non-longitudinal datasets, limited recordings from the prodromal (pre-symptomatic) phase, and reliance on artificial speech tasks that don't reflect natural language use. This gap hinders the development of robust, non-invasive tools for early AD detection.

Methods

The researchers developed a multimodal approach combining acoustic and linguistic features extracted from speech. They utilized pre-trained, non-interpretable embedding models (NIFMs) from state-of-the-art neural networks (like Wav2Vec 2.0, HuBERT for acoustics, and XLM-ROBERTa, Multilingual-e5-large for linguistics) to capture nuanced speech patterns. These embeddings were then fed into classification models. A key aspect was the fusion of predictions from acoustic and linguistic models to leverage their complementary strengths. They employed a nested cross-validation scheme for rigorous evaluation, using Principal Component Analysis (PCA) for dimensionality reduction and Probabilistic Linear Discriminant Analysis (PLDA) for classification.

Datasets

The primary contribution is the introduction of ADCeleb, a new longitudinal speech dataset.

  • ADCeleb: Contains 5347 recordings (approx. 25 hours) from 80 public figures (40 diagnosed with AD, 40 cognitively normal controls - CNs).
  • Recordings span 1 to 10 years before the AD diagnosis (Year of Diagnosis - YoD).
  • Participants are matched by age, sex, and ethnicity.
  • Speech samples were extracted from publicly available YouTube videos, reflecting natural language use in various settings (interviews, conferences, etc.), though with inherent real-world noise.
  • The dataset is related to the ParkCeleb dataset, which was used as a pipeline inspiration.

Key Insights and Ideas

  • Public figures as a valuable data source: Leveraging public figures who later disclosed an AD diagnosis provides a unique opportunity to study longitudinal speech changes in a real-world, albeit specific, population.
  • Prodromal speech holds predictive power: The study strongly suggests that subtle changes in speech, detectable years before a formal diagnosis, can be indicative of AD.
  • Linguistic vs. Acoustic markers: Linguistic features appear more dominant in detecting AD in the earlier prodromal stages (6-10 years pre-diagnosis), while acoustic features become increasingly relevant closer to the diagnosis (1-5 years pre-diagnosis).
  • Multimodal fusion is key: Combining acoustic and linguistic information significantly boosts classification accuracy, especially in the later prodromal stages, highlighting the complementary nature of these modalities.
  • NIFMs are effective: Using pre-trained, non-interpretable embeddings from large neural models proves highly effective for capturing subtle speech alterations, outperforming traditional feature engineering for this task.

Key Results

  • Time Interval -2 (6-10 years before YoD):
    • Best acoustic model (Wav2Vec 2.0): Accuracy 0.67
    • Best linguistic model (Multilingual-e5-large): Accuracy 0.73
    • Fusion model: Accuracy 0.72 (slightly lower than the best unimodal linguistic model, suggesting linguistic markers are sufficient at this stage).
  • Time Interval -1 (1-5 years before YoD):
    • Best acoustic model (Wav2Vec 2.0): Accuracy 0.72
    • Best linguistic model (Cross-en-de-roberta): Accuracy 0.75
    • Fusion model: Accuracy 0.80 (outperforming both unimodal approaches, demonstrating the power of multimodal analysis as diagnosis nears).

Limitations and Future Work

  • Dataset limitations: While ADCeleb is valuable, it primarily features public figures, which may introduce biases (e.g., gender, socioeconomic status, access to media). The dataset also includes real-world noise and variability, which can be challenging.
  • Future directions:
    • Expand the dataset with more recordings to enable deeper longitudinal analysis.
  • Increase data diversity by including speakers from various linguistic, national, and ethnic backgrounds to improve model generalizability.
  • Investigate interpretable models to gain deeper insights into the specific speech changes associated with AD.
  • Explore the potential of these models for clinical trial triage and early screening in broader populations.

Summary updated

BibTeX
@inproceedings{gao2025adceleb,
  author = {Kunxiao Gao and Anna Favaro and Najim Dehak and Laureano Moro Velazquez},
  title = {{ADCeleb: A Longitudinal Speech Dataset from Public Figures for Early Detection of Alzheimer's Disease}},
  booktitle = {Interspeech},
  year = {2025},
}
Figure