CDA: A Contrastive Data Augmentation Method for Alzheimer's Disease Detection

Junwen Duan, Fangyuan Wei, Jin Liu, Hongdong Li, Tianming Liu, Jianxin Wang

Introduction

Early detection of Alzheimer’s Disease (AD) is a critical clinical goal, yet traditional diagnostic methods remain resource-intensive and expensive. While deep learning models utilizing spontaneous speech transcripts offer a non-invasive alternative, they are severely hampered by the scarcity of medical data. Standard datasets often contain fewer than 200 samples, leading to overfitting in complex architectures like BERT. To address this, researchers have proposed Contrastive Data Augmentation (CDA), a novel framework that simulates cognitive decline to improve model generalization.

Methodology

The CDA framework enhances the training process by teaching the model to recognize the specific linguistic markers of cognitive degradation. The architecture consists of three primary components: text encoding, sample generation, and a specialized loss function.

The CDA framework architecture

1. Negative Sample Generation (Random Deletion)

To mimic the word-finding difficulties and loss of semantic coherence typical in AD patients, the method employs Random Deletion (RD). By removing a proportion of words (e.g., 30%) from original transcripts, the system creates "corrupted" negative samples. These are logically expected to appear "more like AD" than the original text.

2. Positive Sample Generation (R-Drop)

To ensure feature stability, the model uses R-Drop. The same input is passed through the network twice with different dropout masks. A Kullback-Leibler (KL) divergence loss penalizes inconsistencies between these two outputs, forcing the model to learn more robust representations.

3. Contrastive Margin Loss

The core innovation is the Contrastive Margin Loss. The model is trained to ensure that the prediction for an original transcript is "healthier" than its corrupted counterpart by at least a predefined margin ($m$). This directional constraint forces the model to learn the process of decline rather than just static patterns.

Dataset

The study was evaluated using the ADReSS Challenge dataset, derived from the DementiaBank’s English Pitt Corpus. It consists of transcripts from 156 participants (78 AD, 78 non-AD) describing the "Cookie Theft" picture. The dataset is balanced for age and gender to minimize demographic bias.

Results and Discussion

The CDA method achieved state-of-the-art performance on the ADReSS benchmark. A single CDA-enhanced model reached 87.5% accuracy, while an ensemble of five models achieved 91.7% accuracy with 100% precision for the AD class. This outperforms larger models like ERNIE 3.0 and multimodal approaches that utilize raw audio features.

An ablation study highlights the critical role of the Random Deletion module. As shown in the table below, removing RD leads to a significant drop in accuracy from 77.5% to 72.3%, confirming that simulating cognitive impairment is the primary driver of the model's success.

Ablation study of CDA components

Conclusion

CDA demonstrates that domain-specific data augmentation—specifically simulating the symptoms of the target disease—is more effective than generic NLP techniques for small-scale medical datasets. While currently focused on text transcripts, future work could integrate acoustic features to further capture the nuances of AD speech patterns. Although not yet a clinical diagnostic tool, CDA provides a powerful foundation for non-invasive AD screening research.

Summary updated

Abstract

Alzheimer’s Disease (AD) is a neurodegenerative disorder that significantly impacts a patient’s ability to communicate and organize language. Traditional methods for detecting AD, such as physical screening or neurological testing, can be challenging and time-consuming. Recent research has explored the use of deep learning techniques to distinguish AD patients from non-AD patients by analysing the spontaneous speech. These models, however, are limited by the availability of data. To address this, we propose a novel contrastive data augmentation method, which simulates the cognitive impairment of a patient by randomly deleting a proportion of text from the transcript to create negative samples. The corrupted samples are expected to be in worse conditions than the original by a margin. Experimental results on the benchmark ADReSS Challenge dataset demonstrate that our model achieves the best performance among language-based models.
BibTeX
@inproceedings{duan2023cda,
  author = {Junwen Duan and Fangyuan Wei and Jin Liu and Hongdong Li and Tianming Liu and Jianxin Wang},
  title = {{CDA: A Contrastive Data Augmentation Method for Alzheimer's Disease Detection}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
  doi = {10.18653/v1/2023.findings-acl.114},
}
Figure