[Paper Review] Explainable CNN-attention Networks (C-Attention Network) for Automated Detection of Alzheimer's Disease

Ning Wang, Mingxuan Chen+1 more|arXiv (Cornell University)|Jun 25, 2020

Machine Learning in Healthcare4 citations

TL;DR

This paper proposes C-Attention Networks, a novel explainable deep learning framework that combines 1D CNNs and self-attention mechanisms to detect Alzheimer’s disease from speech transcripts using part-of-speech (PoS) tags and universal sentence embeddings. The model achieves 92.2% accuracy and 0.952 F1-score on the DementiaBank dataset while providing both intra-feature and inter-feature class explanations through attention weights.

ABSTRACT

In this work, we propose three explainable deep learning architectures to automatically detect patients with Alzheimer`s disease based on their language abilities. The architectures use: (1) only the part-of-speech features; (2) only language embedding features and (3) both of these feature classes via a unified architecture. We use self-attention mechanisms and interpretable 1-dimensional ConvolutionalNeural Network (CNN) to generate two types of explanations of the model`s action: intra-class explanation and inter-class explanation. The inter-class explanation captures the relative importance of each of the different features in that class, while the inter-class explanation captures the relative importance between the classes. Note that although we have considered two classes of features in this paper, the architecture is easily expandable to more classes because of its modularity. Extensive experimentation and comparison with several recent models show that our method outperforms these methods with an accuracy of 92.2% and F1 score of 0.952on the DementiaBank dataset while being able to generate explanations. We show by examples, how to generate these explanations using attention values.

Motivation & Objective

• To develop an explainable deep learning framework for early detection of Alzheimer’s disease using non-invasive speech transcripts.
• To address the lack of interpretability in existing deep learning models for AD detection by introducing attention-based explanations.
• To enable modular integration of multiple linguistic feature types (e.g., PoS and embeddings) in a unified architecture.
• To improve model performance and clinical interpretability over prior methods using only linguistic features or non-explainable models.
• To demonstrate the feasibility of deploying such models in real-world settings via mobile applications for home-based screening.

Proposed method

• The model uses a 1D Convolutional Neural Network (1D-CNN) to extract hierarchical patterns from sequential linguistic features.
• Self-attention mechanisms are applied at two levels: intra-feature (within PoS or embedding features) and inter-feature-class (between PoS and embeddings).
• The architecture is modular, allowing easy extension to additional feature classes beyond PoS and embeddings.
• Intra-class attention computes relative importance of individual features within a feature class (e.g., specific PoS tags).
• Inter-class attention computes the relative contribution of different feature classes (e.g., PoS vs. embeddings) to the final prediction.
• The final prediction is generated by a fusion layer that combines attended representations from all feature streams.

Experimental results

Research questions

• RQ1 Can a unified deep learning architecture effectively combine part-of-speech tags and pre-trained sentence embeddings for Alzheimer’s disease detection?
• RQ2 How can attention mechanisms be used to generate both intra-feature and inter-feature class explanations in a clinically interpretable way?
• RQ3 To what extent do PoS features and sentence embeddings contribute differently to model decisions across patient subgroups?
• RQ4 Can the model achieve high performance while maintaining interpretability comparable to clinical insights?
• RQ5 How does the model’s attention distribution compare with known linguistic biomarkers of Alzheimer’s disease, such as pronoun usage?

Key findings

• The C-Attention-FT+Embedding model achieved 92.2% accuracy and 0.952 F1-score on the DementiaBank dataset, outperforming several recent models.
• The PoS features NNPS (proper noun, plural), MD (modal), EX (existential), and PRP (personal pronoun) were identified as the most influential, collectively accounting for 80.9% of attention in PoS features.
• For 65.1% of test cases, the PoS feature stream received higher attention weights than the sentence embedding stream, indicating its dominant role in decision-making.
• The model correctly classified all three tested patients (including two AD patients and one healthy control), with attention weights reflecting clinical plausibility—e.g., higher embedding attention in AD cases with complex linguistic deficits.
• The model’s attention patterns align with prior clinical findings: AD patients tend to use more pronouns (PRP) and fewer content nouns, which is reflected in high attention to PRP and NNPS.
• The modular design enables extension to additional feature types, and the attention mechanism provides actionable, human-readable explanations for model predictions.