Voice Analysis-Based Early Screening System for Alzheimer's Disease: Innovative Application of Deep Learning in Cognitive Health

This article presents an open-source multimodal early screening system for Alzheimer's disease, which uses voice analysis and deep learning technologies (supporting architectures like LSTM and Transformer) to achieve early detection of cognitive decline. The system adopts a modular design, supporting two modes: voice audio screening and Kaggle tabular data prediction. It has a complete training and evaluation system and an interactive demonstration interface, making it of great application value in the field of cognitive health monitoring.

Alzheimer's disease is a progressive neurodegenerative disease, and early diagnosis is crucial for delaying its progression. Traditional diagnosis relies on expensive brain scans and complex tests, while voice analysis technology provides a non-invasive, low-cost alternative. Studies have shown that cognitive decline leaves recognizable traces in language expression, speech rate, vocabulary choice, etc., creating possibilities for automatic screening using machine learning.

This project builds a complete multimodal Alzheimer's disease detection system, whose core goal is to identify early signals of cognitive decline using voice analysis and deep learning. The system adopts a modular architecture and supports two data modes: voice audio screening mode and Kaggle tabular data prediction mode. This allows researchers to flexibly choose analysis paths based on data, suitable for clinical research (voice samples) and retrospective analysis (structured data).

The project's tech stack covers multiple deep learning architectures: for voice processing, it supports Long Short-Term Memory (LSTM) networks, Transformer architectures, and dense neural networks, which can extract temporal features and semantic patterns from audio to capture cognitive indicators; for tabular data, it provides a dedicated dense neural network baseline model to process structured features such as demographic data and cognitive test scores.

The system's data processing workflow is rigorous: the voice mode requires audio files to be paired with CSV metadata (including core fields such as recording ID, audio path, diagnostic label, and supports covariates like age and gender); for controlled-access voice datasets such as ADReSS, ADReSSo, and DementiaBank, a dedicated import tool is provided, which can automatically infer label categories based on folder names to simplify data preparation.

The project is equipped with a complete training and evaluation script system: the training process is driven by YAML configuration files (supporting presets such as default, Kaggle tabular, and ADReSS voice); the evaluation module provides comprehensive performance analysis, and the prediction module supports real-time inference for single audio files; it emphasizes the use of subject-level data partitioning to avoid data leakage and ensure the generalization ability of medical AI models.

To lower the barrier to use, the project integrates a Streamlit interactive demonstration interface. Users can upload voice samples via the web page to get real-time prediction results, experience the functions without code, which is convenient for clinical researchers to quickly verify ideas and also provides a reference architecture for product deployment.

This project has significant value in the field of cognitive health monitoring: it provides researchers with a reproducible and scalable benchmark framework, promoting standardization in the field of voice biomarkers; the modular design supports the integration of new models/features, facilitating technological iteration; the open-source nature ensures algorithm transparency and auditability, enhancing the credibility of medical AI. In the future, it is expected to integrate voice, text, and vision to build a more comprehensive cognitive assessment system.