Multi-dimensional Speech Feature Fusion: A New Machine Learning Approach for Early Screening of Alzheimer's Disease

This article focuses on integrating three complementary features—acoustic, prosodic, and phonetic—with machine learning technology to achieve automatic detection of Alzheimer's Disease (AD), providing a non-invasive solution for early diagnosis. This method addresses the limitations of traditional diagnostic methods, such as high invasiveness and cost, and the integrated model achieves an F1-score of 0.89, providing technical support for large-scale population screening and clinical applications.

Current Status and Challenges of AD

Alzheimer's disease is the most common neurodegenerative disease globally, accounting for 60%-70% of dementia cases, with approximately 55 million patients worldwide. Early symptoms are difficult to detect, and irreversible brain damage has already occurred when obvious cognitive impairment appears.

Limitations of Traditional Diagnosis

Relying on neuropsychological assessments, cerebrospinal fluid testing, etc., it has problems such as high invasiveness, high cost, and need for professional equipment, making large-scale screening difficult.

Theoretical Basis of Speech Analysis

Subtle changes in language ability are early manifestations of AD, such as degradation in lexical retrieval and semantic comprehension, which provide a basis for automatic detection via speech analysis.

Acoustic Features (Physical Properties)

Including fundamental frequency, formants, energy envelope, speech rate and pauses, etc. AD patients show slowed speech rate, prolonged and irregular pauses, and reduced fundamental frequency variability, reflecting the decline of the nervous system's control over vocal organs.

Prosodic Features (Rhythm and Melody)

Covering intonation, stress, rhythm, etc. AD patients exhibit monotonous and flat prosody ("prosodic flattening"), which is related to the degeneration of the right hemisphere of the brain and the limbic system.

Phonetic Features (Structural Units)

Focusing on the accuracy of phoneme pronunciation and error patterns (substitution/omission/repetition), the type and frequency of errors are related to the severity of the disease, which can distinguish between normal aging and pathological decline.

Feature Engineering

After preprocessing, more than 200 low-level acoustic features are extracted. A discriminative subset is selected through recursive feature elimination and tree model feature importance, improving efficiency and interpretability.

Ensemble Learning Strategy

The integrated model that fuses the three types of features outperforms single-feature models, achieving an F1-score of 0.89.

Interpretability Analysis

Key features are identified through SHAP values: number of pauses, coefficient of variation of fundamental frequency, error rate of specific phoneme pronunciation, etc., providing clues for the mechanism of language pathology.

Dataset Selection

The public dataset from the ADReSS challenge is used, which contains spontaneous speech samples from cognitively normal elderly, patients with mild cognitive impairment, and AD patients, with strong sample representativeness and high annotation quality.

Significance of Longitudinal Tracking

Some participants have undergone longitudinal evaluation for several years, allowing observation of the evolution trajectory of speech features from normal aging to AD, which helps in the establishment of early warning models.

Prospect of Cross-language Validation

The method has the potential of language independence; in the future, it can be validated in populations using Chinese, Spanish, etc., to promote global application.

Application Scenarios

• Community/family self-screening tool: Smartphones record speech to generate risk reports, lowering the threshold for screening;
• Clinical auxiliary tool: Provides objective and quantitative references for doctors.

Challenges

• Influencing factors: Interferences such as age, education, dialect, and emotion require personalized benchmark models;
• Privacy and ethics: Data collection and storage need security protocols and ethical reviews;
• Result positioning: Only a risk prompt, not a substitute for professional diagnosis.

Limitations

Limited sample size, short span of longitudinal data, unvalidated generalization ability across datasets, and need to improve the sensitivity of early mild cognitive impairment recognition.

Future Directions

• Integrate more language features such as lexical semantics and syntactic complexity;
• Explore deep learning modeling;
• Conduct large-scale prospective cohort studies to verify clinical utility;
• Develop user-friendly applications to promote the translation of research findings into practical applications.

Early screening of AD is a key link in healthy aging. The multi-dimensional speech feature fusion method provides technical support for low-cost, non-invasive large-scale screening tools. With the progress of AI and data accumulation, speech analysis is expected to move from the laboratory to clinical practice, benefiting millions of families and deepening the understanding of the relationship between human language and the brain.