Application of Chinese Pre-trained Language Models in Early Detection ofCognitive Impairment: A Comparative Study Based on Spoken Text

Introduction: Degenerative cognitive disorders, such as Alzheimer’s Disease (AD), impose a substantial burden on societies and families worldwide. Currently, no definitive treatments or curative medications exist, and the academic consensus emphasizes the critical importance of early detection and intervention to mitigate disease progression. With advancements in artificial intelligence, particularly the rapid evolution of Natural Language Processing (NLP) technologies, novel approaches for the early identification of cognitive impairments have emerged. Text embeddings derived from Pre-Trained Language Models (PLMs) offer a promising means to classify spoken language samples, enabling objective assessment of cognitive status. However, research on the application of Chinese PLMs in this domain remains relatively scarce. Materials and Methods: Six representative Chinese Pre-Trained Language Models (PLMs) were used as feature extractors to generate text embeddings from transcribed spoken texts. The corpus included 45 healthy young adults, 46 elderly individuals with Mild Cognitive Impairment (MCI), and 48 patients diagnosed with Alzheimer's Disease (AD). These embeddings were combined with four classic machine learning algorithms, Support Vector Machines (SVM), Random Forests (RF), K-Nearest Neighbors (KNN), and Logistic Regression (LR), to conduct classification experiments. Results: Results showed RoBERTa performed best, achieving 95.71% accuracy with SVM, followed by BERT. MacBERT, SimCSE, ERNIE, and BGE had decreasing performance. Among classifiers, SVM and LR outperformed RF and KNN. Discussion: The results of this study not only verify the strong ability of Chinese pre-trained language models in mining semantic degradation features but also indicate that traditional machine learning algorithms still have competitiveness in scenarios with small samples and high-dimensional data. Compared with traditional methods that rely on manually designed language features, the text embedding-based classification strategy in this study undoubtedly shows higher performance Conclusion: These findings highlight the potential of Chinese PLMs in facilitating early detection of cognitive impairment, providing a technical foundation for developing accessible screening tools for Chinese-speaking populations.