TransformerCARE: A novel speech analysis pipeline using transformer-based models and audio augmentation techniques...

Area of Science:

• Neurology
• Computational Linguistics
• Artificial Intelligence

Background:

• Early diagnosis of cognitive impairment, including Alzheimer's disease, is crucial but challenging, with over 50% of cases undiagnosed until advanced stages.
• Speech impairments are recognized as early indicators of cognitive decline, highlighting the need for novel diagnostic methods.

Purpose of the Study:

• To evaluate the utility of speech analysis using advanced speech transformer models for the early detection of cognitive impairment.
• To introduce and assess TransformerCARE, a novel speech processing pipeline for identifying early signs of cognitive decline.

Main Methods:

• The TransformerCARE pipeline involved preprocessing, speech segmentation, fine-tuning of four state-of-the-art speech transformer models (Wav2vec 2.0, HuBERT, WavLM, DistilHuBERT), segment aggregation, and performance evaluation.
• Data augmentation techniques, particularly frequency masking, were employed to enhance the model's ability to detect subtle acoustic cues.
• Performance was measured on the ADReSSo Challenge dataset, comprising 237 subjects.

Main Results:

• TransformerCARE achieved the highest performance with the HuBERT model, yielding an AUC of 81.80.
• Frequency masking data augmentation improved performance by 5%, reaching an AUC of 86.11.
• Incorporating clinicians' speech into the analysis further enhanced the pipeline's performance. Error analysis revealed distinct acoustic pattern differences in misclassified cases.

Conclusions:

• TransformerCARE demonstrates significant potential as a clinical screening tool for early detection of cognitive impairment.
• The pipeline can aid in timely diagnosis and appropriate patient care, facilitating integration into clinical workflows.