Biomarkers.

Area of Science:

• Artificial Intelligence
• Medical Diagnostics
• Neuroscience

Background:

• Early and accurate diagnosis of dementia is crucial for effective management.
• Traditional screening methods can be subjective and time-consuming.
• Integrating diverse data modalities offers potential for improved diagnostic accuracy.

Purpose of the Study:

• To develop a multimodal deep learning model for dementia diagnosis.
• To integrate voice and drawing data from screening tests.
• To evaluate the impact of different data modalities on classification performance.

Main Methods:

• A multimodal deep learning model was developed using voice and pentagon drawing data from 1,091 participants.
• Voice data was converted to Mel Frequency Cepstral Coefficient (MFCC) spectrograms; drawings were preprocessed into grayscale images.
• DenseNet and Multilayer Perceptron (MLP) were used for feature extraction, with weighted ensemble learning for the final model.

Main Results:

• The multimodal model achieved 66.3% accuracy and 0.73 AUC for three-group classification (normal, MCI, dementia).
• For binary classification (normal vs. dementia), accuracy reached 86.9% and AUC 0.86.
• Voice data alone demonstrated strong performance; drawing data improved multi-class classification.

Conclusions:

• The multimodal deep learning model integrating voice, drawing, and clinical data shows significant potential for diagnosing cognitive impairment and dementia.
• These findings highlight the value of combining diverse data modalities for enhanced diagnostic accuracy.
• The approach offers a scalable solution for early dementia detection in clinical and real-world settings.