A multimodal screening system for elderly neurological diseases based on deep learning.

Area of Science:

• Artificial Intelligence
• Neurology
• Medical Diagnostics

Background:

• Neurological diseases require early screening for effective management.
• Current screening methods can be invasive or require specialized equipment.
• There is a need for accessible and accurate screening tools for neurological disorders.

Purpose of the Study:

• To develop and validate a deep learning algorithm for screening neurological diseases.
• To integrate multi-modal data, including human landmarks and voice, for improved diagnostic accuracy.
• To assess the effectiveness of various examination protocols and data modalities.

Main Methods:

• A deep learning algorithm was designed using subnetworks for specific protocols and a feature aggregator.
• Video data was converted into human landmarks, and voice data was collected.
• Pre-trained models, including graph neural networks and time-delay neural networks, were used for feature extraction.
• Multitask learning and data augmentation techniques were employed to enhance model performance.
• A frame-length staticizer was utilized to capture subtle movements like tremors.

Main Results:

• The algorithm achieved an Area Under the Curve (AUC) of 0.802 for stroke screening.
• The system demonstrated an AUC of 0.780 for Parkinson's disease screening.
• Experiments confirmed the effectiveness of different protocols and data modalities (body parts, voice).

Conclusions:

• The proposed deep learning approach offers an effective method for screening neurological diseases.
• Integrating human landmark and voice data enhances screening capabilities.
• The system shows promise as an accessible and accurate tool for early detection of neurological conditions.