Mild cognitive impairment prediction based on multi-stream convolutional neural networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Diagnostics

Background:

• Mild cognitive impairment (MCI) represents a transitional stage between normal aging and dementia.
• Early MCI diagnosis is crucial for effective healthcare interventions.
• Current diagnostic methods, including cognitive and neuroimaging tests, are often costly and time-consuming.

Purpose of the Study:

• To develop and evaluate a novel multi-stream convolutional neural network (MCNN) model for predicting MCI using facial videos.
• To explore the potential of facial data as a non-invasive biomarker for MCI detection.
• To establish an automated, cost-effective, and objective method for MCI screening.

Main Methods:

• Utilized a dataset of 48 facial videos from 45 participants (35 normal, 13 MCI).
• Employed a multi-stream convolutional neural network (MCNN) to extract spatial and dynamic facial features from video segments.
• Evaluated 27 MCNN model combinations, varying ResNet architectures, optimizers, and activation functions.

Main Results:

• The MCNN model achieved an F1-score of 89% at the segment level using ResNet-50, Swish activation, and Ranger optimizer.
• A ResNet-18 backbone with Swish and Ranger achieved a perfect 100% F1-score at the participant level.
• ResNet-50 demonstrated stability across different optimizers, indicating robustness for hyperparameter tuning.

Conclusions:

• Facial videos can be effectively utilized for predicting MCI, serving as a valuable biomarker.
• The developed MCNN approach offers a promising, automated, non-invasive, and inexpensive alternative to traditional MCI screening methods.
• The study provides insights into hyperparameter optimization for enhancing MCI prediction accuracy using facial data.