An explainable self-attention deep neural network for detecting mild cognitive impairment using multi-input digital...

Area of Science:

• Artificial Intelligence in Medicine
• Neuroscience
• Medical Imaging Analysis

Background:

• Mild cognitive impairment (MCI) is an early stage of cognitive decline, a precursor to dementia.
• Early detection of MCI is critical for timely prevention and intervention strategies.
• Current deep learning models for MCI detection using the clock drawing test (CDT) face challenges in early prediction and clinical interpretability.

Purpose of the Study:

• To develop a novel, interpretable deep learning framework for early detection of MCI.
• To improve the classification performance of MCI detection beyond single-task models.
• To provide visual explanations for model predictions to enhance clinical trust and adoption.

Main Methods:

• Recruited 918 subjects (651 healthy, 267 MCI patients).
• Developed a deep learning framework integrating data from Clock Drawing Test (CDT), cube-copying, and trail-making tests.
• Employed soft labels and self-attention mechanisms for improved performance and interpretability, validated with Grad-CAM and expert evaluation.

Main Results:

• The proposed multi-input model significantly improved MCI classification accuracy from 0.75 (baseline) to 0.81.
• Achieved higher F1-score (0.65 vs. 0.36) and AUC (0.84 vs. 0.74) compared to the baseline VGG16 model.
• The model's interpretability was rated higher by medical experts and showed better quantitative evaluation (IoU) than Grad-CAM.

Conclusions:

• The developed deep learning model demonstrates superior classification performance for MCI detection.
• The model provides enhanced visual explanations, addressing the 'black box' challenge in AI medical diagnosis.
• This interpretable AI offers a promising tool for clinical settings, crucial for accurate and trustworthy MCI diagnosis.