An Explainable Self-Attention Deep Neural Network for Detecting Mild Cognitive Impairment Using Multi-input Digital Drawing Tasks

Abstract

Background: Mild cognitive impairment (MCI) is an early stage of cognitive decline beyond the normal aging process, which could develop into dementia. An early detection of MCI is a crucial step for timely prevention and intervention. Recent studies have developed deep learning models to detect MCI and dementia using a bedside task like the classic clock-drawing test (CDT). While these models succeed at distinguishing severe forms of dementia, it remains a challenge to predict the early stage of the disease using the CDT data alone. Moreover, the state-of-the-art deep learning techniques still face the black-box challenges, making it questionable to implement them in the clinical setting.Methods: We consecutively recruited 981 subjects from Geriatric Clinic at King Chulalongkorn Memorial Hospital (651 healthy subjects and 267 MCI patients according to Petersen’s criteria). We propose a novel deep learning modeling framework that incorporates data from multiple drawing tasks including the CDT, cube-copying, and trail-making tasks obtained from a digital platform. Soft-label method and self-attention were applied to improve the model performance and provide visual explanation that aids the interpretation of the deep learning model. The performance of the model was measured on an independent dataset. The visualization of our model and the traditional grad-CAM approach were scored by experienced medical personnels. Results: Using multiple drawings as inputs improves classification between healthy versus MCI in terms of accuracy (0.75 ± 0.01 vs. 0.80 ± 0.01), F1 score (0.36 ± 0.04 vs. 0.59 ± 0.02) and AUC (0.74 ± 0.01 vs. 0.81 ± 0.02). By adding soft-label and self-attention methods improve the accuracy (0.81 ± 0.01) F1 score (0.65 ± 0.01) and AUC (0.84 ± 0.01). The experienced medical personnels preferred the visualization of our model over the grad-CAM model (2.82 ± 0.49 vs. 1.52 ± 0.54). Conclusions: Our model achieves better classification performance at detecting MCI compared to a well-established convolutional neural network model. Moreover, our model can highlight features of the MCI data that considerably deviate from those of the healthy aging population, offering accurate predictions for detecting MCI along with visual explanation that aids the interpretation of the deep learning model.