Deep learning inspired game-based cognitive assessment for early dementia detection

This paper introduces a gaming approach inspired by deep learning for the early detection of dementia. This research employs a convolutional neural network (CNN) model to analyze health metrics and facial images via a cognitive assessment gaming application. We have collected 1000 samples of health metric data from Apollo Diagnostic Center and hospitals, labeled "demented" or "nondemented," to train a modified 1-dimensional convolutional neural network (MOD-1D-CNN) for game level 1. Additionally, a dataset of 1800 facial images, also labeled "demented" or "non-demented," is collected in our work to train a modified 2-dimensional convolutional neural network (MOD-2D-CNN) for game level 2. The MOD-1D-CNN has achieved a loss of 0.2692 and an accuracy of 70.50% in identifying dementia traits via health metric data; in comparison, the MOD-2D-CNN has achieved a loss of 0.1755 and an accuracy of 95.72% in distinguishing dementia from facial images. A rule-based linear weightage method combines these models and provides a final decision. In addition, a better fusion neural network strategy is also explored in the results analysis with an ablation study. The proposed models are computationally efficient alternatives with significantly fewer parameters than other state-of-the-art models. The performance and parameter counts of these models are compared with those of existing deep learning models, emphasizing the role of AI in enhancing early dementia.