Deep Neural Network Ensembles for the Detection of Alzheimer's Disease Using Imbalanced Clock Drawing Test Images

This paper explores innovative machine learning technologies for the detection of Alzheimer's Disease (AD) from large and imbalanced clock drawing test (CDT) images. We present a deep ensemble learning framework, 2Level MM-MD, which includes two-level ensemble of MultiModel deep neural networks (DNN) trained on MultiData blocks. The data blocks are generated by applying partition based and bootstrap statistical data resampling methods to the large but imbalanced CDT image data. To produce a more robust system that is better adapted to imbalanced CDT data, the 2Level MM-MD ensemble framework attempts to leverages both 1) the richness of large data blocks generated using effective data re-sampling schemes, and 2) multiple deep neural network systems that are trained on the resembled data blocks through transfer learning to produce a more robust system that is better adapted to more generalized CDT image domains. Three deep neural network models (EfficientNet, ResNet101, and ViT), along with six different sampling methods are used to generate re-balanced data blocks, and five different decision schemes are implemented and evaluated. The ensemble systems are trained and evaluated using a large CDT image dataset from Rounds 1 to 9 of the National Health and Aging Trends Study (NHATS), which contains over 47,723 CDT images, where 92% are class 0 (no dementia) and 8% are class 1 (probable dementia). The top classification system is a 2Level MM-MD system E5 with a two-level decision rule SoM. The system generated 0.67 Recall and 0.35 F1-Score on the test data. This represents a 123% improvement in Recall and a 25% improvement in F1-Score compared to the best baseline system, which is a classifier trained using EfficientNet with the training data balanced by oversampling the minority class.