Variational autoencoder provides proof of concept that compressing CDT to extremely low-dimensional space retains its ability of distinguishing dementia

Abstract

The clock drawing test (CDT) is an inexpensive tool to screen for dementia. In this study, we examined if a variational autoencoder (VAE) with only two latent variables can capture and encode clock drawing anomalies from a large dataset of unannotated CDTs (n = 13,580) using self-supervised pre-training and use them to classify dementia CDTs (n = 18) from non-dementia CDTs (n = 20). The model was independently validated using a larger cohort consisting of 41 dementia and 50 non-dementia clocks. The classification model built with the parsimonious VAE latent space adequately classified dementia from non-dementia (0.78 area under receiver operating characteristics (AUROC) in the original test dataset and 0.77 AUROC in the secondary validation dataset). The VAE-identified atypical clock features were then reviewed by domain experts and compared with existing literature on clock drawing errors. This study shows that a very small number of latent variables are sufficient to encode important clock drawing anomalies that are predictive of dementia.

Explainable semi-supervised deep learning shows that dementia is associated with small, avocado-shaped clocks with irregularly placed hands

“…For instance, a smaller clockface area is associated with subcortical disease profiles with primary executive dysfunction (e.g., micrographia in Parkinson's disease) 9 , and misplacement of clock hands is associated with visual attention deficits and disinhibition 2 . In comparison to a previously published VAE encoding 27 , the RF-VAE encoding reported in this study achieved significantly better results on the same classification dataset using identical training methods. This improvement is due to diversification of the latent space, and disentangling the latent dimensions.…”

“…The feed-forward neural network classifier combines these features in a non-linear way to discriminate dementia from controls. A previous study attempted to classify dementia from non-dementia using a two-dimensional latent space VAE network 27 . This work provided proof of concept that compressed CDT representations retain their ability to distinguish dementia.…”

Using digital assessment technology to detect neuropsychological problems in primary care settings

“…The horizontal and vertical placement of the drawing in the space provided loaded on other factors. Using a relevance factor variational autoencoder (RF-VAE), a deep neural network, Bandyopadhyay et al (2022) found that an irregularly drawn clock face was able to distinguish between dementia versus non- dementia patients. As described above, Dion et al (2022) found that errors in number placement was associated with poorer performance on selected neuropsychological tests and negatively associated with connectivity from the basal nucleus of Meynert (BNM) to the anterior cingulate cortex (ACC).…”

A deep learning approach for morphological feature extraction based on variational auto-encoder: an application to mandible shape

“…While hybrid architectures of Variational Autoencoder (VAE)-based unsupervised learning and classifier module have been investigated for solving classification tasks with limited labeled data and a large number of non-labeled data 78 , 79 , their application to dimensionality reduction and feature extraction has been studied more recently. For example, Bandyopadhyay et al 56 utilized this architecture to extract features from drawings by dementia patients to distinguish between dementia and non-dementia cases. Similar architectures have been extended to handle multimodal inputs for anomaly detection in robotic vehicles under uncertain environments 55 , or for classification of diverse cancer types using omics data 57 .…”