Applying Variational Autoencoder to Explore Changes in Clock Drawing With Hip Fracture Surgery. (20th December 2022)

• Record Type: Journal Article
• Authors: Chapin, Benjamin Albert; Bandyopadhyay, Sabyasachi; Rodriguez, Katie; Amini, Shawna; Radhakrishnan, Nila; Tighe, Patrick; Rashidi, Parisa; Price, Catherine C.
• Is Part Of: Alzheimer's & dementia. Volume 18(2022)Supplement 7
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-12-20
• Subjects: Alzheimer's disease -- Periodicals; Alzheimer Disease -- Periodicals; Dementia -- Periodicals; Démence; Maladie d'Alzheimer; Périodique électronique (Descripteur de forme); Ressource Internet (Descripteur de forme); 616.83
• DOI: 10.1002/alz.067136
• Languages: English
• ISSNs: 1552-5260

Abstract

Abstract: Background: Older adults with acute hip fractures are particularly vulnerable to delirium and postoperative cognitive decline. Hospital staff, however, do not screen preoperative cognition or systematically monitor for delirium in these at‐risk individuals. Using the clock drawing test (CDT) our team demonstrated the potential for semi‐supervised deep learning using Variational Autoencoder (VAE) to extract clock drawing anomalies predictive of dementia (Bandyopadhyay, in review). We applied the VAE technique to assess preoperative to postoperative changes in CDTs in patients with acute hip fractures. Methods: Post‐injury, pre‐operative hip fracture patients completed screening including the CDT to command and copy, and a baseline delirium assessment with the Confusion Assessment Method for the Intensive Care Unit (CAM‐ICU). Postoperatively, delirium was assessed twice daily. Post‐surgery CDTs were collected after two consecutive negative CAM‐ICUs. A VAE with two latent dimensions was trained using 13, 580 unlabeled clocks in an unsupervised manner to create a parsimonious 2D latent space to encode the constructional aspects of the CDT. This latent space was operationalized using a k‐Nearest Neighbor Classifier on a classification dataset consisting of 71 Dementia and 80 Control samples using 3‐fold cross validation. This separated the latent space into the "Red" and "Blue" areas corresponding to "Dementia" and "Control" groups, respectively. We projected the pre and postoperative hip fracture dataset CDTs onto this latent space to investigate changes in CDT relative constructional features post‐surgery for those with and without postoperative delirium. Result: 27 participants completed our study (age 82.96 ± 9.28; education 13.19 ± 2.90; 85% female). 2 were delirious preoperatively and excluded from analysis, and 4 developed delirium postoperatively (16%). All preoperative command and copy clocks were in the "Dementia" region of the VAE latent space. Copy, but not command clocks moved further into the "Dementia" region. Error type changed in the post‐operative command condition to resemble the copy condition more closely. Trends suggest preoperative copy clock performance may predict delirium risk after surgery. Conclusion: Semi‐supervised deep learning with VAE applied to the CDT, particularly the copy condition, shows promise for automated preoperative cognitive screening and postoperative monitoring in at‐risk hip fracture patients. Future studies are needed with larger participant samples.