The Perioperative Cognitive Anesthesia Network (PeCAN) group has developed a novel machine learning tool to enhance the scoring of the clocking drawing test, a method of screening patients for dementia.
The study is the first to use semi-supervised deep learning methods to analyze digital clock drawings. The findings, which were published in May in Scientific Reports, could have important implications for identifying patients at risk for dementia after surgical procedures.
Sabyasachi Bandyopadhyay, a Ph.D. candidate in the Department of Biomedical Engineering, supported by the group’s NIH grant, is the first author of the paper and the main developer of the program. He was advised by Patrick Tighe, M.D., M.S., associate professor of anesthesiology, the Donn M. Dennis, M.D., Professor of Anesthetic Innovation, associate dean for AI Application & Implementation in the UF College of Medicine, and codirector of PeCAN; Parisa Rashidi, Ph.D., associate professor in the J. Crayton Pruitt Family Department of Biomedical Engineering, UF term professor, Pruitt Family Endowed Fellow, and Bandyopadhyay’s Ph.D. adviser; and Catherine Price, Ph.D., ABPP-CN, associate professor, Paul Satz Term Professor, and codirector of PeCAN.
Other coauthors were Catherine Dion, M.S., a Ph.D. candidate in the Department of Clinical and Health Psychology, and David Libon, Ph.D., ABN, FACPN, a professor in the Departments of Geriatrics and Gerontology and Psychology at Rowan University who provided the well-characterized dementia cohort used in the study.
We caught up with Sabyasachi and Dr. Price to ask some questions about the new research.
Q: What does this research add to past work on the clock drawing test?
CP: Clinicians have historically used the clock drawing test to assess for dementia. But scoring can be challenging. This study used techniques from machine learning to help rapidly identify features seen in dementia patients with Alzheimer’s disease or due to vascular disease.
SB: This work represents differences in several key graphomotor anomalies in the clock drawing test (CDT) inside a single differentiable non-linear 2D manifold. This is important for two main reasons: a) It proves that construction-wise, several graphomotor anomalies are related (e.g., shape is related to size, size is related to placement of hands) and b) these relationships give researchers the opportunity to test changes in clock drawing over time and disparities (e.g., education). Previously it was difficult to associate graphomotor errors at different time points (for an individual) and across education (for different groups), but now we can simply project these CDTs to the 2D manifold and the manifold will depict evolution-trajectories of CDTs.
Q: How does this tool improve our understanding of patients at risk of dementia?
CP: It shows us unique features produced only by dementia. It improves our understanding of human behaviors that change with different dementia forms.
SB: In this study, we have operationalized the 2D latent manifold by dividing the space into a “dementia” region and a “control” region using clock drawings from dementia/non-dementia participants. Using this operationalized manifold, we can simply project clock drawings into the space and their distance from the “control” region can be used as a proxy to how poorly they performed on the CDT (and hence their risk of dementia).
Q: What are the future implications for patients?
CP: This machine learning approach may help us to rapidly identify changes in cognitive states. For example, we are looking at how the classification scheme can identify new cognitive impairment after orthopedic injuries, orthopedic surgeries, and delirium.
SB: We are currently using this manifold to track longitudinal CDT projections for patients who have undergone surgeries with a high anesthesia load (different orthopedic surgeries) to return trends that might help researchers in assessing their cognitive ability after surgery. We have seen several important trends from preliminary data that we aim to publish in the future.
Q: What are the next steps in this research?
CP: Validation with larger samples of data, application to clinical samples, and then ideally involvement in clinical assessments.
SB: Next steps for this research include augmenting this manifold with digit-based or time-based CDT features in larger samples to see whether it can be involved in clinically assessing dementia. We are also trying to use other neural network models to create a complete, orthogonal, and informative set of automatically extracted graphomotor features that can augment the classical CDT scoring techniques. Eventually, our objective is to improve understanding of the implications, and broaden the applications of the CDT as a primary screening test for at-risk presurgical populations. We eventually want to create an integrated AI score that will inform clinicians about cognitive status and predict long-term cognitive impairments of individuals coming in for surgery.