Research team develops ML using video and wearable tech that detects abnormal movements to help fight against conditions such as Parkinson's Disease.
“Training a neural network model on-premises with our former in-lab motion capture dataset used 13GB of GPU memory and took four days for a single experiment. With Oracle Cloud, we’re running between 4 and 8 GPUs in parallel to vastly accelerate our research progress, meaning we can complete an experiment in just a few hours.”
According to a Parkinson's Foundation-supported investigation conducted in 2022, nearly 90,000 Americans are diagnosed with Parkinson's disease (PD) each year, with the total number of individuals with PD predicted to rise to 1.2 million by 2030. With the world's population aging at a rapid rate, it’s becoming increasingly urgent to treat neurodegenerative conditions that affect movement.
The way scientists measure PD symptoms is essentially unchanged since the late 1980s. Neurologists ask patients with PD to walk, tap fingers, move hands, speak, and balance. Then the neurologist rates how well the patient does each task from 0 to 4. This approach is subjective and so involved that many patients receive the test only once a year, making accurate observation and treatment difficult.
Researchers at Emory University set out to design a machine learning system using videos and wearable devices that would precisely measure PD symptoms every day. The project involved designing deep learning models for 3D motion capture data or video datasets, which demand significant GPU computing and memory usage. However, the researchers’ on-premises computer lacked the computation time needed for the project, and the computer was continuously in use by other research teams.
I was relieved to know our OCI Compute instance was HIPAA-compliant from the beginning, and our ability to spin up bare metal GPUs was critical. This allowed the team to achieve faster compute speeds without a Virtual Machine (VM) layer.
Why Emory University chose Oracle
Because the university’s project involved designing deep learning models for 3D motion capture data or video datasets, it demanded significant GPU computing and memory usage. Getting access to Oracle Cloud Infrastructure Compute and GPU instances would significantly accelerate the team’s research by testing variations of deep learning models in a parallel manner.
For example, training a neural network model on-premises with an in-lab motion capture dataset used 13GB of GPU memory and took four days for a single experiment. Using Oracle Cloud Infrastructure (OCI), the team realized it could run 4 to 8 GPUs in parallel and vastly accelerate research progress—taking only a few hours to complete an experiment.
Assistant Professors J Lucas McKay and Hyeokhyen Kwon were relieved to discover that Oracle’s cloud compute instance was HIPAA-compliant from the beginning. Also the team’s ability to spin up bare metal GPUs was critical, enabling the team to achieve faster compute speeds without a Virtual Machine (VM) layer. Oracle for Research’s cloud solutions architects were also very responsive, making machine learning and cloud technologies more understandable and helping the team with training.
The neural network’s results matched with expert ratings for freezing of gait (FOG) in Parkinson’s disease (PD) at a high precision rate of 96.2%. The team designed a ML model that is also interpretable automatically demonstrating, which exact patterns of movement are most important when determining PD and FOG severity scores. This feature will be particularly useful for physicians, who will have a direct insight into how the ML reaches its conclusions.
Lucas, Hyeokhyen, and the team expect that this project will benefit the Human Activity Recognition research community in at least two ways. First, the researchers plan to release the motion capture and video dataset with dense PD behavior statuses labeled by motion disorder specialists. Because the scale of this dataset is expected to be larger than any currently available public dataset detailing PD and “freezing of gait” or FOG, it should encourage more ML/AI experts to advance PD analysis with their own novel deep learning models.
Second, the team plans to release the ML models with the large-scale dataset used in this project. A model pretrained with large-scale data for a domain typically helps any other downstream tasks in advanced analysis.
About the customer
Machine Learning for BehaVior and HealTh AnaLytics (ViTAL) Lab at Emory Biomedical Informatics strives to develop artificial intelligence systems that are inclusive, accessible, fair, and reliable that will effectively improve the healthcare system. Its mission is to develop efficient computing and machine learning systems on the network edge using distributed ambient, mobile, and wearable devices to monitor patients’ conditions in everyday life.
The Emory Brain Health Center Motion Analysis Laboratory uses comprehensive measurements of body motion to characterize disease phenotypes and predict responses to treatments like deep brain stimulation.
- Oracle Research
- How Human Activity Recognition (HAR), wearables, and AI are helping in the fight against Parkinson’s Disease, opens in new tab
- Research in Action: How Human Activity Recognition (HAR), wearables, and AI are helping in the fight against Parkinson’s Disease, opens in new tab
- ViTAL Lab, opens in new tab