Computer Science Colloquium


Utilizing deep learning to create tailored summary outcome metrics for clinical population

Mark Albert
University of North Texas

Date: Friday October 21, 2022
Time: 5:30pm MDT
Room: Zoom zoom.us, Meeting ID 926 9565 5625, passcode 488975
            The talk will be held in Speare Hall room 19 for the CSE 585 class

   Abstract:

In rehabilitation, there is much more research involved in incremental improvements and comparison across therapies than developing novel therapies. Such work is important as evaluations occur implicitly on a day-to-day basis when clinicians decide the appropriate treatment among a number of available options for impaired mobility, and also explicitly at the level of large, clinical research studies. One important aspect to help such decisions is to collect quality outcome metrics. Once a desired outcome metric is available, we have a broad array of tools to help optimize clinical care. Our lab has historically used machine learning to enable wearable sensor systems to provide convenient, continuous, and objective measurements of subject movements; this is critical for a thorough research-oriented approach to improving mobility. However, with more data and measures available, it can be difficult to give a summary answer to "does this approach improve quality of life?"; this is particularly problematic with small samples of highly variable individuals. In this work, we discuss a research approach which collects a number of relevant outcome measures and creates a single, summary metric to assess the effect of therapies. This has been successfully applied to prosthetic outcome measures for K2 amputees to demonstrate that a microprocessor knee significantly outperforms mechanically knees, despite limited support for such microprocessor knees from insurers. Additionally, we have worked with a system of motion analysis labs analyzing gait of children with Cerebral Palsy before and after surgical intervention. With over 100 metrics, we show how a summary metric approach can provide insights comparing surgical approach outcomes. In short, the overarching goal of much clinical work is to improve "quality of life" rather than disparate sets of narrow, interpretable metrics. This work helps to bridge the gap by creating summary metrics which better capture context-specific overall health.

Slides: https://docs.google.com/presentation/d/1ZbYbDm37nua6oWlJ4MYdUW8c0rcwjgIKNWd_PNS8ems/edit?usp=sharing

Bio:

Dr. Mark V. Albert is the director of the Biomedical AI Lab at the University of North Texas and holds a dual appointment in the Department of Computer Science and Engineering and the Department of Biomedical Engineering. He is also the Associate Chair for Graduate Studies in CSE. His lab leverages machine learning to automate the collection and inference of clinically useful health information from wearable sensors to improve clinical outcomes. Currently funded projects include inference of surgical outcomes from motional analysis data in children with Cerebral Palsy through the Shriners Hospital network, machine learning-based detection of a fall prior to impact for a commercially available wearable airbag belt with the Shirley Ryan AbilityLab, and an NSF funded REU program for hardware-accelerated deep learning. His projects in wearable sensor analytics have improved the measurement of health outcomes for individuals with Parkinson's disease, stroke, and transfemoral amputations, and cerebral palsy. Since his PhD in Computational Biology from Cornell University, he also maintains a series of projects related to efficient coding in sensory systems.