AI for Human Thriving (AI4HT)

thrive verb

to grow or develop well or vigoursly; to flourish; to prosper.

People thrive when they are safe, secure and healthy. Problems of human thriving span a broad range from national and international security to environmental safety to individual health and wellness. We want to help solve problems in these areas and advance the state of the art in AI and machine learning in support of that goal.

Collaborators

AI for human thriving is necessarily interdisciplinary. Please contact Professor Simpkins at christopher.simpkins@kennesaw.edu if you are interested in discussing opportunities.

Application Areas

Health and Wellness

The health challenges facing our country are great. In 2019 life expectancy started decreasing for the first time in 100 years. Americans are less physically fit than at any time in our history, which not only decreases our quality of life but threatens our security as our military faces a historic recruiting crisis. Nearly half of all Americans suffer from at least one chronic illness. And our population is aging rapidly. Yet against all these challenges there is great opportunity. We have more data and better data collection and distribution systems than ever before, and over a third of Americans wear fitness trackers and health monitors. We have barely begun to tap the enormous potential of AI to improve the health of our population, and the potential impact can hardly be overstated.

Some particular goals include:

• improving clinical practice via AI-powered decision support,
• improving the efficiency and effectiveness of health practitioners and researchers by building tools that automate parts of their work,
• improving health outcomes for patients and baseline health for everyone, and
• reducing workload, stress and burnout in our indispensable and overworked health care workers by making progress on the goals above.

Current Projects in Health and Wellness

AI for Fall Risk Assessment and Reduction

The CDC estimates the total annual cost of falls to be more than $50 billion annually. In older adults these falls often lead to life-threatening complications. As our population ages, this problem will only increase. We are currently working with biomechanics researchers who specialize in fall risk assessment and prevention in older adults and in clinical populations, e.g., people suffering from neuro-muscular disorders, on the following early-stage projects:

• Automated step labeling for Vicon motion capture data. Biomechanics researchers apply markers to landmarks on study subjects, such as joints and limb mid-points. A camera system captures the positions of these markers over time as the subjects walk or run and are subjected to carefully controlled trips and slips on a specially-equipped treadmill. After these data are collected, humans must examine the Vicon output -- which measures in the thousands of records for a single study subject -- to identify and label important points in the motion of study subjects, such as heel strikes and foot lift-offs. This process is labor-intensive, error-prone, and subject to variation between different human labelers. We are training machine learning models using thousands of pre-labeled trials. Our goal is to use these models in an AI system that performs this labeling automatically, which should increase the consistency of the data and save thousands of hours of valuable researcher time.

• Automated missing marker injection. It is common for motion capture markers on study subjects to be missing in Vicon output data. Markers can fall off, be blocked from the cameras' views by various obstructions, or be forgotten by researchers during trial preparation. Before the data can be further processed, e.g., step labeling, researchers must manually examine the data for missing markers and add the missing markers to the Vicon data using their best judgment. We are training machine learning models on thousands of trials containing complete marker data, some of which was manually added post-capture. Using a self-supervised training process for each marker, these models will impute the location of each marker given the positions of subsets of other markers. Our goal is to use this model in an AI system that will perform missing marker injection, saving thousands of researcher hours.

The ultimate goal of the work above is to produce AI systems to mitigate fall risk:

• On-device assessment. The biomechanics researchers we're working with have hundreds of trials assessing the fall risk of various populations under carefully controlled conditions. If we got these same study participants, or similar subjects to wear a fitness tracker for an extended period of time, we may be able to develop machine learning models that would assess fall risk from everyday movement data. Additionally, these biomechanics researchers are developing fall risk-reducing interventions based on their detailed fall assessments. If the data collected from fitness trackers is of sufficiently high quality and resolution, subject-specific coaching could be provided automatically to reduce their fall risk.

Future Ideas in Health and Wellness

• Hospice prediction. Often when a patient with a combination of conditions nears the end of their life, the various signals available to doctors do not individually indicate that the end is near and hospice care is appropriate. In these cases patients may be subjected to interventions that will not prolong their lives but subject them to considerable discomfort, pain and stress -- just at the time when care should be focused on making the patient comfortable and peaceful. It may be the case that there is sufficient data in existing databases to train an AI system to detect these cases and provide doctors with an additional synthesized signal to use in determining the most appropriate care. This AI system would, of course, be only one part of an ensemble of inputs available to doctors, much in the same way that AI systems provide one of the many inputs radiologists use to evaluate imaging for the potential presence of cancer. Such a system could help doctors not only improve the final stages of the lives of patients, but also the trauma that can be experienced by loved ones who witness fruitless but painful interventions while not getting the chance to peacefully say their good-byes in their loved-ones' final hours.

• Therapy design. In a previous project, the PI led a small team of researchers that implemented the MIMIC algorithm and used it to design planar-array antennas. His lab at Georgia Tech Research Institute had been using genetic algorithms for this purpose, but the genetic algorithms frequently got stuck in local minima and had to be "re-seeded" with human-designed candidates. Our system found optimal designs using far fewer computational resources than the genetic algorithm and did not suffer the same local-minimum problem. The MIMIC algorithm develops an information-theoretic model of the relationships between components of candidate solutions, and uses this model to generate new candidates. There are at least two requirements for applying such algorithms to design problems: (1) a feature vector formulation of candidate solutions, and (2) the availability of a computational assessment of candidates, such as a high-fidelity simulation or machine learning model. There are likely many problems in the health sciences, such as drug design, in which an AI system can design promising candidate treatments for further assessment using traditional health science procedures. We are interested in making advancements in evolutionary optimization algorithms and applying them to problems in health care.

Funding

We are currently funded with a little bit of the PI's startup funding and a mountain of passion. We are seeking external funding from a variety of agencies. Please contact Professor Simpkins at christopher.simpkins@kennesaw.edu if you are interested in funding our work.