Watch the Video of the Lightning Round Presentations:
Integrating AI to Facilitate Clinical Decision Making in Speech Pathology Graduate Programs
Jessica Wacker, Speech Language Pathology, College of Science and Health
The rising intricacy of clinical caseloads, combined with the heightened emphasis on evidence-based practice (EBP) in speech-language pathology (SLP) education, calls for creative strategies to enhance the efficiency of clinical decision-making. Integrating artificial intelligence (AI) into graduate speech pathology programs can enhance clinical operations, improve decision-making, and optimize student learning. AI solutions can help students and clinical educators make informed decisions, leading to better client outcomes and giving students valuable real-world experience in a structured setting. The main goal of incorporating AI into the Speech Pathology Graduate Program is to improve clinical education by providing a helpful tool for:
- Data-Driven Clinical Decision-Making: AI-powered analysis of patient data and research to support diagnosis, treatment planning, and EBP recommendations.
- Streamlined Clinical Operations: AI will reduce administrative burden by assisting with documentation, case management, and scheduling.
- Optimized Student Learning: Real-time feedback on student clinical decisions, promoting critical thinking and improving diagnostic accuracy.
- Evidence-Based Practice Integration: AI will support the incorporation of the latest research and clinical guidelines into student training.
Leveraging Latent Space Diffusion Models for Enhanced Drug Discovery
Tianxiang (Adam) Gao, School of Computing, Jarvis College of Computing and Digital Media
This proposal introduces a novel framework that integrates Variational Autoencoders (VAEs) and diffusion models to enhance drug discovery through AI. The approach begins by embedding molecular graphs into a latent space using a VAE, which captures essential molecular features and structures. In this latent space, a diffusion model is employed to generate new compounds, allowing for the exploration and optimization of molecular properties within a continuous representation. This method effectively addresses challenges in drug discovery, such as ensuring compound validity and exploring diverse chemical spaces. By conditioning the diffusion model on specific molecular attributes, we can guide the generation toward drug-like compounds with desired therapeutic characteristics. This approach not only reduces computational costs but also accelerates the discovery of potential drug candidates. Ultimately, it provides a scalable and innovative solution for AI-assisted drug design, offering substantial potential for interdisciplinary collaboration and advancements in healthcare and scientific research.
Using AI to Study the Cognitive Effects of Bilingualism
Casey Bennett, School of Computing, Jarvis College of Computing and Digital Media
In my past work, we’ve done a lot of work using artificial speech systems (i.e. conversational AI) to study the effects of bilingualism on cognition, particularly looking at Korean and English, as well as a few other languages like Japanese. That was done by creating robots and virtual avatars that could interact with people through conversation in different languages, then modifying that interaction somehow to see if the effects were the same across languages. However, we are interested in exploring other languages, e.g. Spanish, including bilingual and monolingual speakers in those languages. In particular, we are interested in how such language differences in cognition affect certain health conditions where verbal communication difficulties, such as older adults with Dementia & Alzheimer’s as well as younger people with Autism.
Preliminary Analyses Using Machine Learning to Predict Early-Career Police Patrol Officer Retention
Kayla Freemon, Department of Criminology, College of Liberal Arts and Sciences
Police departments are facing a historic crisis in recruiting and retaining officers. While researchers have identified numerous factors that influence officers’ decisions to stay or leave their position – including their relationship with their supervisor, compensation, burnout, and work stress – most are subjective and can only be collected qualitatively through conversations and interviews or surveys. To address this gap, the current study employs a data-driven machine-learning approach using indicators from department-collected administrative records to predict officer retention for patrol officers in their first three years on the job. In preliminary analyses using a random forest approach, I examine 1,105 new hires in the Phoenix Police Department between 2015 and 2022, with 11% leaving their role in the first year and 20% having left their role by their third year employed. Early identification of officers at risk of leaving may help departments address individual workplace concerns and keep good officers on the job.
A Knowledge Base for Running Distributed Science Applications
Joseph Phillips, School of Computing, Jarvis College of Computing and Digital Media
Scientists have a wealth of analysis tools. This is because small Python programs can attack interesting problems, and because GitHub serves as the universal location where scientist-developers distribute their code. However, not all scientists have savvy to use these tools. Potential users must install the necessary software. And, if operating systems differ, they must administer virtual machines. Our knowledge base web application addresses this. Beyond serving as a web interface between users and command-line driven software, the system organizes and helps interpret output. The system is:
- Distributed. Requests are sent to machines (physical or virtual) without users knowing or caring how.
- Multi-functional. The system is designed as a general interface to many specialized applications. The output of one program may be used as input to another.
- Customizable. We handle disagreements about the organization of ontologies by allowing users to pick one or customize their own.
- Uniform. Users see one interface for all the software. This leverages and extends the user-base of existing software and APIs.
The system will be used by undergraduate biologists at DePaul University to assay the distribution of aquatic organisms around Chicago. Environmental DNA (eDNA) is used to track the spread of invasive species, to track changes in the distribution of rare and endangered species, and to better grasp the structure of aquatic communities. Our tool is a knowledge-aware interface for relevant tools, programs and APIs, to perform these tasks, and allows bench scientists and undergraduates to analyze thousands to millions of sequences.
Constrained SVD for Biological Structure and Function: A Ten-Year Retrospective
Eric Landahl, Physics and Astronomy, College of Science and Health
Just over ten years ago a DePaul / Northwestern Feinberg School of Medicine collaboration released constrained Singular Value Decomposition (cSVD), an early machine learning algorithm for analysis of 2D biochemical and biophysical data. I will review some of the successes of the algorithm and its impact on fields from x-ray scattering to vaccine development. Although cSVD has been supplanted by modern ML and AI methods and is no longer a workhorse tool, its record of accomplishment should inspire those looking to apply new data analysis approaches to health and science research. Furthermore, the algorithm’s lightweight computational demands and intuitive approach may still make it attractive for some problems.
Learning as a Replacement of Formal Methods
Umer Huzaifa, School of Computing, Jarvis College of Computing and Digital Media
Formal methods provide the foundation for modeling and designing engineering systems, ensuring reliable and systematic control for safe operation. However, these methods often require specialized expertise to develop accurate models and can be overly idealistic. Take the example of a robot arm. This system consists of complex mechanical hardware with interconnected joints and movement possibilities. The first step in controlling the arm is to accurately model these connections, allowing us to predict how a link will move when a specific torque is applied to a joint. Accurately predicting the outcome requires deriving the system’s movement space using tools from mechanics and multivariable calculus, leading to what is known as the equation of motion. Using this equation, we can predict the movement for any combination of input torques, assuming the system is not affected by factors such as unmodeled friction, flexibility, or wear. To address the limitations of idealism and the need for extensive theoretical background, machine learning (ML) and artificial intelligence (AI) offer an alternative approach. or do they? ML and AI rely on experimental data, tuning models based on actual performance rather than mathematically predicted behaviors like those found in equations of motion. The key question, then, is whether machine learning can effectively tackle this complex problem. Specifically, can we model a robot solely based on systematically collected data and derive general rules for how the links should move in response to input torques? In this lightning talk, I will explore this concept with examples from robotics and biomedical engineering, highlighting the importance and potential of machine learning in this context.
Using AI Projects with Community Partners to Promote Student Learning and Engagement
Mark Potosnak, Environmental Science, College of Science and Health
The Metropolitan Chicago Data-science Corps (MCDC) is a collaboration of non-profit or community organizations and data science students and experts from several Chicago-area universities funded by the National Science Foundation addressing challenges and opportunities around extracting insights, deemed relevant by the community, from data. Each summer project consists of a primary contact at the community organization, two to four undergraduate students, an expert faculty mentor, and optionally an expert co-mentor who is studying in a Master’s program. Teams are assembled so they have appropriate and complementary experience and expertise. Community partners can also work with a practicum course at one of our dedicated universities throughout the school year. There the learning will be facilitated within the classroom by the mentor with a larger number of students. These projects fall within the schedules of the universities with which they are paired. We strive to encourage early communication around these projects so that there is ample time to prepare for the course needs and structure.