Current Research

Research Spotlight

Decision Support Systems

Novel physiological monitoring systems producing highly informative signals and images, as well as machines collecting molecular data with high-throughput techniques, have provided researchers with a wealth of new knowledge. However, in order to conduct an in-depth study of the causes and mechanisms for many diseases and illnesses and design reliable decision support systems, the analysis of the data from a single high-throughput system or a single image/signal modality may not suffice. It has become evident that integration and processing of all sources of information, from molecular to macroscopic level is the only method of forming an accurate model and therefore generating reliable clinical recommendations.

On multiple clinical applications and in collaboration with researchers from diverse fields of research, we have been designing integrative frameworks to create computational models that integrate and process all typical sources of information including clinical images, clinical signals, proteomic and genomic data, and prior clinical knowledge, in order to form a single multi-faceted model of the process under study. Examples of such ensemble decision support systems include designing computational models for the wound healing process, signal processing methods to predict hemodynamic instability, computer aided decision support systems for traumatic brain injuries and computer aided decision support systems for traumatic pelvic/abdominal injuries.

A Fully-Automated Endoscopic Scoring System for Ulcerative Colitis

In this project, a computer-aided system for colonoscopy video analysis is under development, which will facilitate disease severity measurement and aid in treatment selection and clinical outcome prediction.

Clinical Lead: Ryan Stidham

Project Team: Lingrui Cai, Cristian Minoccheri, Flora Rajaei, Lara King, Dian Jiao, Sophia Tesic

Inactive Project Team: Heming Yao, Reza Soroushmehr, Jonathan Gryak, Jie Gui, Shuyang Cheng

Colonoscopy is a standard medical examination used to inspect the mucosal surface and detect abnormalities of the colon. Objective assessment and scoring of disease features in the colon are important in conditions such as colorectal cancer and inflammatory bowel disease. However, human disease assessment and measurement is hampered by interobserver variation and several biases.

In this project, a computer-aided system for colonoscopy video analysis is under development, which will facilitate disease severity measurement and aid in treatment selection and clinical outcome prediction. The proposed system consists of non-informativeness classification, Mayo severity estimation, image-wise biopsy forceps detection, image-wise location estimation, image segmentation and camera’s motion tracking. By analyzing image features and camera motion, we could derive the contextual understanding of the colonoscopy video such as severity and surgical regions distribution, which can provide richer information for patient’s condition evaluation and outcome prediction.

Improving Care for Heart Failure Patients Using Tropical Geometry and Soft Computing

In this project, we aim to develop an automated decision-making system that can identify patients eligible for an HT/MCS that would facilitate primary care physicians or general cardiologists referring those patients for consideration of advanced therapies.

Project Lead: Yufeng Zhang, Oviyan Anbarasu

Clinical Leads: Jessie Golbus,  Keith Aaronson

Inactive Project Team: Heming Yao, Jonathan Gryak, Renaid Kim, Harm Derksen, Shuyang Cheng, Cristian Minoccheri

Determining the appropriate timing of mechanical circulatory support (MCS) device implantation or heart transplantation (HT) for patients with advanced heart failure is essential as there may be a mortality cost to delayed care. Recognizing patients eligible for an HT/MCS requires expertise from providers trained in transplant cardiology. In this project, we aim to develop an automated decision-making system that can identify patients eligible for an HT/MCS that would facilitate primary care physicians or general cardiologists referring those patients for consideration of advanced therapies. We have developed a hybrid of fuzzy logic and neural networks with a genetic algorithm to identify patients that are eligible for HT/LVAD. Our model achieved better classification performance than other traditional ML algorithms. To overcome the optimization difficulty and computational burden in the existing model, we are developing a novel soft computing architecture with tropical geometry. The proposed network can be not only optimized by gradient-based methods but also close to piecewise linear to ensure the model’s interpretability.

Non-Invasive Measures of Intracranial Pressure in Traumatic Brain Injury Patients

The proposed system would automatically estimate intracranial pressure from head CT scans for traumatic brain injury patients. 

Project Lead: Emily Wittrup

Clinical Leads: Erica Stein, Craig Williamson

Project Team: Haoyuan Ma, Christine Geng, Ethan Schnathorst

This research project focuses on improving how doctors diagnose and treat severe brain injuries, which can be caused by accidents such as car crashes. When someone sustains a brain injury, it is crucial for the doctor to know the pressure inside their skull within the first few hours after an injury occurred. High pressure inside the skull can cause additional harm, so measuring it accurately is important and can help doctors make decisions about how to treat the patient. Right now, doctors use invasive methods that are risky and time-consuming. The researchers want to find a safer and quicker way to measure the pressure inside the skill using medical scans of the head that are already done when a patient arrives at the hospital with a brain injury. The researchers are developing a computer program that can automatically analyze these scans and measure a specific part of the eye that swells when there is high pressure in the skull. This new method aims to give doctors a fast and reliable estimate of the skull pressure, helping them make better decisions about the patient's treatment. It can also help predict negative outcomes so that the doctor can prioritize interventions leading to a better recovery. 

Physiologic Closed Loop Control System for Diuretic Management

This project aims to develop a closed loop control system using explainable artificial intelligence methods to regulate diuretics for pediatric patients in the intensive care unit.  

Project Lead: Matthew Hodgman

Clinical Lead: Daniel E. Ehrmann

Fluid overload is extremely harmful to patients and health systems and is associated with morbidity, mortality, and long-term sequalae that extend beyond discharge from the cardiac intensive care unit. Each day of failure to achieve net negative fluid balance after congenital cardiac surgery prolongs length of mechanical ventilation and hospital stay, which significantly increases hospital costs. The cornerstone of therapy for post-operative fluid overload is initiation and titration of diuretics (e.g., furosemide). However, this practice is imprecise and varies significantly between providers and heart centers, commonly leading to either under- or over-diuresis. Our multidisciplinary team proposes to solve this problem by building a system that improves how diuretic infusions are managed after congenital cardiac surgery. Specifically, we will leverage innovative computational techniques to develop a physiologic closed loop control system capable of automatically titrating a furosemide infusion to manage fluid overload. By using a PCLCS to titrate a diuretic infusion consistently and precisely, we aim to manage fluid overload more safely, effectively, and efficiently than the current standard of care. 


In-Vehicle Systems

The race is on to develop autonomous vehicles, with traditional automotive manufactures competing against technology companies both small and large. There is a wide range of what is meant by an "autonomous vehicle". The Society of Automotive Engineers (SAE) has sought to standardize this terminology by devising a scale (J3016) that delineates the degree of automation within a vehicle. The scale ranges from 0 (no automation) to 5 (full automation). Levels 1 and 2 contain technologies that are already commonplace within vehicles, such as cruise control and parking assistance. However, these levels of autonomy still assume that the driver is ultimately responsible for monitoring the driving environment. From level three and up, the responsibility for monitor the environment, including the driver's fitness to operate the vehicle, is placed upon the in-vehicle system.

The ability of a driver to safely operate the vehicle is comprised of many factors, including the driver's attention to the road, level of alertness, and health. By integrating various imaging and health sensors, along with analytical software, into the vehicle’s onboard computing systems, vehicles at all levels of autonomy can improve driver safety and reduced accidents and mortality. Even when vehicles reach full autonomy (level 5), there will still be a need for vehicles to monitor the health and wellbeing of its passengers, and potentially alert public safety officials or health experts of threats to physical safety or medical emergencies.

Development and Assessment of an In-Vehicle Cardiac Monitoring and Severe Event Prediction System

The proposed system will accurately predict or detect the onset of severe cardiac events (e.g., hemodynamic instability) prior to complications experienced by the driver in order to ensure driver safety.

Project Team: Matthew Hodgman, Emily Wittrup, Emmalyn Campau

Clinical Leads: Kevin Ward, Michael Mathis

Inactive Project Team: Jonathan Gryak, Zhi Li, Alexander Wood, Julia Pagnucco, Gang Liu, Zixue Zeng, Harm Derksen, Hamid Ghanbari, Winston Zhang, Shuyang Cheng, Cristian Minoccheri, Ali Zare

Drug Development

The realm of drug discovery is undergoing a transformative evolution through the integration of machine learning. By fusing the precision of computational algorithms with the complexities of biochemical interactions, we are ushering in a new era of pharmaceutical innovation. Our approach involves leveraging vast datasets encompassing molecular structures, biological activities, and clinical outcomes to train sophisticated models. These models, powered by machine learning, unravel intricate patterns and hidden insights that human analysis alone might overlook. This synergy between artificial intelligence and pharmaceutical expertise expedites the identification of promising drug candidates, significantly expediting the research process. As we stand at the intersection of technology and medicine, our mission is to reshape the landscape of drug discovery, accelerating the delivery of safe and effective treatments to patients in need.

Utilizing Fully Homomorphic Encryption for Privacy Preserving Machine Learning in Drug Development

The goal of this project is to develop a privacy-preserving machine learning system that would allow institutions to benefit from shared knowledge without sharing sensitive data. 

Project Team: Cristian Minoccheri, Lara King, Chaewon Lim, Dian Jiao

Machine learning has found great success in developing clinical decision support systems for a variety of drug design, repositioning, and repurposing applications. To properly train and validate ML methods, large and diverse datasets are needed, however these datasets are rarely available to a single entity (i.e., a pharmaceutical company) which may not be large and diverse enough to create reliable and robust ML models. This suggests that reliable ML models can be achieved often only when multiple entities can share their datasets to form larger training and validation databases. This requirement is one of the main reasons why the use of such ML systems in the multi-institutional settings is challenging, i.e., ML methods require sharing sensitive data whose access is severely restricted by informed consent, institutional guidelines, intellectual property restrictions and federal regulations. The proposed solution would create a privacy-preserving machine learning (PPML) system which can be readily adapted for detection of drug design, efficacy, toxicity, and adverse events in multicenter / multi-institutional trials without compromising patient privacy, data security, and proprietary clinical data.