Past Projects

We developed signal processing and machine learning algorithms to 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.

In this project, we detected the presence of the acute respiratory distress syndrome (ARDS) at early stages utilizing electronic health records. The project combined the capabilities of two emerging fields of machine learning, privileged learning and learning from uncertain data, both highly relevant to healthcare applications. While the project focused on improving diagnosis of ARDS, the proposed learning methods will generalize across healthcare settings, allowing for better characterization of patient health status in both in-hospital and in-home settings via portable electronic monitoring devices.

 Traumatic brain injury is a serious public health problem in the U.S. contributing to a large portion of permanent disability. However, its early management and treatment could limit the impact of the injury, save lives and reduce the burden of cost for patients as well as healthcare systems. Subdural hematoma is one of the most common types of TBI, which its visual detection and quantitative evaluation are time consuming and prone to error. In this study, we propose a fully automated machine learning based approach for 3D segmentation of convexity subdural hematomas. Textural, statistical and geometrical features of sample points from intracranial region are extracted based on head Computed Tomography (CT) images.

We have developed wound segmentation and deep neural network models that can accurately predict infection and wound healing. The image segmentation methods provide an accurate “hands-free” measurement of wound surface area. Our objective here was to refine our algorithms and test the suitability of image analysis as a biomarker for healing or infection. We did this by comparing the resulting image features of diabetes-related foot ulcers treated by total contact cast (TCC) and those not treated by TCC.

This project developed construction of AngioAid, a fully automated computer-based platform to assist with the interpretation of coronary angiogram videos. It also proposed a new vessel segmentation method for angiograms using deep learning by formulating a new segmentation probability map achieving higher accuracy, sensitivity, and Dice-score measures than other methods. 

Diagnosis of temporomandibular joint (TMJ) Osteoarthritis (OA) before serious degradation of cartilage and subchondral bone occurs can help prevent chronic pain and disability. Clinical, radiomic, and protein markers collected from TMJ OA patients have been shown to be predictive of OA onset. Since protein data can often be unavailable for clinical diagnosis, we harnessed the learning using privileged information (LUPI) paradigm to make use of protein markers only during classifier training.

The project developed an innovative, real-time clinical decision support (DSS) platform. The platform utilized Big Data analytic methods, novel algorithms, and software tools to integrate and analyze disparate sources of continuous and non-continuous patient data to understand and predict potential complications and recovery trends for patients following cardiac surgery, other open vascular operations in the chest and abdomen, and other trauma surgeries involving the cardiovascular system.

The project aim is to establish a prognostic predictive model using non-convex kernel models to distinguish between benign and malignant tumor patients. This model will provide adjuvant information to stratify osteosarcoma patients and subsequently to guide clinical decision.

The project is a collaboration between CareProgress and the University of Michigan. The project aims to develop computational algorithms to analyze the clinical and self-reported data collected by CareProgress to assess the recovery and effectiveness of treatment and health resource consumption. Improving the accuracy and reliability of the predictions on the recovery and health resource consumption by using the physiological data collected by wearable sensors, including through the creation of patient risk scores.

In this project, the project team at U-M hereafter, in collaboration with DENSO, will create computational methods, in particular deep learning algorithms, to understand and predict human behavior and affective state (distracted/focused, drowsy/active, etc.). The team will implement the above-mentioned algorithms in MATLAB and Python framework.

his project continues the development of the Piezo Ring. The ring portably measures the vascular tone and reactivity of small blood vessels in the finger, which gauge a patient’s physiology and hypotension. Using computational methods, the ring predicts hypotension ranging from 9-108 minutes in advance. 

This is useful for hemodialysis patients who greatly benefit from constant blood pressure monitoring, as well as in combat and critical care situations.

The Piezo Ring is easily portable, provides continuous hemodynamic monitoring, and predicts hypotension up to 108 minutes in advance. The Piezo Ring also provides more relevant data to physicians, making it a better indicator of patient physiology than current devices such as pulse oximeters.

While sleep disruption and disorders are common in individuals with spinal cord injury (SCI), little is known about how these disturbances impact symptoms (fatigue and pain), cognition (subjective and objective), and functioning (physical activity and social participation) in these individuals. Because sleep disturbance has been shown to be associated with poor functional outcomes (i.e., worse symptoms, poorer cognition, less physical activity, and reduced social participation) in other clinical populations such as multiple sclerosis, and these disturbances are readily treatable, there is a need to better understand these relationships in individuals with SCI.