DISCIPLINE Projects

The most common non-cardiac cause of readmission after cardiac surgery is infection, which is associated with increased morbidity and mortality, hospital cost, length of stay, and readmissions. Using large-scale national databases from the Society of Thoracic Surgeons (STS) in pediatric and adult cardiac surgery, we will use novel machine learning (ML) methods to predict the risk of infection and infection-related readmissions. Our team has successfully collaborated on multiple federally funded projects with extensive experience in both real-time ML integrating within health systems and implementation trials leveraging near-realtime ML models for surveillance. This project will create a predictive toolkit (iPACS: Infection after Pediatric and Adult Cardiac Surgery) that can be embedded into electronic health records (EHR) to better identify pediatric and adult patients at high risk for 30-day infection after cardiac surgery.

There is wide variation of clinical practice among providers across regions, specialties, and over time.  It should be noted, while providers agree that practice variation should be reduced to improve overall care, some variation is not bad and reflects patient preferences and clinical practice in the absence of evidence-based data to support such standardization.  Once assessment of practice variation and estimation of concordance with guideline-based practice is done, encouragement for provider behavior change needs to be done.  Mixed modalities such as didactic education, academic detailing, and being allowed time to quickly look up relevant details surrounding an encounter, all helped promote care delivery alignment with guidelines.  In addition, clinical reminders to providers during encounters help to improve delivery of care but are taxing on providers’ time and are not scalable to the volume and breadth of care.  While most provider care variation is performed between providers, there is a growing awareness of implicit bias within a provider’s practice that results in variation within a provider’s practice.  In both cases, machine learning (ML) and data science may be able to synthesize a provider’s practice and provide relevant, timely, and personalized information that does not need a high amount of manual rule-based development and maintenance to impact their clinical care.  This project seeks to develop a generalizable framework to assist a provider in understanding and improving their clinical practice patterns leveraging novel data science and ML methods and tools.

Failing to proactively address deteriorating accuracy of clinical machine learning (ML) and artificial intelligence (AI) models risks patient safety, undermines user trust, wastes limited infrastructure and implementation resources, and fails to deliver on the promise of predictive analytics to improve patient outcomes. An extension of the learning health system paradigm of data-driven continuous improvement, learning prediction systems (LPS) would conduct post-deployment surveillance to collect evidence of model success or deterioration and recommend changes to improve prospective performance and outcomes for patients, providers, and health systems. Our overarching goal is to inform the design and implementation of LPS that respond to the dynamic nature of clinical environments through active model maintenance that enables sustainable, robust, and equitable (EnSuRE) AI/ML-based clinical decision support. In the EnSuRE project, we will apply advanced mixed methods and sequential analyses to develop novel analytic approaches to model surveillance and updating; characterize and support information needs across stakeholders; and establish best practices for LPS.

Chronic Obstructive Pulmonary Disease (COPD) is a leading cause of mortality and morbidity in the United States, affecting around ~24 million people. People with COPD  face many barriers that interfere with or prevent therapeutic management and medication adherence compliance. These barriers can lead to exacerbations and episodes of worsening symptoms requiring hospitalization. In this research project will develop a Clinical Decision Support System to support healthcare providers in medication non-adherence (MNA) management in patient with COPD. Our system will use natural language processing (NLP) to extract relevant SDoH data and interpretable machine learning (ML) models to predict MNA risks and rank contributing factors. By providing healthcare providers with insights into the SDoH factors and clinical characteristics associated with MNA, the system can help providers better understand and address their patients’ needs. This can lead to improved medication adherence, reduced exacerbations, and better overall outcomes for COPD patients.