Informatics techniques using clinical data and advanced analytic methods demonstrate significant promise in answering health services questions. There is also great promise in applying machine learning methods to improve clinical practice. However, there are significant challenges in harmonizing the different types of data and harnessing the information throughout the electronic medical record. Part of our diverse portfolio is dedicated to combining statistical and machine learning techniques to analyze claims data and data from electronic hospital systems to improve clinical practices and patient outcomes.

Physician Judgment and Machine Predictions: Improving Medical Decisions Using Machine Learning

In this Robert Wood Johnson Foundation (RWJF 7283) funded project, we study how machine learning based predictive models can be used to improve emergency room physician decisions for sepsis patients.

Smarter Big Data for a Healthy Pennsylvania: Changing the Paradigm of Healthcare

In this project, funded by the State of Pennsylvania, we study the use of big data methods to combine multi-modal data to study clinical event prediction in the hospital, during transitions to the outpatient setting, and at the community level.

UPenn-CareMore Research Collaboration on High-Need Patients and Care Delivery Management

In this project, funded by Anthem, Inc., we utilize CareMore data to identify and categorize high needs patients based on their clinical and demographic characteristics, as well as utilization patterns. We examine the impact of CareMore’s end stage renal disease (ESRD) care model on the quality outcomes and cost of care that is patients receive.

Impact of COVID-19 Pandemic on Cancer Care within the Veterans Health Administration

The objective of this VA-funded pilot study is to assess the impact of the COVID pandemic on access to and utilization of cancer care among Veterans with localized, curable cancers that are primarily treated with surgery and/or radiation.

Segmenting High-Need, High-Cost Veterans into Potentially Actionable Subgroups

The main objective of this VA funded study is to apply statistical and machine learning clustering methods to classify high-need, high-cost (HNHC) Veterans into clinically actionable subgroups based on detailed clinical information extending beyond diagnosis codes.

Racial Bias in a VA Algorithm for High-Risk Veterans (Planning Award)

The objective of this project is to examine unfairness in the VA Care Assessment Needs (CAN) Algorithm and develop approaches to mitigate it, with the long-term goal of developing clinically-applicable, fair risk algorithms that reduce health disparities by more effectively allocating resources to high-risk Veterans.

Smarter Big Data for a Healthy Pennsylvania: Changing the Paradigm of Healthcare

The objective of this project is to improve the value of care and reduce health disparities by developing a set of powerful algorithms to consistently improve upon human clinical judgments. We will begin by addressing the foundational question: how can machine learning methods be used on claims and clinical data to accurately predict future clinical events? Our test case will be detecting sepsis in patients in the emergency department (ED), using existing data infrastructure on sepsis care from the University of Pennsylvania Health System. 

This test case will help us 1) develop a framework for identifying applications of ML to clinical problems related to clear clinical need—areas where doctors currently make poor decisions, 2) develop a generalizable framework for identifying where in clinical decision-making and workflow machine intelligence will be most impactful and most accepted, and 3) identify decisions where knowledge possessed by humans and inaccessible to algorithms improves predictions.