Projects

Across all projects we use tools such as Python, Pandas, Jupyter notebooks, Logistic Regression, Random Forest Models, Decision Trees, Boosted Tree Forests and TPOT, an automated machine learning tool.

Predicting Readmission Risk

Principal Investigator: Eric L. Hume, MD
Associate Professor of Clinical Orthopaedic Surgery

We used electronic health records of patients admitted for hip and knee joint replacement surgery to estimate risk of hospital readmission within 30 days of discharge.  We identified key predictors of readmission risk from a wealth of data including financial information, lab results, zipcodes and demographics.

 

Precision Phenotyping of Hypertension

Daniel S Herman, MD, PhD
Assistant Professor Of Pathology And Laboratory Medicine

We have made significant steps toward phenotyping primary aldosteronism, a subset of hypertension that has a distinct treatment protocol.  We have used PennMedicine data including medications, diagnoses, chart notes, and lab results to phenotype hypertension with a higher degree of accuracy than is available through the existing Penn hypertension registry.  Moving forward we are developing workflows to apply our custom algorithm to the entire EPIC database.  We are also exploring available data and data ontologies to identify new features that will lend more precision to our phenotyping efforts.

 

Assessing Risk of Falls in Elderly Home Healthcare Patients

Kathryn H. Bowles, PhD, RN, FAAN, FACMI
Professor of Nursing and vanAmeringen Chair in Nursing Excellence

Together with Dr Bowles and her team we analyzed records of elderly home health care patients to model and predict fall risk.  Our analyses incorporated OASIS clinical survey data and a myriad of other clinical notes, both free text and structured records.  We were able to predict fall at the start of care with an accuracy significantly higher than the existing MAHC questionnaire assessment benchmark.  Our work was incorporated as a critical section in Dr Bowles’ R01 grant which focuses on assessing the efficacy of interventions for preventing falls.

 

Other Work

Additionally we provide service for large-scale data selection and processing, such as identifying from a database with millions of records a demographically matched control cohort of size >100,000 for Dr. Nadia Penrod.