ADS Capstone Chronicles Revised
1
Adverse Drug Reaction Surveillance: A Precision Public Health Model
Halee Staggs Applied Data Science Master’s Program Shiley Marcos School of Engineering/University of San Diego hstaggs@sandiego.edu
Vicky van der Wagt Applied Data Science Master’s Program
Shiley Marcos School of Engineering/University of San Diego vvanderwagt@sandiego.edu
ABSTRACT Almost half of Americans take prescription pharmaceutical drugs every month. Side effect profiles and warnings issued bydrug manufacturers are limited by clinical trial data, and therefore lack variance for individual differences. To track the epidemiological impacts of drugs, theFood and DrugAdministration(FDA)createdthe Adverse Event Reporting System (FAERS) whichreceivesmillionsofreportseachyear of adverse reactions to pharmaceutical drugs, underscoring the size of the public health burden. Many systems have been developedtomodelsideeffectsandadverse drug reactions based on FAERS data, but they lack interpretability of individual differences and economic impacts. In this project,anextract-transform-loadpipelineis implemented that synthesizes multiple sources ofpublicdataaboutpharmaceutical drugs into a publicly accessible SQL database (FAERS, Medicaid drug prices, RxNorm).Dataisqueriedfromthedatabase to train, tune, and test machine learning modelstoclassifyoutcomesofadversedrug reactions (ADRs). The optimal pre-trained model (random forest) is stored in GitHub and deployed in a user-friendly Streamlit application. The data is also loaded into PowerBI as an interactive dashboard. The model reveals novel insights that are clinicallyrelevantandaddaprecisionpublic
health approach to adverse drug event outcomes. Individual differences like age, weight, and sex, and economic factors like drug prices, are the primary, significant features for classifying adverse outcomes, with pharmaceutical drugs adding less contributiontothemodel.Morespecifically, younger, overweight females are at highest risk for death, compared to older, lower weight males, implying that differences in pharmacokinetics are related to outcome seriousness. These results show that individual difference input features can informpatient-focuseddecisions,ratherthan modeling side effects and drug compounds only whichlackcontext.Itisrecommended that the FAERS should require more biopsychosocial variables for reporting so that other factors can be assessed in ADR outcomes like socioeconomic status, underlying health conditions, stress, and substance use. Other recommendations for improvement in the FAERS system and tracking true population rates ofADRs are discussed. KEYWORDS adverse drug reactions, pharmaceutical drugs, machine learning, data mining, data engineering, precision public health, side effects, drug prices
151
Made with FlippingBook - Online Brochure Maker