ADS Capstone Chronicles Revised
8
medication use module (HHS, n.d.) to represent the true underlyingfrequenciesof adverse events for particular drugs in relation to severity of outcome. Data was sourced from ADReCS (Cai et al., 2015), FAERS, MEPS from 2012 to 2019, PubChem (NCBI, n.d.),DrugBank(Wishart et al., 2018), Anatomical Therapeutic Chemical (ATC) system (World Health Organization,n.d.),andKyotoEncyclopedia of Genes and Genomes (Kanehisa & Goto, 2000). The variable of interest was clinical severity of ADRs: recovered, recovering, not recovered, resolved with sequelae, and fatal. Using a penalized, cumulativemodel, and common terminology criteria for adverse events’ 5-grade system (HHS, 2017),theauthorsassignednumericalscores of 1-5 to the outcomes, which correspond with mild, moderate, severe, life-threatening, and death. The results provide a new framework of drug classificationandseveritythatwasaddedas an update to ADReCS (Cai et al., 2015). 3.8 Current Limitations All of the models and tools above have similar themes of limitations which are addressed in this project. First, the models lackinformationaboutpatientdemographics like sex, age, and weight. These are often important factorstoconsiderwhenitcomes to adverse reactions to drugs, especially because clinical research has a historical gendergapandpharmacokineticswasfound to be studied 90x more in males than femalesinclinicaltrials(Staggs,etal.,n.d.). Figure 3 shows that adverse events in females are reported at a higher frequency than males, on average. The models explainedabovecannotbeinterpretedinthe context of patients’ needs and circumstances. Second, the models do not consider pricing of drugs as a feature. Controlling for demographic and economic features will addabiopsychosocialangleto
adverse drug events from which individual differences and financial burden can be inferred. Finally, while deep learning approaches are robust, the computational load and model complexity are not accessible, and most of the published models appear to be theoretical and static, instead ofpracticalwithmodelretrainingto stay up to date. The model architecture cannot be interpreted and the high performance is due to modeling obvious relationships between adverse events and druginteractionsandsideeffects,whichare established causal variables by drug manufacturers and chemistry research. The models are also tested with the same benchmark datasets which limits the ability to identify novel insights. Most of the models have low performance metrics for the testing set for multiclass classification, which highlights poor choices in data cleaning,featureengineering,andbalancing
outcome classes during training. Figure 3 ADRSex Representation
3.9 Addressing Limitations In this project, interpretable models are producedwherecoefficientvaluesareuseful and informative. Input features capture individual differences in people and economic features (U.S. Centers for Medicare & Medicaid Services [CMMS], 2024) in addition to specific drug compounds. This is the first step towards a precisionpublichealthapproachtodrugrisk assessment with these specific input features.
158
Made with FlippingBook - Online Brochure Maker