ADS Capstone Chronicles Revised

‭24‬

‭Figure 18‬ ‭Predicting Patient Risk‬

‭5.4 Apache Airflow Model Maintenance‬ ‭Because‬ ‭the‬ ‭FAERS‬ ‭data‬ ‭is‬ ‭refreshed‬ ‭two-to-four‬ ‭times‬ ‭per‬ ‭year,‬ ‭the‬ ‭optimal‬ ‭classification‬ ‭model‬ ‭will‬ ‭be‬ ‭retrained‬ ‭and‬ ‭deployed‬ ‭to‬ ‭the‬ ‭application‬ ‭and‬ ‭dashboard‬ ‭based‬ ‭on‬ ‭the‬ ‭most‬ ‭recent‬ ‭data.‬ ‭Apache‬ ‭Airflow‬ ‭is‬ ‭used‬ ‭to‬ ‭refresh‬ ‭the‬ ‭data‬ ‭every‬‭4‬‭months‬‭starting‬‭on‬‭August‬‭23,‬‭2024‬ ‭using‬‭airflow‬‭scheduler‬‭with‬‭directed‬‭acyclic‬ ‭graph‬ ‭(DAG)‬ ‭parameters‬ ‭on‬ ‭a‬ ‭local‬ ‭command‬ ‭line‬‭interface.‬‭Model‬‭drift‬‭will‬‭be‬ ‭assessed‬ ‭by‬ ‭reviewing‬ ‭a‬ ‭table‬ ‭of‬ ‭outputted‬ ‭model‬ ‭performance‬ ‭metrics.‬ ‭See‬ ‭CodeLibrary/ModelMaintenance‬ ‭for‬ ‭associated‬ ‭files.‬ ‭ChatGPT4.o‬ ‭was‬ ‭used‬ ‭as‬‭a‬ ‭tool for figuring out airflow implementation.‬ ‭6 Discussion‬ ‭6.1 Conclusion‬ ‭This‬ ‭project‬ ‭created‬ ‭an‬ ‭end-to-end‬ ‭data‬ ‭science‬ ‭ETL‬ ‭pipeline‬ ‭that‬ ‭creates‬ ‭a‬ ‭persistent‬ ‭data‬ ‭store‬ ‭of‬ ‭Adverse‬ ‭Drug‬ ‭Reaction‬ ‭events‬ ‭from‬ ‭FAERS‬ ‭in‬ ‭an‬ ‭SQL‬ ‭database.‬ ‭The‬ ‭API-sourced‬ ‭data‬ ‭in‬ ‭the‬ ‭database‬ ‭supports‬ ‭machine‬ ‭learning‬ ‭objectives‬ ‭and‬ ‭a‬ ‭dashboard‬ ‭application‬ ‭in‬ ‭PowerBI‬ ‭and‬ ‭Streamlit.‬ ‭The‬ ‭quality‬ ‭of‬ ‭the‬ ‭tools‬ ‭are‬ ‭assessed‬ ‭four‬ ‭times‬ ‭per‬ ‭year‬ ‭to‬ ‭maintain relevance.‬ ‭This‬ ‭system‬ ‭can‬ ‭be‬ ‭used‬‭for‬‭surveillance‬‭of‬ ‭actively‬ ‭marketed‬ ‭drugs.‬ ‭The‬ ‭FDA,‬ ‭prescribing‬ ‭professionals,‬ ‭and‬ ‭the‬ ‭public‬ ‭could‬‭benefit‬‭from‬‭identifying‬‭drugs‬‭causing‬ ‭serious‬ ‭outcomes‬ ‭or‬ ‭death,‬ ‭including‬ ‭the‬ ‭frequency‬ ‭and‬ ‭demographics‬ ‭of‬ ‭people‬ ‭suffering‬ ‭from‬ ‭adverse‬ ‭reactions.‬ ‭This‬ ‭precision‬ ‭public‬ ‭health‬ ‭model‬‭offers‬‭a‬‭more‬ ‭useful‬‭clinical‬‭application‬‭compared‬‭to‬‭other‬ ‭models‬ ‭assessing‬ ‭only‬ ‭side‬ ‭effects,‬ ‭which‬ ‭cannot‬ ‭be‬ ‭used‬ ‭for‬ ‭preventative,‬ ‭precise‬ ‭healthcare decisions.‬ ‭The‬ ‭importance‬ ‭and‬ ‭interpretability‬ ‭of‬ ‭this‬ ‭system‬ ‭is‬ ‭that‬ ‭it‬ ‭provides‬ ‭evidence‬ ‭that‬

‭Table 7‬ ‭Top Five Drugs by Outcome‬ ‭Outcome‬ ‭Drug‬ ‭NDC 9‬

‭p̂‬

‭NonSerious‬

‭Dupixent‬

‭0024-5911‬

‭.72‬

‭NonSerious‬

‭Ezetimibe‬

‭68462-226‬

‭.081‬

‭NonSerious‬

‭Ocrevus‬

‭50242-150‬

‭.073‬

‭NonSerious‬

‭Cabometyx‬

‭42388-023‬

‭.066‬

‭NonSerious‬

‭Revlimid‬

‭59572-402‬

‭.064‬

‭Serious‬

‭Carboplatin‬

‭60505-6282‬

‭.27‬

‭Serious‬

‭Cyclophosphamide‬

‭82943-100‬

‭.21‬

‭Serious‬

‭Paclitaxel‬

‭68817-134‬

‭.18‬

‭Serious‬

‭Rituximab‬

‭50242-051‬

‭.18‬

‭Serious‬

‭Immunoglobulin g‬

‭69800-0250‬

‭.17‬

‭Death‬

‭Rituximab‬

‭50242-051‬

‭.24‬

‭Death‬

‭Prednisone‬

‭75987-020‬

‭.22‬

‭Death‬

‭Acetaminophen‬

‭72288-405‬

‭.19‬

‭Death‬

‭Carboplatin‬

‭60505-6282‬

‭.19‬

‭Death‬

‭Methotrexate‬

‭62135-772‬

‭.17‬

174