This project will create, test, implement, and evaluate a real-time Adverse Drug Effect (ADE) alerting system at Vanderbilt University Hospital (VUH). Utilizing information extracted from admission history and physical examination notes (H&Ps) stored in the electronic medical record (EMR), our system will detect adult inpatients' previously unrecognized symptomatic ADEs and alert appropriate care providers. The project will create an ADE knowledge base, mined from multiple publically available sources including MEDLINE, RxNorm, and the product labels for human prescription drugs. Next, we will apply natural language processing (NLP) to EMR-based H&P texts to identify mentions of patients' current medications and dosages. We will similarly detect EMR-based H&P documentation of patients' clinical manifestations (CMs - diseases, symptoms, findings, etc.), and then represent them using the Unified Medical Language System's (UMLS) Concept Unique Identifiers (CUIs). The system will then compare each patient's recognized medications and CMs against the aforementioned ADE knowledgebase, generating appropriate patient-specific alerts for potential adverse effects related to the patient's current medications. Each alert will identify the offendin medication, the suspected nature of the ADE, and evidence supporting the assertion. We will independently evaluate the accuracy of the ADE knowledgebase using expert manual review and comparison to multiple other sources. Before implementing the real-time ADE monitoring system, we will conduct a pilot implementation using retrospective EMR data from the Vanderbilt Synthetic Derivative (SD), a de-identified version of the Vanderbilt EMR. After successful testing and any necessary iterative improvements, we will implement the real-time detection system at VUH, initially monitoring newly admitted inpatients presenting to the Internal Medicine service. Using the methods described above, the system will detect potential ADEs each time a new inpatient H&P is generated, and alert appropriate clinicians via additions to an existing dashboard that the clinicians already utilize. We will survey clinicians immediately upon receipt of an ADE alert to determine if the alerting condition was already known or not, whether the alert seems plausible, and whether it requires intervention. Then we will compare discharge medications to admission medications to determine what actions occurred post-alert, independent of physician survey results. We will thus evaluate both the effectiveness of the system in improving ADE recognition and its perceived usefulness according to the physician-subjects. We hypothesize that our system will improve clinicians' awareness of ADEs in a manner applicable to any facility that stores admission H&Ps electronically. Addressing previously unrecognized ADEs has the potential to reduce costs and improve patient care.
Unrecognized adverse drug effects (ADEs), a serious clinical problem, cause preventable hospitalizations, increase healthcare costs, and worsen health outcomes. To potentially improve quality of care, my dissertation project will develop and evaluate a novel system to detect adult inpatients' previously unrecognized symptomatic ADEs and alert appropriate care providers. Using automated natural language processing of clinician-generated electronic admission history and physical exam (H&P) notes and a locally-developed ADE knowledgebase derived from publically available sources, the system, once validated, could improve both recognition and treatment of ADEs in a generalizable manner - applicable in hospital environments using electronic medical records.
