Methods: RxNorm was serialized into FHIR Medication, MedicationKnowledge, and Substance resources. The new system, MedXN-FHIR, used the FHIR resources to build a drug vocabulary. In addition, the FHIR data models were utilized for ontology-based reasoning. Both MedXN and MedXN-FHIR were used to annotate a test set of 251 discharge summaries. The system annotations were evaluated against i2b2 ground truth annotations for precision, recall, and F1-measure.
Conclusions: A drug terminology was transformed into FHIR resources, and the same resources were reused as a drug ontology to normalize unstructured medication names and attributes through NLP techniques. Further insight into the performance of MedXN-FHIR in assigning RxNorm identifiers would be useful.
Results: Both systems exhibited good performance with F1-measures above 0.8 on medication name, dose, and frequency fields. For medication name, MedXN-FHIR produced an increase in precision of 0.0373 over MedXN. Overall, MedXN-FHIR produced a higher precision but lower recall than MedXN for most attributes. MedXN-FHIR performed worst for duration, with a decrease in F1-measure of 0.204.
Objectives: To develop a medication natural language processing (NLP) system that was derived from MedXN and to evaluate its performance in annotating medication name, dose, route, frequency, and duration using the 2009 i2b2 Medication Extraction Challenge dataset.
|In Administrative Set:|