Natural language processing models demonstrated accuracy in detecting axial SpA concepts
Click Here to Manage Email Alerts
Axial spondyloarthritis concepts can be feasibly and accurately identified in the free text of electronic medical record documents, according to findings.
Researchers gleaned historic data on U.S. veterans enrolled in the Veterans Health Administration. Data were collected through routine care encounters and administrative procedures. Patients eligible for analysis had documents in text integration utilities clinical notes or radiology findings within the corporate data warehouse.
Researchers generated methods of identifying and extracting meaningful terms that were likely to represent concepts expected to be predictors of axial spondyloarthritis (SpA) diagnosis. They assessed and extracted sections of text that included these meaningful terms (snippets). The snippets were then annotated and categorized based on whether these exemplified the intended SpA concept. Natural language processing (NLP) tools were then used to teach computers to reproduce the snippet classifications rendered by the clinical experts. The clinical experts selected three axial SpA concepts, including sacroiliitis, terms beginning with the prefix “spond*,” and positive HLA-B27 status.
For the three axial SpA concepts, the researchers developed three NLP models. Using 10-fold cross-validations, they estimated classification accuracies of 91.1% for the sacroiliitis NLP model; 93.5% for spond*; and 97.2% for the HLA-B27+ model.
Independent validation revealed classification accuracies of 92% for the sacroiliitis model, 91% for the spond* model, and 99% for HLA-B27+
“The novel methods developed with this research generated accurate NLP models for identifying axial SpA concepts in clinical notes. Additional work is necessary to further validate these methods in external data sets and to determine combinations of concepts that will identify patients with axial SpA and axial SpA subtypes. These methods will facilitate observational research in axial SpA and may be applied to research with other diseases.” - by Jennifer Byrne
Disclosures: Walsh reports consulting fees from Novartis and AbbVie (less than $10,000 each). Clegg reports consulting fees from Janssen (less than $10,000).