david odgers - bmi retreat 2015 poster
TRANSCRIPT
Introduc)on
Outcome Variables
Phenotypic Profiling for Rheumatoid Arthri7s Treatment Efficacy Using Electronic Health Records
Stanford University, Department of Biomedical Informa)cs David J. Odgers, Natalie Tellis, Heather Hall, Michel Dumon)er
Results
Rheumatoid arthri)s (RA) accounts for one-‐fiKh of the deaths due to arthri)s, the leading cause of disability in the United States. Finding effec)ve treatments for managing arthri)s symptoms are a major challenge, since the mechanisms of autoimmune disorders are not fully understood and disease presenta)on differs for each pa)ent. The American College of Rheumatology clinical guidelines for treatment consider the severity of the disease when deciding treatment, but do not include any predic)on of drug efficacy. Using Electronic Health Records and Biomedical Linked Open Data (LOD), we hypothesize that LASSO regression classifiers can be created from base clinical features consis)ng of 482 co-‐prescrip)ons and 3,210 comorbidi)es for the top three treatment regimes for RA pa)ents. Addi)onally, we hypothesize that extending the co-‐prescrip)ons and comorbidi)es into features directly rela)ng to basic biomedical knowledge using LOD will improve the overall LASSO classifier performance.
Cohort Selec)on
Conclusion
The LASSO classifiers with extended features showed rela)ve improvement in almost all metrics over the baseline of strictly clinical features. All the AUCs were between .71 and .81 with a 7% rela)ve increase in AUC over baseline for DMARDS features. Test error improved across all models from 3-‐12% over baseline. Sensi)vity and Specificity showed both improvements and decline between the models, even though the classifier performance described by the Mathews Correla)on Coefficient improved or stayed stable for all three treatments regimes when compared to baseline. These results warrant further inves)ga)on, tes)ng addi)onal therapeu)c areas with addi)onal LOD resources.
Extended Features