Developing Computable Phenotype for Chronic Diseases using PEDSnet
Download as pptx, pdf0 likes135 views
This document describes the development of a computable phenotype for Crohn's disease using data from the PEDSnet network. It discusses defining case and non-case definitions using various combinations of diagnoses, medications, procedures. It then evaluates the phenotype algorithms through internal validation of consistency across sites and predictive ability compared to procedures, and external validation against a gold standard of manually curated cases. The key lessons are on the importance of evaluating data quality issues across sites and balancing sensitivity vs specificity in selecting the final algorithm.
Developing Computable Phenotype for Chronic Diseases using PEDSnet
- 1. Designing Computable Phenotypes for Chronic Diseases using PEDSnet Ritu Khare, PhD The Childrens Hospital of Philadelphia November 16 2016
- 2. Background Cohort identification Certain conditions or features Key in pediatric research Traditional practice Chart abstractions Administrative data (billing code) Computable Phenotype EHR-based, pragmatic Multiple clinical elements e.g. eMERGE network True Cases Diagnoses Lab Interventions Encounters
- 3. Workflow for Computable Phenotype Development Case Non-case Definition Medications Diagnoses Labs Encounter Exposures Combination Configuration Algorithm Design Internal External Validation Framework Criteria Algorithm Selection Type 2 diabetes Crohns disease Glomerular disease Autism
- 4. Crohns Disease Computable Phenotype Exposures Diagnosis (SNOMED-CT) Methods = string match, concept relationships 49 Crohns disease, 34 ulcerative colitis, 208 other GI concepts Medications (RxNorm) Method Identify ingredient-level codes: balsalazide, mesalamine, sulfasalazine, ciprofloxacin, levofloxacin, (Ritchie et al. 2010) Determine granular codes: Clinical drug form, Clinical drug, Branded drug, Branded pack, 1086 RxNorm concepts for Crohns disease medications Others Encounter type Clinic or provider specialty Diagnoses type
- 5. Crohns Disease Computable Phenotype Algorithm design Version Exposures #Case Patients in PEDSnet Diagnosis* Meds Exclusion Diagnosis Case v1 >=1 Crohns disease Ulcerative colitis 10,669 Case v2 >=1 Crohns disease Crohns disease medications Ulcerative colitis 8,567 Case v3 >=3 Crohns disease Ulcerative colitis 7,570 Case v4 >=3 Crohns disease Crohns disease medications Ulcerative colitis 7,030 Noncase v1 >=1 Other GI disease Crohns disease, ulcerative colitis 1,169,861 Noncase v2 >= 1Other GI disease** Crohns disease, ulcerative colitis 290,081 *limited to office visits or problem lists **limited to specialty clinics
- 6. Crohns Disease Computable Phenotype Internal Validation: cross-site comparison Site Case v1 Case v2 Case v3 Case v4 Non- case v1 Non- case v2 s1 0.34% 0.21% 0.24% 0.19% 22.09% 8.28% s2 0.09% 0.06% 0.07% 0.06% 26.25% 4.93% s3 0.16% 0.12% 0.13% 0.11% 22.45% 4.73% s4 0.21% 0.17% 0.17% 0.15% 15.86% 4.60% s5 0.23% 0.16% 0.19% 0.16% 19.24% 7.13% s6 0.22% 0.16% 0.16% 0.13% 24.67% 0% s7 0.26% 0.19% 0.24% 0.18% 24.60% 6.55% s8 0.16% 0.12% 0.13% 0.11% 22.13% 7.86%
- 7. Crohns Disease Computable Phenotype Internal validation: Logistic Regression Cohort = Patients with GI disease (#1,060,127) Endoscopy GI Radiological exam #Crohns disease encounters (1,2,3,4,5+) 1.24 (1.21-1.27) 1.56 (1.54 1.58) Crohns disease medications (yes/no) 3.80 (3.65-3.97) 3.39 (3.27-3.51) ExposureVariables Outcomes (procedures) Odds ratio (95% confidence interval)
- 8. Crohns Disease Computable Phenotype Internal Validation: Longitudinal Reliability 60.00% 65.00% 70.00% 75.00% 80.00% 85.00% 90.00% 95.00% 100.00% Y1 -> Y2 % case patients with a follow- up visit 60.00% 65.00% 70.00% 75.00% 80.00% 85.00% 90.00% 95.00% 100.00% Y1 -> Y2 % case patients with a follow-up Crohns disease visit v1 v2 v3 v4
- 9. Crohns Disease Computable Phenotype External Validation: Gold Standard (CHOP) 1 Dx 3Dx + Meds Manually curated list Manually curated list Precision (PPV) = 0.66 Recall(sensitivity) = 0.99 Precision (PPV) = 0.81 Recall (sensitivity) = 0.98
- 10. Conclusions and Lessons Learned Evaluation patient-level chart reviews are expensive several possibilities of internal validation Data quality issues sites with no GI providers or care sites! sites with no patients on Crohns disease medications! Algorithm Selection known data quality issues internal validation findings sensitivity vs. specificity balance
Editor's Notes
- #3: --- what is cohort identification and why is it important ? - Traditional methods, and computable phenotypes. Diagnosis codes alone are not sufficient in developing optimal patient identification algorithms Using additional clinical data elements beyond diagnosis codes can often be better predictors of disease-specific outcomes Diagnosis codes Accurate Cohort identification is the key for conducting any research What are the challemges of cohort identification More accurate than methods relying on diagnosis billing code data alone Less prone to instances of mis-coding Using multiple clinical elements improves accuracy of case identification (stronger evidence) CP Using multiple clinical elements improves accuracy of case identification Allows for more complex patient identification Can use an algorithmic approach to patient identification with multiple inclusion and exclusion criteria Can aid in new patient identification Patients may meet clinical criteria for conditions but not yet identified by clinician or have no documented diagnosis code in the HER Automatic identification of patients with hypertension and obesity based on biometric data or by disease-specific medications Integration into EHRs to offer clinical decision support to alert providers to undiagnosed conditions Sources: diagnosis codes, procedure codes, laboratory test results and biometric data Phenotypes developed with domain experts who understand the clinical phenotype and how it is represented in EHR data Clinical informaticians identify and extract the information
- #4: Describe higher level workflow. Based on development of at least 3 different computable phenotypes in PEDSnet. For three different chronic conditions. : type 2 diabetes, crohns disease and glomerular disease. We will use Crohn;s disease as a running example for this talk.
- #8: 7 sites Endo.... Would expect a higher odds ratio Ever recieved one of those exams.... All exposures in one equation. New model --- one dummy for has CD medications. And -- one dummy --- 1,2,3,4 or more.
- #9: Average 89% in the first graph and 79% in the second graph Alogrithm version Most specific algorithm, For first graphs did they shift to UC diagnosis. ? At chart review level. Look at patients who we are losing from one graph to another. Take Chriss feedback.