Day 03
Phenotype eval is with the usual suspects: Sensitivity, Specificity, etc.
Building the Contingency Table is complicated: You don’t have a gold standard!
Sensitivity is hard. You can have a high PPV but low Sensitivity.
Incidences can vanish after 2015. Why? ICD9 → ICD10! Using a standard vocabulary addresses this.
Hardest phenotype to do is diagnosis of exclusion (e.g. TROLY and DILY.)
Don’t just build a cohort and assume it’s good!
It’s not just a single diagnosis code that you use to phenotype. The clinical story is far more complicated. Just understand that. E.g. did they actually have Myocardial Infraction or were they treated as if they had it? How do you know?
Coders will alter the admission diagnosis post fact. Don’t always think that the admission diagnosis will be the same as the discharge diagnosis.
Concept Set Building
For conditions: Common reference standard is SNOMED (ICD9 before 2015 and ICD10 after 2015 and “Read” in the UK map to this.)
MedDRA (Commercial, used in cllinical trials for outcome reporting) sits on top of SNOMED. It’s for “classification terms” (nothing breaks if you don’t use them.) Now SNOMED itself is an ontology that defines relationships. MedDRA provides additional relationships that are not covered (or are a superset) of SNOMED. Again, you may use these or not.
For Drugs: RxNorm, NDC, Local Terminologies, CPT4/HCPCS, ICD10PCS map to → RxNorm and RxNorm Extension.
For Procedures: OPS, CPT4/HCPCS, Local Terminologies map to → SNOMED.
For Labs: ICD10CM, Local Terminologies, Read, LOINC map to → LOINC and SNOMED.
Evidnet allows you to see what other data exists in participating OHDSI sites. Note that this is metadata: it’s aggregate summary counts and not the ‘full’ data.
Just feel like using ICD10 (or some other source vocabulary) and don’t care about the standard vocabularies? The “Mapped” column allows you to do this. It’s not recommended at all though.
Cohort Building
That “Continuous Observation” thing — this is important when, for example, you want to really ensure that this is the first time someone had a stroke. There is no guarantee that they didn’t have a stroke before entering the system! What if they moved from Kentucky → NYC?
You are building distinct cohorts for your Target and Outcome!