How to run a retrospective chart review.

Write every variable with its definition and its source before the review starts: the diagnosis code set, the lab and its LOINC code, the medication class, the date rule. Split the list in two. The coded variables (diagnoses, dates, results, medications, procedures, staging) come out of the EHR as columns. The judgment variables (a reading of a note, an ambiguous stage, a cause) are the medical chart review itself, and they get an abstraction rule and a second reader. The first list becomes data-dictionary.json; the second is the abstraction form.

Most chart reviews start from a list: MRNs from a pathology log, names and dates of birth from a clinic schedule, a CSV from a prior study. Roster pull matches every row against the EHR, an exact MRN first, then the FHIR Patient/$match operation on name, date of birth and sex, then a demographic search, and grades it certain, probable, possible, multiple or none. You confirm the matches, and the confirmed rows become the cohort the export runs on. A study that starts from criteria instead of a list builds the cohort the same way and reads the count first.

The export runs for the cohort and the resource types the variable list names, through the certified bulk FHIR API, in your environment. What comes back is study.csv, one row per patient and one column per coded variable, and data-dictionary.json. For every coded variable the reviewers work in the CSV, and each cell can be traced to the FHIR element it came from.

The judgment variables are the review: two readers, an abstraction rule per variable, an agreement check on a sample. With the coded fields filled, the readers open a chart only for the questions that take a reading of the notes. Abstraction starting from the coded record shows the split for each field and what it does to the minutes per case.