Skip to content

Latest commit

 

History

History
38 lines (20 loc) · 4.26 KB

stratified_statistics_protocol_normal.md

File metadata and controls

38 lines (20 loc) · 4.26 KB

Stratified statistics of biomarker labs: protocol and workflow

Generating statistics for stratified, normal-physiological condition patients

Clinical lab tests for biomarker measurement can be enhanced and targeted for precision diagnostic and prognostic applications by EHR-data driven stratification. Whereas there is general understanding that values indicating normal-physiological, and abnormal-pathological, depend on patient strata, such as by sex and age, there is an unmet need for data-driven guidance on these statistics. This is the need we attempt to begin to address in this approach, initially focusing on molecular biomarkers, mostly plasma proteins, and normal-physiological (healthy) distribution statistics.

For each molecular biomarker, the following steps comprise our algorithm.

  1. SELECT YEAR: Select a year which will be considered representative. There are several reasons for not using the entire HealthFacts database. One is practical, since a subset allows for faster analysis and methods development. Another reason relates to data quality, and uncertainties regarding sampling bias and other issues with some older data. We are using the year 2018, the last full year that HealthFacts was maintained as such, before the acquisition of Cerner by Oracle.

  2. SELECT LABS: Determine the lab procedure IDs and LOINC codes for lab tests for the biomarker of interest. There are generally multiple tests for important biomarkers such as Troponin and PSA.

  3. QUERY DB FOR LABS: Query the db for all test results for the given LOINC codes. We expect ample data quantity, so focus on quality, and filter any rows with NULL (or non-numeric) values, or NULL or uninterpretable units. This query also retrieves encounter IDs, which can be linked to patient IDs.

  4. SELECT UNITS: If a predominant or most-frequent unit (e.g. ng/mL) allows for sufficient data quantity, filter all others.

  5. FILTER EXTREME VALUES: Mindful that EHR data includes noise and errors, and given that we are focusing on normal-physiological statistics, filter values exceeding some percentile from the median. Possibly keeping only 10-90pct-ile. Generate basic statistics for this full dataset: N, mean, median, min, max, stddev.

  6. QUERY DB FOR PATIENTS: Query for all patient IDs from encounter IDs in the dataset. This is the initial patient cohort. It may be advantageous to build a temporary db table with this mapping, and the patient stratification variables of interest.

  7. EXCLUSION CRITERIA: Develop set of diagnostic ICD codes to be exclusion criteria, defining a normal patient criteria.

  8. QUERY DB FOR EXCLUSION DX: Retrieve all exclusion criteria diagnoses for the initial patient cohort. Filter the initial patient cohort using the exclusion criteria, to generate the normal patient cohort. Re-generate descriptive statistics for the stratification variables of interest.

  9. GENERATE STATISTICS: Generate descriptive statistics on the patient cohort for the stratification variables of primary and initial interest: age, sex, race.

    • Note that while DOB is a persistent patient variable, age depends on the date of the encounter and test. This is another reason to use a single year of data, since to some precision this allows for a single age per patient (e.g. age at mid-year).

    • Similarly, patient-type (Emergency, Inpatient, Outpatient, etc.) depends on the encounter and can vary for a given patient. We may wish to filter by patient-type to favor normal-physiological lab results.

  10. REMOVE EXTREME VALUES: Remove extreme values from dataset, based on percentiles. Currently we remove values below 10th percentile and above 90th percentile. Removing spurious values is one justification. Also, given that exclusion criteria are unlikely to exclude all non-healthy patients, this can further ensure normal-physiological values.

  11. AGE RANGES: Define stratification age ranges. Probably age ranges by decade [0-9], [10-19], [20-29], ... [80-89], 90+. Generate patient ID sets for each sub-cohort.

  12. For each patient sub-cohort, each strata, generate statistics: N, mean, median, min, max, stddev.

This workflow should result in data files (and tables) which allow for revision of: exclusion criteria, stratification ranges, and value filtering.