Cheat sheet · Sheet 03
Characterization & Treatment Pathways
Quick reference for the ATLAS pathways module: what to build, in what order, and what the output actually means.
01Core mental models
Concept sets define WHAT. Cohorts define WHO and WHEN. A concept set is a code list with no dates. A pathway is a sequence of dated intervals, so pathways analysis takes cohorts as input, target and event alike.
| Purpose | Logic that belongs here | |
|---|---|---|
| Target cohort | Who is studied | Diagnosis logic, confirmation, exclusions, prior observation |
| Event cohorts | What happens to them over time | Exposure start, persistence, exit |
Diagnosis logic in an event cohort, or exposure logic in the target, makes results very hard to explain later.
02Target cohort
Typical entry
- Condition occurrence
- Initial events limited to earliest per person
- 365–730 days prior observation required
Prior observation is continuous observable data before index, enforced via OBSERVATION_PERIOD. It is what lets you claim the entry event is plausibly incident. Without it, someone's tenth year of therapy reads as their first line.
Answers: were we watching this person long enough to know this is new?
Confirmation (recommended) is one inclusion rule containing OR logic: second diagnosis within ±365d OR relevant drug exposure 0–180d after index. (Rules combine with AND; keep both branches inside one rule.)
Common exclusions are competing or mimicking diagnoses, and often an age floor. Name which are clinical judgment rather than method.
03Event cohorts
Each event needs a start date, an end date, and a persistence rule. A concept set supplies none of those, hence cohorts.
Drug event cohort template
| Element | Setting |
|---|---|
| Entry | Drug exposure of the ingredient |
| Initial event limit | Earliest per person |
| Prior observation | 0d before / 0d after, since the target cohort already establishes observability |
| Inclusion criteria | None, intentionally |
| Exit | End of continuous exposure, 30–60d persistence window, using days supply and exposure end date |
Principle: an event cohort is a pure exposure definition. No diagnosis logic.
04Persistence windows
Controls how refills group into episodes, meaning how large a gap has to be before ATLAS calls it a new exposure rather than a continuation.
- Does not change who received a drug.
- Does change path length, path complexity, and distinct event cohorts per person.
| Window | Effect |
|---|---|
| 0 days | Highly fragmented pathways |
| 30 days | Common default |
| 60–90 days | More consolidated episodes |
| 180 days | Conservative; fewer, longer steps |
Persistence is not adherence and not true duration of therapy. It is a data assumption about gaps, and that belongs next to the number every time it is shown.
Best demo in the module: regenerate with a different persistence window, put the two outputs side by side.
05Session flow
- Build the target cohort
- Dwell on prior observation, the conceptual center of the session
- Build one drug event cohort together
- Add two or three more
- Generate pathways
- Read the sunburst
- Move to the tables
- Change the persistence window, regenerate, compare
06Reading the sunburst
| Element | Is |
|---|---|
| Center | Target cohort entry (index) |
| Inner ring | First observed treatment event |
| Outer rings | Subsequent treatment events |
| Color | Event cohort |
| Arc size | Number of people |
- A pathway advances only when a new event cohort is entered.
- Remain = nothing new recorded after that step. Not "stopped treatment", since it may be continued therapy, lost observability, or the end of the study window.
- Diff = went on to another event cohort.
- Settings that change what you see: maximum path length (truncation), minimum cell count (suppression), event collection window relative to index.
07Tabular output
1a–1d below are this course's shorthand. Newer ATLAS versions may name them differently, so teach the questions rather than the labels.
| Table | Reads as | Watch |
|---|---|---|
| All pathways | One row = one exact sequence, to max path length | "+" = same step, not an ordering. % with pathway compares rows; % of cohort gives prevalence |
| Counts by rank | What tends to come first, second, third | Rank is position, not preference |
| Counts overall | "Ever exposed" across all ranks | Rows not mutually exclusive |
| Distinct cohorts per person | 1 = single observable episode; 2+ = observed treatment change | Highly sensitive to persistence window and prior observation, which is why it teaches well |
08Key takeaways
- Concept sets define what; cohorts define when.
- Prior observation protects us from calling old disease new.
- Pathways move forward only when something new happens.
- Persistence assumptions can change the picture more than behavior does.
- The sunburst shows a pattern; the tables tell you whether to believe it.
- Pathways are about time, sequence, and "firstness" in data collected for care, not research.
09References
A target cohort to practice on: Menopause in the OMOP vocabularies builds one, then notes what a hormone-therapy pathways analysis would use for its event cohorts.
- The Book of OHDSI, Ch. 10 and 11
- ATLAS documentation
- Hripcsak et al., Characterizing treatment pathways at scale using the OHDSI network, PNAS 2016
- Worked example 01, menopause in the OMOP vocabularies (dated 1 Aug 2026)
- Worked example 02, preterm labor in the OMOP vocabularies (dated 4 Aug 2026)
Corrections or suggestions welcome: danielle@boycedatascience.com