Most cardiovascular outcome trials define MACE as a 3-point composite: cardiovascular death, non-fatal myocardial infarction, and non-fatal stroke. Some protocols extend this to a 4-point or 5-point composite by adding hospitalization for heart failure or unstable angina. The protocol specifies what counts, down to the troponin threshold and imaging criteria for MI confirmation. The EHR does not write to that specification, and closing that gap at scale is where most RWE cardiology teams lose significant analyst time.
What the EHR gives you versus what the protocol needs
Discharge summaries, cardiology consultations, emergency department notes, and procedure reports all contain MACE-relevant information. A discharge summary might say: "Pt admitted with anterior STEMI, underwent primary PCI, discharged on DAPT." That maps cleanly to ICD-10 I21.09 and CPT 92928. The MACE definition is satisfied.
The same event documented differently creates extraction problems. A progress note on day two of admission might read: "troponin trending down, MI confirmed, transferred to cath lab." A cardiology referral written six months later might describe "post-STEMI patient managed at regional cardiac center." In a medication reconciliation note, the same history appears as "post-MI on aspirin/statin" with no event date. Three distinct formulations of the same event, none of them a structured field, all of them requiring extraction.
Why ICD codes are a starting point, not an answer
Most RWE teams begin MACE ascertainment with ICD billing codes. It is a reasonable first step: I21 covers STEMI and NSTEMI, I63 covers ischemic stroke, I46 covers cardiac arrest. But billing codes carry well-documented accuracy limitations for clinical research. Upcoding, rule-out admissions coded as confirmed diagnoses, and coding lag all affect precision. The accepted solution in outcomes research is note-level confirmation, which is exactly the labor-intensive step that automated extraction is designed to reduce.
A team that published a cardiovascular outcomes study with ICD-only endpoint ascertainment will now typically find that regulatory reviewers and peer reviewers ask for confirmation of note-level documentation for primary endpoints. That expectation has been established in the field, and it reflects real accuracy concerns rather than procedural formality.
Three extraction problems that need to be solved simultaneously
Reliable MACE extraction from clinical notes requires handling at least three distinct technical problems at once.
Event confirmation. Is the mention of MI, stroke, or death referring to a confirmed current event, or to a suspected, ruled-out, or historical event? "No evidence of MI" and "NSTEMI confirmed" are syntactically similar but semantically opposite for endpoint counting. Note type carries information: a discharge summary diagnosis list represents a final confirmed diagnosis. A day-one progress note may document a working hypothesis that the discharge summary resolves differently. An extraction system that treats both with equal confidence will inflate false positives.
Temporal attribution. When did the event occur? An MI documented in a note dated November 12 is not necessarily an MI that occurred on November 12. The note may be describing an event from a prior hospitalization. For endpoint analysis, the event date relative to the study index date determines whether it counts. Extracting the event without a reliable event date produces a record that is unusable for time-to-event calculations.
Negation scope. Clinical negation is local, not global. "No fever, chills, or shortness of breath" applies negation across all three entities. "MI ruled out, but new atrial fibrillation identified" negates the MI and asserts a separate finding. Correct negation handling requires identifying which entity falls within the scope of which negating expression, sometimes across complex sentence structures.
A concrete workflow for a 1,400-patient cohort
An evidence team at a cardiovascular outcomes research group had 1,400 patient records for a retrospective cohort study examining a 3-point MACE composite over a 24-month follow-up window. Their raw EHR extract contained discharge summaries, progress notes, cardiology consultations, and procedure reports in unstructured text. Structured billing data had I21 and I63 codes for a subset of patients, but the team had confirmed from a prior study that ICD-only ascertainment missed roughly 15 percent of endpoint events compared to adjudicated chart review.
We configured CliniNote to extract three entity types across the note corpus: confirmed acute MI events (STEMI and NSTEMI, with documentation of troponin criteria or imaging), confirmed ischemic stroke events with laterality where documented, and death events with cause where available. For each extracted entity, the system returned the source sentence, the note type, the note date, an estimated event date where derivable from the text, and a confidence score.
The workflow they adopted had three stages. First, automated extraction flagged all candidate MACE events with source citations. Second, an analyst reviewed only the flagged candidates rather than the full chart. Third, cases where extraction confidence was below a defined threshold went to full manual review. Analyst review time for flagged candidates averaged approximately 8 minutes per case, compared to 25 to 35 minutes for a cold chart read. Total analyst time for endpoint ascertainment on that cohort dropped by approximately 55 percent.
What this approach does not replace
We want to be direct about the limits. Automated MACE extraction functions as pre-screening, not adjudication. Trials with formal endpoint adjudication committees have specific requirements for the evidence package submitted to each committee, and those requirements often specify blinded clinician review of source documents. Extraction narrows the scope of that review and generates citation evidence for the adjudicators, but it does not eliminate the adjudicator role.
There is also a category of outpatient events that automated extraction handles poorly. An MI or stroke that occurred at a different facility, documented only as "prior MI per outside records" without a specific date or supporting hospitalization note, gives the extraction system very little to work with for confident endpoint classification. Those cases still require manual trace-back to source facility records regardless of what automated pre-screening found.
Reproducibility as the core argument
One reason to invest in structured extraction for MACE ascertainment has nothing to do with speed. It is reproducibility. Manual chart review produces variable results across reviewers and across time. Inter-reviewer disagreement rates of 12 to 20 percent for unstructured endpoint definitions are documented in methodology literature and are not outliers. They reflect genuine ambiguity in clinical text evaluated against strict endpoint criteria.
Structured extraction creates a version-controlled, documented process: model version, configuration, confidence thresholds, extraction timestamp, and source citations for every event are logged and exportable. When a regulatory reviewer or a peer reviewer asks how endpoints were ascertained, the answer is a replicable method, not a description of what reviewers were trained to look for. For teams building a portfolio of related RWE studies over multiple years, that audit trail has compounding value as regulatory and publication standards tighten.
