AI Scribe for Interventional Cardiology

Specialty Documentation Architecture

Interventional cardiology documentation demands the simultaneous handling of structural echo metrics, functional symptom narratives, device interrogation data, and guideline-directed medical therapy (GDMT) titration histories within a single encounter. Unlike primary care, where a single problem may dominate, the interventional visit interleaves decompensated heart failure, valvular disease severity, and procedural planning for TAVR or transcatheter edge-to-edge repair. Merry AI is architected to keep these streams discrete, coded, and auditable rather than collapsing them into unstructured prose.

The central design principle here is that NYHA functional class is a clinician-assigned assessment, not an artifact that can be auto-derived from an ejection fraction alone. NYHA describes symptom-based limitation—Class I through IV—entirely independent of the underlying LVEF. Joint Commission abstraction guidance is explicit that NYHA class must be specifically documented in the record and never inferred by an abstractor from symptoms. Our automation therefore assists the clinician by presenting context, but requires an explicit selection before any coded value is written.

Binding echo metrics to functional class is consequently associative rather than deterministic. The system extracts the most recent LVEF with its date and modality (TTE, TEE, MRI, or nuclear), assigns the heart failure phenotype—HFrEF at LVEF ≤40%, HFmrEF at 41–49%, HFpEF at ≥50%—and presents that phenotype alongside the NYHA selection widget. The clinician sees the substrate and confirms the functional class; only then does the note engine assemble the structured narrative.

Anchoring NYHA and LVEF at the Point of Care

Every heart failure encounter should open a dedicated HF Status block that captures phenotype, LVEF with source study, NYHA class at this visit, ACC/AHA stage, recent hospitalizations, the current GDMT regimen with doses, and device status. Best-practice cardiology templates and ACC registry requirements—NCDR and PINNACLE—converge on exactly these discrete fields, which means capturing them cleanly serves both the clinical note and the downstream quality feed simultaneously.

Consider the canonical decompensation scenario: a patient with LVEF 28% presenting with dyspnea on mild exertion and fatigue with ordinary tasks but no symptoms at rest. The system surfaces "Latest echo (TTE 06/01/2026): LVEF 28% → HFrEF" adjacent to the NYHA selector and proposes a plausible range of Class II–III. It does not select for the clinician. Once the interventionalist chooses Class II, the engine drops a structured statement—"Heart failure with reduced ejection fraction, LVEF 28%, NYHA Class II at this visit"—and stores the coded observation mapped to LOINC 33878-0.

Preserving Interval Change and Trajectory

Longitudinal memory is where ambient documentation earns its place in a cardiology clinic. When NYHA class worsens from II to III between visits, that transition is clinically meaningful even in the absence of overt decompensation. Merry AI compares the confirmed class against the prior coded value and generates trajectory language: "NYHA functional class has worsened from II to III since the last visit, indicating progression of heart failure despite current therapy." This single sentence carries disproportionate weight in defending the progression element of high-complexity MDM.

Standardized Coding via LOINC 33878-0

Terminology discipline underpins the entire architecture. The NYHA field is modeled as a single coded observation with values 1 through 4 plus "not documented," mapped bidirectionally to LOINC 33878-0 and to local EHR codes. Peer-reviewed work on automatically extracting NYHA classification from clinical notes, indexed in the National Library of Medicine's PMC archive, confirms both the feasibility and the pitfalls of NLP-based extraction—reinforcing why our design keeps the human clinician as the source of truth while the machine handles structuring and coding.

Building the MDM Engine for 99215 and G2211

High-complexity medical decision making for an established patient supports 99215, selectable either by MDM or by 40–54 minutes of total time on the encounter date. The MDM path generally requires one or more chronic illnesses with severe exacerbation or progression, or one acute or chronic illness posing a threat to life or bodily function. Decompensated heart failure with reduced ejection fraction sits squarely in this territory, particularly when combined with device interrogation and possible hospitalization.

The engine operationalizes complexity across three domains in sequence. First, it identifies the chronic condition and its severity—flagging HF as progressive or life-threatening when LVEF ≤30%, NYHA III–IV, or a recent HF hospitalization is present. Second, it counts the categories of data reviewed: echo, laboratory trends, device interrogation reports, and independently interpreted external cardiology notes. Third, it assesses management risk, recognizing multi-agent GDMT titration across ARNI, beta-blocker, MRA, SGLT2i, and diuretics, or explicit hospitalization thresholds discussed with the patient.

When all three domains resolve to high, the system surfaces "MDM level: High — supports 99215," always as decision support and always overridable. Critically, it ensures the narrative text matches the structured data used to justify the level, preserving an audit trail that shows the exact logic path. Separately, because interventional cardiology so often serves as the continuing focal point for a patient's single serious HF condition, the engine flags G2211 add-on eligibility consistent with CMS guidance on the visit complexity add-on, capturing longitudinal complexity revenue that is otherwise routinely left unclaimed. You can pull the full MDM scaffold and HF status generators from Access Specialty Prompts at templates.scribing.io.

Compliance, Auditability, and Clinician Control

Regulatory posture is non-negotiable in this design. Automation is treated strictly as clinical decision support, never as autonomous coding—CPT and payer guidance both expect physician judgment on final E/M selection. The system therefore requires clinician confirmation of NYHA class, permits downgrading of any suggested MDM level, and maintains logs that pair the structured HF findings with the reasoning path that produced each recommendation.

Protected health information is handled under a zero-retention model. Encounter audio is processed transiently in RAM and shredded immediately after the draft is produced, with a Business Associate Agreement executed for every practice. Only the discrete, clinician-signed fields propagate to registries and analytics, which keeps quality reporting to NCDR and PINNACLE both accurate and defensible. To see the LVEF-to-NYHA binding and MDM engine applied to your own structural heart workflow, Schedule a 15-Minute Specialty Workflow Audit.

The Quiet Advantage in a High-Acuity Clinic

The cumulative effect of this architecture is a note that reads as though it were dictated by a meticulous department chair: phenotype and LVEF stated with source and date, NYHA class explicitly assigned and trended, GDMT regimen enumerated with doses, and a plan whose risk language aligns precisely with the billed complexity. Nothing is inferred that should be judged; nothing is judged that could be structured.

For the interventional cardiologist moving between the cath lab, the structural heart clinic, and the advanced HF service, that consistency is the difference between defensible longitudinal documentation and a scattered record. Merry AI removes the clerical friction while preserving the one thing that must remain human—the clinician's functional assessment of the patient in front of them.