By Nettie McFarland, RHIT, CCS-P, CHC
Director, Provider Education Support
Artificial intelligence (AI) is reshaping documentation in emergency medicine. Voice-to-text, ambient scribing, and predictive charting can reclaim clinician time and reduce clerical burden. But these same tools introduce new documentation risks: over- and under-documentation that can cloud clinical reasoning and misrepresent patient acuity. In parallel, the sheer volume of real-time data (labs, imaging, monitors) can push teams to anchor on isolated metrics rather than the evolving, multidimensional picture of a critically ill patient.
Documentation Pitfalls in the AI Era
Over‑documentation: When “More” Obscures the Truth
AI tools, optimized for completeness, often generate expansive, templated language. That can unintentionally inflate findings (e.g., by populating normal systems that were not examined) and bury the clinician’s actual medical decision-making (MDM). The result? Charts that look thorough but don’t truly reflect the encounter—and that may complicate coding, audits, and medicolegal review.
Under‑documentation: When Nuance Gets Lost
At the other extreme, AI may miss subtle but crucial elements—context, patient quotes, bedside reasoning, or rapid clinical changes—especially amid ED chaos. It can capture actions but omit the “why,” weakening the clinical narrative and the justification for resource-intensive care (e.g., critical care). Human review and targeted edits remain essential to preserve intent, nuance, and accuracy.
Bottom line: AI should augment—not replace—clinician judgment. The partnership works only when clinicians actively curate and contextualize autogenerated text.
Clinical Acuity Isn’t a Single Number: The Problem with Isolated Metrics
Emergency deterioration is rarely defined by one abnormal value. While single metrics (e.g., oxygen saturation, lactate, heart rate) can trigger alerts or AI prompts, true decline usually unfolds as a pattern across respiratory, cardiovascular, neurologic, and metabolic domains. Anchoring on a single metric risks over- or underestimating critical care needs.
Pattern Recognition Beats Point‑in‑Time Values
Key indicators that often out-predict any single number include:
- Increasing work of breathing
- Declining mental status
- Rising oxygen or fluid requirements
- Short-interval vital sign trends (minutes to hours)
These dynamic patterns, captured and reassessed at the bedside, usually forecast deterioration more reliably than an isolated outlier.
Why Context Matters: O₂ Saturation Examples
Low SpO₂ that may not support critical care:
Chronic COPD/emphysema at baseline 88–92%; high-altitude physiology; mild pneumonia with stable vitals—patients may be comfortable, conversant, and not in distress despite lower numbers.
Near‑normal SpO₂ that can support critical care:
Pulmonary embolism (with tachycardia, hypotension, chest pain); early sepsis (tachypnea, elevated lactate, altered mentation); carbon monoxide poisoning (SpO₂ 100% yet symptomatic). Here, pattern and physiology—not a single “normal” value—drive the decision.
Bringing It Together: Documenting Critical Care in an AI-Supported ED
To ensure documentation both reflects reality and supports appropriate billing and compliance, align AI use with the whole‑patient assessment:
1) Make AI your scribe, not your storyteller
- Actively edit autogenerated sections to remove unperformed exam elements and irrelevant boilerplate.
- Preserve MDM clarity: explicitly state differential, risk of deterioration, and the rationale behind high-resource interventions.
2) Promote patterns over points
- When an alert or AI suggestion triggers on a single metric, treat it as a prompt to reassess the whole patient, not an endpoint.
- Document the trajectory (e.g., work of breathing, mental status, fluid/oxygen needs, trend in vitals) and how those changes informed interventions.
3) Capture bedside reassessment loops
- Short interval checks often surface deterioration earlier than static values. Note time-stamped reassessments and responses to therapy (e.g., “10:12—RR 32 → 10:28 after fluids/antibiotics—RR 24; mentation improving”).
4) Contextualize “abnormal” and “normal”
- Clarify when a low value is baseline or physiologically expected (COPD, altitude).
- Conversely, underscore red flags despite normal-looking numbers (e.g., sepsis with SpO₂ 97% but acidosis and AMS). This protects against under-calling acuity in charts.
5) Keep the narrative yours
- Use AI to capture quotes and events but author the interpretation: why you suspected PE despite normal SpO₂, why you mobilized critical care resources, and how you judged response and risk after interventions.
Sample Documentation Language (Adapt and Tailor)
MDM—Risk & Rationale: Patient at high risk for rapid decompensation given escalating work of breathing, tachycardia, hypotension, and altered mentation despite near‑normal SpO₂. Pattern concerning distributive shock with impending respiratory failure. Initiated continuous monitoring, high‑flow O₂, large‑volume resuscitation, broad‑spectrum antibiotics; critical care time documented below.
Reassessments & Trends: 09:15—RR 32, HR 128, MAP 58, lactate elevated; WOB increased; oriented ×2. 09:35—after 30 mL/kg fluids: RR 28, MAP 64, mentation slightly improved; continues to require escalating O₂.
Interpretation: Partial response; ongoing risk of deterioration; ICU consult.
AI‑Generated Content Review: Autogenerated normal ROS elements removed; exam limited by acuity. All documented findings are verified and reflect the care delivered.
Quick Checklist for ED Teams
- Remove unperformed templated elements from AI notes.
- Explicitly document the pattern (WOB, mental status, requirements, trends
- Record time-stamped reassessments and treatment responses.
- Explain when “low” is baseline and when “normal” hides danger
- Ensure the MDM narrative is clinician-authored and ties interventions to risk.
Conclusion
Modern ED documentation lives at the intersection of AI efficiency and human clinical judgment. The best records don’t just recount vitals or echo templates; they demonstrate synthesis: why the team escalated care, what changed over minutes to hours, and how bedside assessment outweighed or reinterpreted isolated numbers. By curating AI output and foregrounding pattern recognition, clinicians can produce documentation that is accurate, defensible, and truly representative of emergency medicine at its best.
