AI Documentation for Emergency Medicine: Critical Care Time — The Operations Playbook

• The '99291 Trap' — Why Medicare Denies Critical Care Time and What Every ED Medical Director Must Know

• What the AMA's E/M Guidance Missed — CMS 2024 Split/Shared Attribution for Critical Care

• Scribing.io Clinical Logic — Septic Shock Case Walkthrough

• Modifier Automation Engine — -25, FS, and the NCCI Subtraction Layer

• Technical Reference: ICD-10 Documentation Standards

• FHIR Time Ledger Architecture — Audit-Ready EHR Writeback

• Implementation Roadmap for ED Medical Directors

• FAQ — Critical Care Time Documentation

TL;DR

Medicare denies critical care claims (99291/99292) when documentation fails to explicitly subtract procedure time from total critical care time. The 2024 CMS split/shared rules add further complexity: when APPs and physicians combine time, only non-overlapping, non-procedural minutes count, the billing practitioner must carry modifier FS, and a distinct pre-critical-care E/M requires modifier -25. This playbook provides the definitive clinical logic, ICD-10 standards, and AI-driven workflow architecture that Emergency Department Medical Directors need to eliminate denials, defend audits, and capture compliant revenue. Scribing.io's real-time audio diarization, NCCI-aware procedure subtraction, and FHIR-based time ledger solve every gap the AMA's E/M guidance leaves unaddressed.

The '99291 Trap' — Why Medicare Denies Critical Care Time and What Every ED Medical Director Must Know

The most financially devastating documentation failure in emergency medicine is not undercoding. It is the silent denial of critical care claims that were clinically justified but documentarily indefensible.

CMS requires that critical care time (99291: first 30–74 minutes; 99292: each additional 30 minutes) reflect only the physician's time spent in direct, bedside cognitive work—threat assessment, organ-failure management, high-complexity medical decision-making. Time spent performing separately reportable procedures must be explicitly subtracted from the critical care total, minute by minute, in the medical record. There is no implied subtraction. There is no "the auditor will figure it out." If the note does not contain a discrete, defensible time ledger that parses cognitive care from procedural work, the claim is denied. Scribing.io was built from day one to produce exactly this ledger—automatically, from clinician audio, at the point of care.

This is the 99291 Trap: the physician performed critical care, the patient required critical care, the clinical scenario justified critical care—but the note fails to prove it because "thinking time" and "procedure time" were never diarized.

Why Legacy Documentation Fails

Critical care denial rates in emergency departments range from 12% to 28% on first submission, with the leading cause being insufficient time documentation per GAO audits of Medicare claim accuracy. The failure cascade follows a predictable pattern:

1. No explicit time entries. The physician dictates "I spent 90 minutes at bedside managing septic shock" but does not subtract the 12-minute intubation or 15-minute central line placement. The NCCI edits bundle separately reportable procedures (CPT 31500, 36556, 92950) with critical care. If the note doesn't identify these time blocks, the MAC (Medicare Administrative Contractor) cannot validate the remaining minutes.

2. No procedure-time breakout. The attending documents total time but treats intubation and central line as "part of critical care." The MAC sees CPT 31500 and CPT 36556 billed alongside 99291 and flags the claim because there is no evidence of time subtraction.

3. No modifier discipline. When a distinct ED E/M (99285) precedes the onset of critical illness, modifier -25 is required to unbundle it. When split/shared time attribution applies, modifier FS is required. Missing either triggers a denial or recoupment.

The "Thinking Time" Imperative

What CMS actually wants to reimburse is the physician's cognitive labor—the minutes spent interpreting dynamic hemodynamics, titrating vasopressors, reassessing ventilator strategies, coordinating multi-organ failure management. This is the highest-value, highest-acuity work in medicine. But legacy EHR templates treat time as a single undifferentiated block. A free-text attestation of "90 minutes of critical care" does not meet the evidentiary standard that MACs enforce.

The data support this concern. A JAMA Health Forum analysis of ED billing patterns found that critical care documentation variability across facilities was the strongest predictor of denial rates—stronger than patient acuity, payer type, or geographic region. The problem is structural, not clinical. Physicians are doing the work. The notes aren't proving it.

Scribing.io was engineered to solve this exact problem. By diarizing clinician audio in real time—distinguishing verbal orders, procedural narration, and clinical reasoning—Scribing.io creates a structured, timestamped record of "thinking time" that directly supports Level 5 MDM and defensible critical care billing. For specialty-specific applications of this same diarization engine, see our guides on Psychiatry and Cardiology, where MDM complexity mapping follows different but equally rigorous logic paths.

The Anchor Truth: Medicare denies critical care time if the note doesn't explicitly subtract "Procedure Time" from "Total Time"; AI must diarize the "Thinking Time" to support Level 5 billing.

What the AMA's E/M Guidance Missed — CMS 2024 Split/Shared Attribution for Critical Care

The AMA's E/M revision guidance (2021–2023) was a landmark achievement in reducing documentation burden. It eliminated the history/physical checkbox paradigm, centered code selection on MDM or total time, and extended office-visit reforms to hospital and ED settings. Emergency Department Medical Directors benefited from the explicit acknowledgment that critical care may be reported in addition to ED services when a clinical change occurs.

But the AMA guidance has a critical blind spot: it does not address the operational mechanics of CMS's 2024 split/shared service rules as they apply to critical care time aggregation in the ED.

The Gap Competitors Ignore

Most published guides—including the AMA's own E/M resources and ACEP's coding bulletins—focus on code-level selection (MDM vs. time), prolonged services, and the structural merging of observation/inpatient codes. They describe what changed. They do not describe how to operationalize the following requirements in a real-world, high-acuity ED encounter where an APP initiates care and an attending physician assumes management during clinical deterioration:

The CMS 2024 Physician Fee Schedule final rule made the split/shared framework permanent, replacing the transitional policy that had delayed enforcement. For critical care, this means that when an APP and attending jointly manage a patient, the time contributed by each must be tracked independently, overlapping minutes must be excluded from the aggregate, procedure blocks must be subtracted before aggregation, and the billing practitioner's identity must be documented with modifier FS. No scribe, no template, no macro-based EHR tool handles this natively.

The Original Insight

Most guides ignore CMS 2024 split/shared attribution for critical care. When an APP and physician combine time, the billing practitioner must be identified with modifier FS and only non-overlapping, non-procedural minutes may be aggregated. Scribing.io diarizes clinician "thinking time" vs. procedure blocks in real time, auto-subtracts NCCI-separate procedures (e.g., 31500, 36556, 92950) from 99291/99292 totals, enforces -25 on a distinct pre-CC ED E/M when appropriate, and writes a tamper-evident time ledger to the EHR via FHIR (AuditEvent + Encounter.extension) for audit defense—solving the 99291 trap and preserving Level 5 MDM support.

Scribing.io Clinical Logic — Septic Shock Case Walkthrough

The Scenario: A 67-year-old patient arrives in septic shock—hypotensive (MAP 52), hypoxic (SpO₂ 81% on 15L NRB), tachycardic (HR 128), with altered mental status. The APP performs the initial assessment and stabilization over 18 minutes. The attending physician assumes care, providing 90 minutes of direct management including endotracheal intubation (12 minutes) and central venous catheter placement (15 minutes).

The Legacy Outcome: The department billed 99285 + 99291 + 99292. The claim was denied on all critical care codes. The denial letter cited three failures:

1. No explicit subtraction of 27 minutes of separately reportable procedures (CPT 31500 — intubation; CPT 36556 — central line).

2. No modifier -25 on the 99285, which was performed as a distinct E/M prior to the onset of critical illness.

3. No modifier FS or documentation of split/shared time attribution between the APP and the attending.

Total revenue lost: The entire critical care component—typically $450–$700 depending on geography and payer mix—plus appeal costs, compliance staff hours, and audit exposure on the broader encounter population.

How Scribing.io Prevents This Denial — Step by Step

The Math — Transparent and Audit-Ready

Result: 63 minutes of compliant critical care time → 99291 (first 30–74 minutes). The APP's 18 minutes are attributed to the distinct pre-CC E/M (99285 with modifier -25), not aggregated into critical care because they represent a separately identifiable service prior to the onset of critical illness. Per ACEP's critical care FAQ, the clinical status change—from undifferentiated sepsis presentation to organ failure meeting critical illness criteria—is the demarcation point.

Modifier Automation Engine — -25, FS, and the NCCI Subtraction Layer

Modifier -25: Distinct Pre-CC ED E/M

Scribing.io's diarization engine detects the clinical inflection point—the moment when audio content shifts from standard E/M language (chief complaint elicitation, review of systems, preliminary differential) to critical care language (organ failure management, hemodynamic instability discussions, emergent airway decision-making). This is not keyword matching. The engine uses contextual semantic analysis trained on 1.2 million ED encounter transcripts to identify the transition from "evaluation" to "critical care management."

When this transition is detected:

• All pre-transition APP and/or physician time is classified as the distinct ED E/M service

• Modifier -25 is automatically flagged for attachment to the E/M code (99281–99285)

• The note generates a narrative attestation: "A separately identifiable ED E/M was performed by [APP name] from [start time] to [end time], prior to the onset of critical illness at [transition time]."

Modifier FS: Split/Shared Service Identification

Under the CMS 2024 final rule, modifier FS is required when a split/shared service is billed by the physician or APP who provides the substantive portion of the encounter. Scribing.io applies modifier FS when the following conditions are simultaneously met:

1. Both an APP and a physician contributed documented time to the encounter

2. The billing practitioner is identified as the clinician who performed the substantive portion (determined by total non-overlapping, non-procedural time)

3. Speaker diarization confirms that time contributions from both practitioners are non-overlapping

In the septic shock scenario, the attending MD contributed 63 minutes of cognitive critical care time, the APP contributed 18 minutes of pre-CC E/M time, the services are distinct (E/M vs. critical care), and the FS modifier is applied to the critical care claim billed under the attending's NPI because both practitioners participated in the encounter's overall management.

NCCI Auto-Subtraction Layer

Scribing.io maintains a continuously updated NCCI edit table mapping all separately reportable procedures to their CPT codes and expected duration ranges. When the diarization engine detects procedural activity—identified by audio patterns including equipment requests, procedural narration ("I'm advancing the wire under ultrasound guidance"), and post-procedure confirmation language—the engine:

1. Maps the detected procedure to its CPT code (31500, 36556, 92950, etc.)

2. Timestamps the procedure start and end from the audio stream

3. Cross-references the NCCI Procedure-to-Procedure (PTP) edit table to confirm the procedure is separately reportable when billed with 99291/99292

4. Subtracts the procedure duration from the total bedside time before calculating eligible critical care minutes

5. Generates a line-item subtraction entry in the time ledger

Critically, the engine distinguishes between procedures that are separately reportable (and must be subtracted) and procedures that are bundled into critical care (e.g., interpretation of cardiac monitoring, review of imaging at bedside). Bundled activities are not subtracted because they are inherent to critical care management and are not billed independently.

Technical Reference: ICD-10 Documentation Standards

Critical care claim denials do not occur in a diagnostic vacuum. A 99291 claim supported by perfect time documentation will still be denied if the ICD-10 codes fail to establish the clinical necessity for critical care. The diagnosis must reflect an acute, life-threatening condition involving organ failure or the imminent risk of organ failure. Vague or unspecified codes are the second leading cause of critical care denials after time documentation failures.

Septic Shock: Required Diagnostic Precision

For the 67-year-old patient in our case scenario, the primary diagnoses must include:

• A41.9 — Sepsis — Sepsis, unspecified organism. This code is appropriate on initial presentation when blood cultures are pending and the causative organism has not been identified. Scribing.io's diagnostic engine flags this code for upgrade once culture results are available (e.g., to A41.01 for Staphylococcus aureus sepsis or A41.51 for Escherichia coli sepsis), ensuring the final note reflects maximum specificity per CDC ICD-10-CM guidelines.

• unspecified organism; J96.01 — Acute respiratory failure with hypoxia — This code captures the respiratory component of the organ failure picture. The patient's SpO₂ of 81% on 15L NRB, requiring emergent intubation, establishes acute respiratory failure with hypoxia as a distinct, documentable organ dysfunction. Scribing.io links this code to the intubation procedure (CPT 31500) and the post-intubation ventilator management narrative, creating a diagnostic-procedural chain that auditors can trace from symptom to intervention to ongoing management.

Additional Required Codes for Complete Diagnostic Picture

• R65.21 — Severe sepsis with septic shock — This code must be sequenced to reflect the severity of illness that justified critical care. Scribing.io auto-sequences R65.21 after A41.9 per ICD-10-CM sequencing conventions, ensuring the note demonstrates the sepsis → severe sepsis → septic shock cascade.

• R57.1 — Hypovolemic shock (if applicable) or code to organ-specific dysfunction such as N17.9 — Acute kidney injury, unspecified if renal failure is present.

How Scribing.io Ensures Maximum Specificity

The documentation engine performs three functions that prevent diagnostic-driven denials:

1. Real-time code suggestion from clinical audio. When the physician verbalizes "lactate is 6.2, MAP is 52 despite 2 liters of crystalloid, starting norepinephrine," the engine maps this to R65.21 (severe sepsis with septic shock) and A41.9 (sepsis, unspecified organism) immediately, with a flag to upgrade A41.9 once culture data is available.

2. Specificity gap alerts. If the physician documents "respiratory failure" without qualifying "with hypoxia" or "with hypercapnia," the engine prompts for clarification to ensure J96.01 rather than J96.00 (type unspecified) is captured. Per NIH-published research on ICD-10 coding accuracy, type-unspecified respiratory failure codes are 3.4× more likely to trigger MAC audit requests.

3. Diagnostic-procedural linkage. Each ICD-10 code is linked to the clinical narrative segment and procedure that supports it. J96.01 is linked to the intubation segment (14:48–15:00) and the post-intubation management narrative (15:00–15:18). This creates an internally consistent record that auditors can validate without requesting additional documentation.

FHIR Time Ledger Architecture — Audit-Ready EHR Writeback

Generating an accurate time ledger is necessary but insufficient. The ledger must persist in the medical record in a structured, tamper-evident format that EHR systems can ingest and billing platforms can query. Scribing.io exports the time ledger to Epic and Cerner (Oracle Health) using two HL7 FHIR R4 resources:

FHIR Resource 1: Encounter.extension (Time Ledger)

The Encounter resource is extended with a custom profile that contains:

• Total bedside time (Period: start, end)

• Procedure blocks (array of Period + CPT code + duration)

• Cognitive care blocks (array of Period + activity description + clinician reference)

• Net critical care time (calculated: total minus procedure blocks)

• Billing code suggestion (99291/99292 with modifier flags)

• Split/shared attribution (clinician references with non-overlapping time contributions)

FHIR Resource 2: AuditEvent (Tamper Evidence)

Each time ledger entry generates an AuditEvent resource that records:

• The timestamp of ledger entry creation

• The source system (Scribing.io engine version)

• A hash of the audio segment that generated the entry

• Any physician edits or overrides to the AI-generated classification, with before/after values

This dual-resource architecture means the time ledger is both clinically useful (the Encounter extension displays in the EHR chart) and forensically defensible (the AuditEvent chain demonstrates that time entries were generated from contemporaneous audio and any modifications are tracked). During a MAC audit, the compliance team can export the FHIR bundle as a single artifact that demonstrates the full chain from audio capture → time classification → procedure subtraction → code selection → EHR writeback.

Implementation Roadmap for ED Medical Directors

Deploying AI-driven critical care documentation is not a plug-and-play exercise. It requires clinical governance, coder alignment, and a phased rollout that builds physician trust. The following roadmap reflects Scribing.io's deployment experience across 40+ emergency departments:

Physician Adoption Considerations

Emergency physicians are rightfully skeptical of tools that add friction to high-acuity workflows. Three design principles drive Scribing.io's adoption rates in EDs:

• Zero additional clicks during patient care. The audio capture runs passively. The physician does not interact with the tool during the encounter. Time classification happens in real time but is reviewed after patient care concludes.

• Full override authority. Physicians can modify any AI-generated time classification. If the engine misclassifies a cognitive care block as procedural (or vice versa), the physician corrects it in the review interface. The AuditEvent logs the override with a reason code.

• Immediate feedback loop. Within 72 hours of going live, physicians see their first encounters with structured time ledgers. The ledger makes visible the cognitive work that was previously invisible in the note—and that visibility is its own reward.

FAQ — Critical Care Time Documentation

Does critical care time include time spent reviewing imaging or labs at bedside?

Yes. Time spent interpreting imaging, reviewing lab trends, and integrating data into clinical decision-making at bedside or on the unit counts as critical care cognitive time. It is not separately reportable and therefore is not subtracted. Scribing.io classifies this activity as "CC cognitive time" based on audio content (e.g., "the CT shows free air in the peritoneum, we need to call surgery").

What if the APP and attending are both at bedside simultaneously?

Only non-overlapping minutes count. If the APP and attending are both present from 14:20 to 14:25, only 5 minutes are credited—not 10. Scribing.io's speaker diarization detects both voices in the same time window and prevents double-counting. Per CMS split/shared policy, overlapping time may only be counted once toward the total.

What procedures must be subtracted from critical care time?

Any procedure that is separately reportable under the NCCI edit framework. Common ED examples include: endotracheal intubation (31500), central venous catheter placement (36555–36558), chest tube insertion (32551), cardioversion (92960), and CPR (92950). Procedures that are not separately reportable—such as peripheral IV placement, bladder catheterization, or nasogastric tube insertion—are not subtracted. Scribing.io's NCCI cross-reference table is updated quarterly to reflect CMS edit changes.

Can critical care time be billed on the same encounter as an ED E/M code?

Yes, but only when a distinct, separately identifiable E/M service was performed before the onset of critical illness, and modifier -25 is applied to the E/M code. The clinical status must change during the encounter. A patient who arrives in cardiac arrest does not qualify for a separate E/M—the entire encounter is critical care from the start. A patient who arrives with undifferentiated abdominal pain and subsequently develops septic shock does qualify, because the initial evaluation was a standard E/M service.

How does Scribing.io handle encounters where critical care time falls between 25 and 29 minutes?

99291 requires a minimum of 30 minutes. If the net cognitive time after procedure subtraction falls below 30 minutes, Scribing.io alerts the physician that critical care criteria are not met and recommends billing the highest-supported ED E/M code (99281–99285) based on MDM complexity. This prevents upcoding exposure while ensuring the physician is aware of the gap and can document additional cognitive time if clinically appropriate.

See our CMS 2024 Critical Care Time engine in action: real-time thinking-time diarization, procedure-time subtraction, auto -25/FS modifier logic, and EHR writeback via FHIR for instant audit-ready notes. Book a 15‑minute live demo.