Critical Care and Pulmonary Medicine

Critical care generates dense streams of ventilator, hemodynamic, laboratory, medication, and monitoring data. That density creates opportunities for pattern recognition while amplifying alert burden, label ambiguity, and the consequences of delayed or inappropriate action. Published alarm reviews describe hundreds of alerts per patient-day in some ICU environments (Sendelbach and Funk, 2013). Before adding an AI alert, measure the local alert volume and define who must respond, how quickly, and with what action.

Learning Objectives

After reading this chapter, you will be able to:

  • Evaluate AI-powered early warning systems for patient deterioration
  • Understand sepsis prediction algorithms, their benefits and significant limitations
  • Assess AI tools for mechanical ventilation optimization and weaning
  • Navigate AI applications in pulmonary imaging, including chest CT analysis
  • Evaluate AI for pulmonary function test interpretation and COPD exacerbation prediction
  • Recognize implementation challenges specific to ICU environments
  • Address alert fatigue and its impact on AI adoption in critical care
  • Apply evidence-based frameworks for ICU AI implementation

The Clinical Context: Critical care generates more continuous data than any other medical setting: ventilator parameters, hemodynamic monitoring, continuous vital signs, frequent labs, and medication infusions. This data-rich environment creates opportunities for AI pattern recognition while also posing challenges: high alert burden, rapidly changing patient status, and life-or-death consequences of both false positives and false negatives.

What Works Well:

Application Systems Evidence Level Key Benefit
Automated closed-loop ventilation INTELLiVENT-ASV Large RCT No significant increase in ventilator-free days; earlier studies reported fewer manual adjustments
Chest CT triage Product-specific systems Observational and diagnostic Evidence and intended use vary by finding, product, and workflow
PFT interpretation Evaluated software Bounded multicenter case comparison Stronger performance than participating clinicians on 50 selected cases
AKI prediction Multiple research models Retrospective prediction Earlier risk estimation without demonstrated prevention benefit
Cardiac output monitoring Retia Argos Infinity (K253092) Non-AI comparator Conventional arterial-waveform signal processing

What’s Problematic:

Application Concern Reality
Epic Sepsis Model 33% sensitivity Widely deployed despite poor external validation
Generic deterioration indices Transportability and alert burden Require version- and threshold-specific local evaluation
Autonomous ICU management Not validated Research stage only

Critical Insights:

  • Alert burden is a major barrier. AI can add false or redundant alerts unless the complete response pathway is designed and measured.
  • External validation addresses transportability. It does not guarantee clinical utility, and failure is not universal across every model or institution.
  • Sepsis prediction and sepsis quality abstraction are different tasks. A 2026 cluster-randomized trial improved SEP-1 process compliance through LLM-enabled audit and feedback, without significant differences in ICU admission or 30-day mortality.
  • Retrospective AI-beats-clinician ICU prediction did not transfer cleanly to the bedside. Subspecialized physicians generally outperformed the model prospectively, and hybrid predictions were often best (Kuang et al., 2026).
  • A geriatric ICU severity model can beat static scores retrospectively. Elder-DDB reported external AUROC 0.801–0.834 and a lower alarm rate than SOFA at matched sensitivity; that is not a mortality-benefit claim (Gao et al., 2026).

The Bottom Line: ICU AI evidence must be separated into detection, prediction, automated control, quality abstraction, workflow, and patient outcomes. Local acceptance testing, a defined response pathway, and prospective monitoring are essential before deployment.

Introduction

Critical care appears well suited to algorithmic assistance because it combines continuous physiologic monitoring, structured data streams, and time-critical decisions. Yet the same environment has shifting baselines, treatment-confounded data, multiple simultaneous devices, and high alert burden. The external validation of the Epic Sepsis Model remains a prominent lesson in version-, site-, and threshold-specific evaluation.

Alert burden is a central implementation challenge. A system that adds notifications without a measured response benefit can dilute attention even when its technical discrimination appears acceptable.

The Sepsis Prediction Controversy

External Validation of the Epic Sepsis Model

The Epic Sepsis Model (ESM) represents a cautionary case for deployment before independent, setting-specific validation. Wong et al. evaluated the model at Michigan Medicine and reported results for the version, threshold, data availability, and workflow used there.

The Validation Study:

Wong et al. (2021) published the external validation in JAMA Internal Medicine (Wong et al., 2021):

  • Study population: 27,697 patients, 38,455 hospitalizations at Michigan Medicine
  • Sepsis prevalence: 7% of hospitalizations

Performance results:

Metric ESM Performance Clinical Implication
Sensitivity 33% Missed 67% of sepsis cases
Specificity 83%
Positive Predictive Value 12% 88% of alerts were false positives
AUC 0.63 Limited discrimination in this external evaluation

What this means clinically:

  • The ESM generated alerts for 18% of all hospitalized patients (6,971 of 38,455)
  • Of 2,552 patients who developed sepsis, the ESM failed to identify 1,709 (67%)
  • Only 183 patients with sepsis (7%) were identified by ESM who had not already received timely antibiotics

The potential alert-fatigue pathway:

  1. ESM generates alert for 18% of patients
  2. 88% are false positives
  3. Repeated nonactionable alerts can reduce attention and trust
  4. The organization may fail to measure whether true-positive alerts receive an appropriate response
  5. The study establishes poor external performance at the evaluated threshold; it did not randomize deployment or prove that the model worsened outcomes.

Why Sepsis Prediction Is Fundamentally Difficult

The definitional problem:

  • Sepsis criteria have changed (Sepsis-2 vs. Sepsis-3)
  • Model trained on one definition may not generalize to another
  • “Sepsis” is a syndrome, not a single disease

The timing problem:

  • By the time vital signs deteriorate enough for AI detection, sepsis is often clinically obvious
  • The value would be in pre-symptomatic detection, which remains elusive

The heterogeneity problem:

  • Sepsis from pneumonia differs from urosepsis, differs from soft tissue infection
  • Single algorithm struggles with diverse presentations

Alternative Approaches

Some institutions have developed local sepsis prediction models calibrated to their populations. The relevant implementation principles are:

  • Train on local data
  • Validate prospectively before deployment
  • Monitor continuously for drift
  • Adjust thresholds based on alert fatigue metrics

Multi-modal integration:

Emerging approaches combine:

  • Vital signs trends (not just thresholds)
  • Laboratory trajectory (lactate trend, WBC changes)
  • Nursing documentation (altered mental status, mottled skin)
  • Medication patterns (vasopressor initiation)

Current status: No class-wide conclusion follows from one sepsis model. Each model version needs external validation, prospective workflow evaluation, and an intervention pathway tied to patient-relevant endpoints.

2026 Cluster-Randomized Evidence: Quality Feedback, Not Sepsis Prediction

Boussina et al. randomized 66 attending emergency physicians at two academic emergency departments to receive near-real-time feedback based on LLM abstraction of SEP-1 bundle compliance or the usual delayed feedback process. Among 301 eligible patient encounters, modeled SEP-1 compliance was 82.9% in the intervention group and 70.1% in the control group, a 13.0-percentage-point absolute improvement (95% CI, 2.5–23.4 percentage points) (Boussina et al., 2026).

This was AI-enabled quality-measure abstraction plus targeted feedback, not a sepsis-onset prediction model. There were no significant differences in ICU admission or 30-day mortality. The trial was retrospectively registered at the journal editors’ request, and the higher SEP-1 compliance was driven largely by a documentation-sensitive fluid-bolus component. The defensible claim is improved process-measure compliance in this two-site workflow, not improved survival or proof that sepsis prediction works.

Part 2: ICU Early Warning Systems

Deterioration Prediction

Beyond sepsis, ICU AI attempts to predict clinical deterioration: cardiac arrest, respiratory failure, hemodynamic instability.

Epic Deterioration Index (EDI):

  • Produces a proprietary deterioration-risk score from EHR data
  • Inputs, refresh cadence, version, threshold, and routing must be verified in the deployed configuration
  • Deployment counts and performance from vendor materials should not be treated as peer-reviewed clinical evidence

Evidence:

  • Retrospective discrimination varies by institution, outcome definition, prediction horizon, model version, and threshold
  • A reported lead time can reflect when the label becomes observable rather than when an intervention can change the outcome
  • Prospective comparative evidence is required before attributing reductions in escalation, ICU transfer, cardiac arrest, or mortality to the score

Gao et al. trained an interpretable time-series model on 41,050 MIMIC-III/IV admissions aged 65 or older and externally validated on 47,335 admissions from US eICU, European, and Chinese cohorts (Gao et al., 2026). Cumulative 4-hour windows predicted in-hospital death within 24 h (external AUROC 0.801 US / 0.816 Europe / 0.834 Asia; internal 0.873 vs SOFA 0.650). At matched 80% sensitivity in the US external set, patient-level alarm rate was 29.7% vs 55.8% for SOFA. This is retrospective discrimination and lower alert volume at a matched threshold, not evidence that deploying Elder-DDB reduces deterioration, cardiac arrest, or mortality.

Kuang et al. compared site-trained models with clinicians on the same ICU outcomes (in-hospital and 30-day mortality, AKI, delirium, ICU and hospital length of stay) in a mixed retrospective arm (models trained on 46,631 admissions; 990-admission comparison from chart data only) and a prospective arm (238 admissions; clinician predictions collected during active care) (Kuang et al., 2026). In the retrospective arm, AI generally outperformed individual clinicians (one-clinician in-hospital mortality accuracy 0.806, 95% CI 0.774–0.838, vs AI 0.864, 0.836–0.891), while aggregated predictions from seven clinicians could outperform AI (accuracy 0.800 vs 0.760 on n = 50 overlapping cases); in the prospective arm, subspecialized physicians generally outperformed AI (in-hospital mortality accuracy 0.852 vs 0.820) whereas physicians in training and nurses generally did not, and clinician-AI combinations often performed best. This is a discrimination and calibration comparison, not an alert-deployment or patient-outcome trial.

Critical limitations:

  • Predictive value is local: Translate sensitivity, specificity, and prevalence into the expected number of true and false alerts at the deployed threshold
  • Alert fatigue: Nurses begin ignoring high-frequency alerts
  • Response burden: Every alert requires RRT evaluation
  • No RCT evidence for mortality benefit

Alert Fatigue: The Core Problem

Quantifying the crisis:

  • A review described hundreds of alarms per patient-day in some ICU environments and reported that many were false or clinically irrelevant (Sendelbach and Funk, 2013)
  • Alarm burden varies by unit, monitoring configuration, thresholds, patient acuity, and what investigators count as actionable
  • Measure acknowledgment, response, and patient outcome rather than assuming a universal override rate

AI’s contribution:

  • Potential: Reduce false alerts through intelligent filtering
  • Reality: Often adds to alert burden without solving root problem

What works:

  1. Tiered alert systems: Critical vs. warning vs. informational
  2. Alert grouping: Combine related alerts into single notification
  3. Threshold optimization: Site-specific calibration to reduce false positives
  4. Alert resolution: Automatic silencing when trigger condition resolves
  5. Regular audits: Disable low-value alerts quarterly

The Alert Fatigue Paradox

Adding an AI system to an ICU already overwhelmed by alerts may worsen outcomes by:

  1. Increasing total alert volume
  2. Diluting attention from critical alarms
  3. Creating “alert lottery” where responses are random
  4. Desensitizing staff to all warning systems

Before deploying any AI alert system, audit the current alert burden, define the response pathway, and identify redundant low-value notifications that can be safely removed through the institution’s alarm-governance process.

Part 3: Automated Mechanical Ventilation and AI Research

Closed-Loop Ventilation

INTELLiVENT-ASV (Hamilton Medical), a non-ML comparator:

INTELLiVENT-ASV is an automated closed-loop ventilation system. The cited clinical literature does not establish it as a machine-learning model. It is included because it shows what conventional feedback control can and cannot accomplish while newer AI ventilation research remains investigational. The system automatically adjusts:

  • Tidal volume
  • Respiratory rate
  • PEEP
  • FiO2
  • Inspiratory pressure

Based on continuous SpO2 and end-tidal CO2 monitoring.

Evidence from systematic reviews:

A systematic review identified 10 RCTs examining INTELLiVENT-ASV (Botta et al., 2021):

What earlier studies reported:

  • Fewer manual ventilator adjustments in evaluated settings, including a reported comparison of 5 versus 10 adjustments per day
  • Maintains safe tidal volumes and airway pressures
  • Effective titration of SpO2 and PETCO2 targets
  • Nurses and physicians find it easier to use (P<0.001)

What remains uncertain:

  • No proven benefit for duration of ventilation
  • No proven mortality benefit
  • Studies underpowered for patient-centered outcomes
  • Efficacy in severe ARDS not established

ACTiVE Trial results (JAMA, March 2026):

The ACTiVE trial enrolled 1,201 patients (602 closed-loop, 599 conventional) across multiple centers. The primary endpoint, median ventilator-free days at day 28, was 16.7 versus 16.3 days and did not differ significantly. The trial found no evidence of an increase or decrease in ventilator-free days with automated closed-loop ventilation compared with protocolized conventional ventilation (Sinnige et al., 2026).

This supports a narrower framing: the system can automate adjustments and improve selected physiologic-control measures, but the large randomized trial did not improve its patient-centered primary endpoint. Workflow efficiency should be measured directly rather than inferred from the presence of automation.

Ongoing trials:

  • POSITiVE II Trial: A registered cardiac-surgery comparison should be described from its current registry record before status, enrollment, or completion dates are treated as current facts.

Weaning Prediction

AI models attempt to predict readiness for extubation:

Approaches:

  • RSBI (Rapid Shallow Breathing Index) threshold optimization
  • Multi-parameter prediction combining strength, secretions, mental status
  • Trend analysis of ventilator settings over time

Current status:

  • Promising research findings
  • No authorized weaning-prediction record is cited here; any marketing claim should be checked against the exact FDA record and intended use
  • Clinical judgment remains standard of care

ARDS Management Support

Emerging applications:

  • Optimal PEEP titration based on lung mechanics
  • Prone positioning decision support
  • Driving pressure monitoring and alerts

Limitations:

  • ARDS is heterogeneous (phenotypes vary widely)
  • Optimal ventilator settings depend on underlying cause
  • AI cannot replace bedside assessment of patient tolerance

Part 4: Pulmonary Imaging AI

Chest CT Analysis

Critical care frequently requires emergent chest imaging. AI assists with:

Pulmonary Embolism Detection:

  • Multiple products have authorization records for defined pulmonary-embolism triage functions; each product, finding, population, and scanner configuration requires its own evidence
  • Do not transfer sensitivity, time-to-notification, or outcome results from one product to Aidoc, Avicenna.AI, RapidAI, or a class label
  • Workflow: Automated triage, flagging critical studies for immediate review
  • Value question: Does prioritization shorten verified treatment intervals without increasing harmful false alerts or missed findings?

Pneumothorax Detection:

  • Multiple product-specific authorization records exist for defined pneumothorax triage functions
  • Performance depends on study design, radiograph type, patient population, pneumothorax size, and hardware
  • May miss small pneumothoraces, especially in complex post-surgical patients
  • Value must be measured in the deployed ED, trauma, or ICU workflow

COVID-19 and Pneumonia:

  • Hype exceeded reality during pandemic
  • Many AI models learned confounders (portable vs. PA, patient positioning) not pathology (DeGrave et al., 2021)
  • Current role: workflow triage, not diagnostic

Chest X-Ray AI in the ICU

Applications:

  • Endotracheal tube position verification
  • Line and catheter placement assessment
  • Pneumothorax detection
  • Pulmonary edema quantification

Limitations:

  • ICU portable films have lower image quality
  • Patient positioning varies
  • Overlying hardware (lines, tubes, monitors) obscures anatomy
  • AI trained on standard PA films may perform poorly on portable AP images

ICU-Specific Imaging Challenges

AI systems trained on outpatient imaging data often fail in ICU settings due to:

  1. Portable AP vs. PA technique: Different magnification, positioning
  2. Overlying hardware: Central lines, NG tubes, ECG leads
  3. Motion artifact: Agitated patients, respiratory motion
  4. Prior imaging comparison: ICU patients have complex imaging histories
  5. Clinical context: Post-operative changes, known pathology progression

Demand ICU-specific validation before deploying any imaging AI in the unit. Product-specific regulatory and evidence details are maintained in the Radiology chapter.

Pulmonary Function Test (PFT) Interpretation AI

AI interpretation of spirometry addresses a gap in primary care and general pulmonology, where spirometry is often available but expert interpretation is not.

Performance vs. pulmonologists:

  • In a 50-case comparison involving 120 pulmonologists from 16 European hospitals, the evaluated software matched the reference PFT pattern in 100% of cases versus a mean of 74.4% for participating pulmonologists; the reference diagnosis was assigned in 82% versus 44.6% of interpretations (Topalovic et al., 2019). The result is bounded to this software, case sample, reference standard, and information supplied.
  • Pulmonologist interrater variability for diagnosis was high (kappa = 0.35), while AI produced consistent interpretations
  • AI spirometry diagnostic support achieved high sensitivity and specificity for identifying COPD in a blinded primary care validation study (European Respiratory Journal Open Research, 2025)
  • ML models predicting restrictive patterns from spirometry improve both accuracy and equity, addressing known disparities in FVC lower limit of normal criteria across racial groups

Chest X-ray to PFT estimation:

A deep-learning model estimated selected pulmonary-function measures from chest radiographs across multiple Japanese institutions (Lancet Digital Health, 2024). Estimated values are not equivalent to a complete pulmonary-function study and should not be used to claim that radiography replaces spirometry, lung volumes, or diffusing-capacity measurement.

Clinical utility:

  • Greatest value in primary care, where misdiagnosis of COPD is common and access to pulmonologists is limited
  • May reduce unnecessary referrals while improving detection of restrictive disease
  • Does not replace full PFTs with body plethysmography when clinically indicated

COPD and Asthma AI Beyond the ICU

For outpatient COPD exacerbation prediction and its limitations, see Internal Medicine and Hospital Medicine.

2025 meta-analysis update: A systematic review of 13 studies (2017–2025) found pooled AUC of 0.77 (95% CI 0.74–0.80) for AI/ML prediction of COPD exacerbations and readmissions; readmission subgroup AUC 0.73 and readmissions, though high heterogeneity and limited external validation restrict generalizability (Frontiers in Digital Health, 2025).

Digital inhaler sensors: A 12-week clinical study using digital dry powder inhalers with integrated sensors demonstrated that ML algorithms could detect declining peak inspiratory flow and inhalation volume in the days preceding an exacerbation, offering a novel data source beyond traditional patient-reported symptoms (PMC, 2025).

Asthma AI: Research and commercial platforms combine wearable sensors, environmental measurements, symptoms, and prediction models. Product names and regulatory status change, so each intended use should be verified from an exact primary record. The cited evidence is dominated by feasibility and prediction studies, with limited prospective patient-centered outcome evidence.

Part 5: Acute Kidney Injury Prediction

The Clinical Need

Acute kidney injury is common in critical care and is associated with worse outcomes. Earlier risk identification could support:

  • Nephrotoxin avoidance
  • Hemodynamic optimization
  • Earlier nephrology consultation
  • Reduced dialysis requirement

AI Approaches

DeepMind/Google Health AKI Model:

  • Predicted AKI up to 48 hours before creatinine rise
  • Trained on VA data (700,000+ patients)
  • Published in Nature (Tomašev et al., 2019)

Performance in the retrospective VA evaluation:

  • The model predicted 55.8% of all inpatient AKI episodes and 90.2% of episodes followed by dialysis, with up to 48 hours of lead time
  • At the operating point used for the 55.8% result, the model generated two false alerts for every true alert
  • ROC-AUC was 0.921 for prediction of any AKI within the 48-hour window, while precision-recall AUC was 0.297 (Tomašev et al., 2019)

Limitations:

  • Significant male predominance in VA training data (94% male)
  • External validation in other populations limited
  • Implementation challenges: how should clinicians respond to alerts?
  • The retrospective prediction study did not test whether alerts prevented AKI, reduced dialysis, or improved survival

Implementation Considerations

The intervention gap:

Predicting AKI creates clinical value only when the response improves care without causing offsetting harm. Candidate actions include:

  • Review potentially nephrotoxic exposures and medication dosing
  • Reassess hemodynamics and volume status in clinical context
  • Reconsider avoidable contrast exposure without delaying necessary diagnosis or treatment

Alert design:

  • What action is expected when AKI risk alert fires?
  • Who receives the alert (ICU team, nephrology, pharmacy)?
  • How frequently should alerts repeat for sustained high risk?

Part 6: Hemodynamic Monitoring and Prediction

Hypotension Prediction

Hypotension Prediction Index (HPI) by Edwards Lifesciences, a predictive-monitoring comparator:

  • Uses arterial-waveform analysis to estimate risk of a defined hypotensive event
  • The exact device record, compatible hardware, population, prediction horizon, and intended use must be verified before deployment
  • A predictive index is not automatically evidence of machine learning or improved outcomes

Evidence:

  • Initial studies showed potential for earlier intervention
  • Mixed results in subsequent validation
  • Requires arterial line (not universal in all ICU patients)

Cardiac Output and Fluid Responsiveness

Argos Infinity by Retia Medical:

Argos Infinity received FDA 510(k) clearance under K253092 on February 13, 2026. It is intended for continuous cardiac-output measurement from a radial arterial-pressure signal in adults using specified compatible monitoring configurations. Its Multi-Beat Analysis is conventional hemodynamic signal processing; the FDA record does not establish a machine-learning component. It is included as a non-AI comparator.

Independent validation:

  • In a head-to-head comparison against Edwards Lifesciences FloTrac in 58 off-pump coronary bypass patients (572 paired measurements), the Argos monitor showed lower CO error (RMSE 1.16 vs. 1.54 L/min) and higher trending concordance (87% vs. 72%, p < 0.0167) against pulmonary artery thermodilution reference (Mukkamala et al., 2021). Note: the first author co-developed the MBA algorithm.
  • ICU validation in 31 post-cardiac surgery patients showed reasonable agreement with thermodilution (bias 0.08 ± 1.10 L/min, concordance 88%), though percentage error was 40.7% (Greiwe et al., 2020). Clinical rationale: Cardiac output can change without a corresponding categorical blood-pressure threshold, so continuous waveform-derived monitoring may add information in selected patients. That physiologic rationale does not establish that the monitor improves outcomes.

Limitations: The system requires radial arterial-line access and compatible validated configurations. Suitability should be judged against prespecified criteria for the intended clinical use rather than a universal percentage-error threshold. Most validation studies are in cardiac-surgery populations; generalizability to medical ICU or sepsis populations remains unestablished.

Other approaches in development:

  • Fluid responsiveness prediction from pulse pressure variation (PPV) and stroke volume variation (SVV)
  • Non-invasive cardiac output estimation from photoplethysmography
  • Clinical utility of all approaches depends on integration with goal-directed treatment protocols

Part 7: Implementation Challenges in Critical Care

The ICU Environment

Why ICU AI is harder than outpatient AI:

Challenge ICU Outpatient
Data velocity Continuous streams Intermittent snapshots
Patient stability Rapidly changing Relatively stable
Decision timeframe Minutes Days to weeks
Alert tolerance Already overwhelmed More capacity
Consequence of error Often immediate and life-threatening Often less time-sensitive
Team complexity Multidisciplinary, 24/7 Physician-centric

Workflow Integration

Lessons from failed implementations:

  1. Standalone dashboards can fail when outputs do not reach the clinical workflow.
  2. Alert placement matters: Buried alerts are ignored
  3. Response protocols required: What should happen when AI flags a patient?
  4. Feedback loops essential: Clinicians must be able to indicate false positives

Cost and Value Assessment

Static license ranges, rapid-response-team costs, and a calculated “cost per true positive” are not reliable without a current contract and a measured local workflow. A defensible analysis includes:

  • Current licensing, integration, validation, training, support, monitoring, and upgrade costs
  • Staff time for every alert, including false and duplicate notifications
  • Downstream tests, treatments, delayed care, and rapid-response activations
  • Patient-relevant benefits and harms relative to the existing response system
  • Sensitivity analyses for prevalence, threshold, staffing, and model drift

Earlier detection is not the same as prevention. Economic value depends on whether the alert changes timely care and improves a prespecified endpoint.

Professional Society Guidance and Evidence Reviews

Operationalization Evidence Gap

A 2025 JAMA Network Open systematic review of ICU AI operationalization screened 17,401 records and included 1,263 studies. Most remained early-stage: 936 studies (74%) were technology-readiness level 4 or below, only 24% reached external validation, only 25 studies (2%) reached clinical integration, and none reached full implementation. High risk of bias was identified in 53% of assessed studies (Berkhout et al., 2025).

For intensivists, this is the implementation baseline. ICU AI papers are abundant, but operationalized, prospectively tested tools remain rare. A vendor claiming “validated ICU AI” should be asked whether validation means internal model testing, external validation, silent deployment, clinical integration, or measured patient benefit.

Society of Critical Care Medicine (SCCM)

SCCM has addressed AI through educational material and journal-publication policy. Its current clinical-practice-guideline directory does not identify a general AI deployment guideline for critical care. Educational programming and authorship policy should not be presented as product-specific clinical recommendations.

SCCM Journal Policy on AI (2024)

The SCCM journals established fair use statements for AI in scientific publication:

Core Principles:

  1. Human accountability: AI tools are neither independent nor sentient. Only humans can be held accountable, therefore only humans can be named as authors.

  2. Fair use: Using computers to enhance insight, foster hypothesis generation, facilitate analyses, and improve prose quality is fair. Using AI as a substitute for critical thinking is strongly discouraged.

  3. Transparency required: Failure to acknowledge AI use in submissions is considered a violation of scientific integrity.

Source: (SCCM, 2024)

Consensus Statement on AI Implementation (2025)

An expert consensus published in Critical Care (2025) addressed AI integration in ICUs:

22-Expert Consensus on AI in Critical Care (2025)

Key Recommendations:

  1. The expert panel called for professional societies to establish specific clinical-practice guidance for AI in critical care

  2. Standards needed for:

    • Model validation requirements
    • Clinician-AI collaboration frameworks
    • Accountability structures

  3. Regulatory bodies must adapt oversight to rapidly evolving AI capabilities

  4. Equity, transparency, and the patient-clinician relationship must be prioritized

Source: (Cecconi et al., 2025)

This is expert consensus, not a graded society clinical-practice guideline.

SCCM Educational Initiatives

SCCM’s 2025 Critical Care Congress featured dedicated AI programming:

Deep Dive Course (February 2025):

  • Foundational skills for AI in critical care
  • Real-world demonstrations of patient outcome enhancement
  • Ethical and legal considerations
  • Data privacy, consent, and accountability
  • Addressing biases inherent in AI systems

Interpretation: Conference programming describes topics selected for education. It is not evidence that a product improves outcomes or a formal SCCM recommendation for deployment.

European Society of Intensive Care Medicine (ESICM)

ESICM has educational and scientific activity involving AI. Its current guidelines and consensus directory does not list a general clinical-practice guideline for ICU AI deployment. Relevant topics include:

  • AI for sepsis prediction and management
  • Machine learning in ICU prognostication
  • Ethical frameworks for AI implementation

American Thoracic Society (ATS)

ATS publications and meetings address pulmonary AI applications including:

  • Interstitial lung disease pattern recognition
  • COPD exacerbation prediction (pooled AUC 0.73–0.77 in 2025 meta-analysis; see COPD and Asthma AI)
  • Pulmonary function test interpretation (AI outperforms pulmonologists; see PFT AI)
  • Digital inhaler sensors for exacerbation monitoring

No AI-specific ATS clinical-practice guideline is cited in this chapter. Pulmonary imaging, PFT interpretation, and COPD prediction should therefore be evaluated through their underlying studies rather than represented as society-endorsed practice.

Check Your Understanding

The following cases are fictional decision exercises. They do not describe actual patients, institutions, product results, costs, legal holdings, or standards of care.

Scenario 1: The Sepsis Alert Override

Clinical situation: An overnight ICU attending receives a high-risk sepsis alert for a patient admitted after elective hip replacement. The patient is afebrile and hemodynamically stable, with no localizing sign of infection. This fictional exercise uses an unspecified model rather than asserting how the Epic Sepsis Model behaves in this patient.

Question: How do you respond to this alert?

Answer: Assess the patient and apply the institution’s sepsis pathway. The score alone neither establishes nor excludes sepsis.

Reasoning:

  1. Context matters: Postoperative physiologic changes can overlap with infection, so the alert must be interpreted with the examination, trajectory, source evaluation, and relevant tests
  2. Population performance is not an individual probability: The 12% positive predictive value reported in the Michigan Medicine ESM evaluation cannot be transferred to an unspecified model, threshold, or patient
  3. Clinical assessment: No fever, no hypotension, no localizing infection source

What to do:

  1. Assess the patient clinically (not remotely based on alert alone)
  2. If assessment confirms low suspicion, document: “Sepsis alert reviewed. Clinical assessment shows post-operative changes consistent with normal recovery. No clinical signs of sepsis. Will monitor.”
  3. Do NOT reflexively order blood cultures, broad-spectrum antibiotics, or lactate if not clinically indicated
  4. Provide feedback to informatics team about false positive

Lesson: AI alerts require clinical context. The purpose of AI is to prompt evaluation, not dictate treatment. An alert does not obligate action if clinical assessment contradicts it.

Scenario 2: Evaluating a New Deterioration Prediction Tool

Clinical situation: A hospital is considering a fictional vendor’s deterioration-prediction tool. The vendor presents illustrative development-site results of 92% sensitivity and 0.89 AUC. These numbers are prompts for appraisal, not claims about a real product.

Question: What questions should you ask before agreeing to implementation?

Answer: Demand external validation data, understand false positive rates, and require a pilot protocol.

Key questions:

  1. External validation: “What is performance at institutions outside your development site?”

    • If answer is “none” or “similar,” be skeptical

  2. False positive rate: “What is the positive predictive value at your recommended threshold?”

    • Sensitivity alone does not determine the false-alert burden
    • Ask: “For every true deterioration detected, how many false alerts?”

  3. Alert volume: “How many alerts per nurse shift should we expect?”

    • Set an acceptable burden through local workflow testing rather than a universal cutoff

  4. Local pilot: “Can a silent evaluation run long enough to capture representative case mix, staffing, and temporal variation before clinical deployment?”

    • Compare predictions to actual outcomes on your population

  5. Integration: “How does this integrate with our EHR workflow?”

    • A standalone dashboard may fail if the responsible team does not see or act on it

  6. Threshold control: “Can we adjust sensitivity/specificity thresholds for our population?”

    • One-size-fits-all thresholds rarely work

  7. Performance monitoring: “What tools exist to track real-world performance after deployment?”

  8. Exit strategy: “If performance is poor, can we turn it off without penalty?”

Red flags:

  • Vendor cannot provide external validation
  • Vendor refuses silent pilot
  • Threshold cannot be adjusted
  • No ongoing performance monitoring capability
Scenario 3: The Ventilator Autonomy Decision

Clinical situation: An ICU has INTELLiVENT-ASV available. A 67-year-old patient with COPD exacerbation is intubated and has moderate ARDS. The respiratory therapist asks whether the automated closed-loop mode fits the unit’s protocol and the patient’s current physiology.

Question: Is INTELLiVENT-ASV appropriate for this patient?

Answer: The decision should follow the device labeling, local ventilation protocol, clinician and respiratory-therapist assessment, and a predefined monitoring and override plan.

When INTELLiVENT-ASV may be appropriate:

  • The patient falls within the device’s intended use and local protocol
  • Required sensors and signal quality are reliable
  • Staff are trained to recognize inappropriate automated settings and override promptly
  • Patient-specific lung mechanics, gas exchange, synchrony, and treatment goals can be monitored

Caveats for this patient:

  1. ARDS consideration: Verify that delivered ventilation remains consistent with the current lung-protective protocol
  2. COPD physiology: Assess auto-PEEP, air trapping, synchrony, and whether automation responds appropriately
  3. Changing severity: A mode acceptable at one time point may require override as physiology or treatment goals change

What to do:

  1. Use the unit’s protocol for initiation and monitoring
  2. Verify delivered tidal volume, pressure, oxygenation, ventilation, synchrony, and alarms against patient-specific targets
  3. Document the conditions that require clinician review or transition to another mode
  4. Maintain immediate manual-override capability
  5. Do not treat automation as unattended care

Key insight: Closed-loop ventilation is a tool, not a substitute for clinical judgment. It reduces manual adjustments but does not eliminate the need for oversight.

Scenario 4: Alert Fatigue Intervention

Clinical situation: An ICU medical director receives reports that staff are routinely silencing alarms without assessment. A fictional near-miss prompts review of whether high alert burden and unclear ownership are degrading response.

Question: How do you address this systemic problem?

Answer: Conduct an alert audit, reduce total alert volume, and redesign the alert response workflow.

Step 1: Quantify the problem

  • How many alerts per patient per day?
  • What percentage are false positives?
  • Which alert types have highest override rates?

Step 2: Reduce alert burden

  • Review or retire alerts with poor actionability through the institution’s alarm-governance and patient-safety process
  • Combine redundant alerts (do not alert for tachycardia AND high heart rate separately)
  • Implement delay periods (do not alert for 30-second SpO2 dip during suctioning)

Step 3: Stratify alerts

  • Critical (must act): V-fib, asystole, severe hypoxia
  • Warning (assess soon): Deterioration trend, vital sign outlier
  • Informational (acknowledge only): Lab results, medication reminders

Step 4: Accountability structures

  • Critical alerts cannot be silenced without documentation
  • Regular review of alert responses at quality meetings
  • Non-punitive reporting system for near-misses

Step 5: Before adding any new AI

  • Commit to removing equivalent alert volume
  • Pilot with limited patient population first
  • Track whether new AI alerts are acted upon

The operational lesson: Adding AI before measuring and redesigning the existing alert system can increase burden. Deployment should include an alert budget, response ownership, stop rules, and prospective monitoring.

Key Takeaways

Clinical Bottom Line for Critical Care AI

  1. The Epic Sepsis Model is a version-specific cautionary case. The Michigan Medicine evaluation reported 33% sensitivity and 12% positive predictive value at the evaluated threshold. External validation matters.

  2. Automated closed-loop ventilation is not the same as AI. The ACTiVE trial found no significant increase in ventilator-free days, despite earlier evidence of fewer manual adjustments.

  3. Alert burden is a major barrier. AI that adds notifications needs measured actionability, response ownership, and stop rules.

  4. Imaging evidence is product- and workflow-specific. Pulmonary-embolism and pneumothorax triage results must remain connected to the exact product, finding, population, and endpoint.

  5. AKI prediction is not AKI prevention. The VA model predicted selected events up to 48 hours early but did not test whether alerts improved outcomes.

  6. Local acceptance testing is essential. Models can improve, preserve, or lose performance after transport; the answer must be measured rather than assumed.

  7. Use a representative silent evaluation. Duration and sample size should follow prevalence, endpoint, uncertainty, subgroup, staffing, and seasonal requirements rather than a universal calendar period.

  8. Do not confuse expert consensus with a clinical-practice guideline. The 2025 ICU AI paper offers expert recommendations; current SCCM and ESICM guideline directories do not identify a general ICU AI deployment guideline.

  9. AI augments, never replaces, clinical judgment. The sickest patients need the most human oversight, not the least.

  10. Economic value is use-case-specific. Measure full deployment cost, alert response, downstream care, displacement, and patient-relevant benefit.

  11. The Topalovic PFT comparison was strong but bounded. One software system outperformed participating pulmonologists on 50 selected cases; that result is not universal superiority across products or settings.

  12. COPD prediction requires a tested response pathway. Pooled discrimination does not show that exacerbations or readmissions are preventable through the model.

  13. Retrospective AI-beats-clinician ICU prediction did not transfer cleanly to the bedside. In Kuang et al. 2026, subspecialized physicians generally outperformed the model prospectively, and hybrid predictions were often best. This is not evidence that an ICU outcome model should be deployed.

Questions About Critical Care and Pulmonary AI

How accurate is the Epic Sepsis Model?

In one external validation at Michigan Medicine, the Epic Sepsis Model had 33% sensitivity and 12% positive predictive value at the evaluated threshold. Those results describe that model version, institution, population, and threshold; they are not class-wide sepsis-model performance.

What is the main barrier to ICU AI adoption?

Alert burden is a major implementation barrier. Published reviews describe very high alarm volumes in critical care, but the local burden must be measured by unit, device, alert type, shift, and response pathway before adding an AI alert.

Does automated closed-loop ventilation improve outcomes?

INTELLiVENT-ASV is automated closed-loop control, not established machine-learning evidence. Earlier studies reported fewer manual adjustments, while the 1,201-patient ACTiVE randomized trial found no significant increase in ventilator-free days at day 28.

Why does sepsis prediction remain difficult for AI?

Sepsis labels, onset times, source conditions, treatment patterns, and case mix vary across institutions. A prediction model also needs an evaluated response pathway, because earlier risk estimation alone does not establish better patient outcomes.

Can AI interpret pulmonary function tests better than pulmonologists?

In Topalovic et al.’s 50-case comparison, the evaluated software matched the reference PFT pattern in 100% of cases versus a mean of 74.4% for 120 pulmonologists, and assigned the reference diagnosis in 82% versus 44.6%. This bounded study does not establish universal superiority across populations or products.

How well can AI predict COPD exacerbations?

A 2025 review of 13 studies reported pooled AUC 0.77 for exacerbation and readmission prediction and 0.73 in the readmission subgroup. Heterogeneity and limited external validation restrict clinical interpretation, and prediction alone does not prove preventability.

What is Retia Medical Argos Infinity for cardiac output monitoring?

Argos Infinity is FDA 510(k)-cleared under K253092 for continuous cardiac-output measurement from a radial arterial-pressure signal in adults. Its Multi-Beat Analysis is conventional hemodynamic signal processing, not documented machine learning, so it is a non-AI comparator in this chapter.