Development and validation of a reinforcement learning algorithm to dynamically optimize mechanical ventilation in critical care
PubMed • Full text
[[Category:]]
Clinical Question
Among patients in a historical ICU database receiving volume-controlled mechanical ventilation, does an AI-driven reinforcement learning algorithm that dynamically optimizes ventilator settings (tidal volume, PEEP, FiO₂) improve probability of 90 day survival or in-hospital survival?
Bottom Line
Among patients in a historical ICU database receiving volume-controlled mechanical ventilation use of an AI-driven reinforcement learning algorithm that dynamically optimizes ventilator settings (tidal volume, PEEP, FiO₂) is associated with improved probability of 90 day survival or in-hospital survival.
Major Points
The use of AI in critical care is evolving. There is a need to develop and externally validate such models prior to their implementation. The present manuscript developed and validated the VentAI model in ICU-specific databases (MIMIC-III and eICU). The model was intended to optimize vent settings and this manuscript describes model development, validation, and estimation of likelihood of 90 day survival and hospital survival in these databases. VentAI outperformed clinicians. A next step would be implementing this AI model at the bedside in a randomized design to determine if it improves clinical outcomes in prospective cohorts.
Guidelines
As of March 2025, no guidelines have been published that reflect the results of this trial.
Design
- Retrospective cohorts used for model development and validation.
- Data sources:
- MIMIC-III (2001-2012) → 61,532 ICU stays, 11,943 mechanical ventilation events
- eICU (2014-2015) → 200,859 ICU stays, 25,086 mechanical ventilation events
- Intervention: AI-driven VentAI model for optimizing ventilation in these historical databases
- Comparison: Standard clinician-guided mechanical ventilation
- Analysis: Performance return measured via reinforcement learning (Q-learning)
- Primary Outcome: 90 day mortality and hospital survival estimation
Population
Inclusion Criteria
- Critically ill adult patients (≥18 years old)
- Receiving volume-controlled mechanical ventilation
- Documented ventilator settings (Vt, PEEP, FiO₂)
- ICU stay data available in MIMIC-III or eICU databases
Exclusion Criteria
- Patients with missing ventilator data
- Patients withdrawn from mechanical ventilation early <72 hours
- Non-volume-controlled ventilation modes (e.g., pressure-controlled ventilation)
Baseline Characteristics
From the MIMIC-III dataset.
- Demographics: Age 67 years, female sex 36%
- 90 day mortality: 15.8%
- In-hospital mortality: 11.1%
- ICU LOS: 9 days
- Vent settings: PEEP 6 cm H2O, FiO2 46%, Vt 8.3 mL/kg
- SOFA at admission: 5.6
Interventions
- AI Model (VentAI):
- Applied reinforcement learning (Q-learning) to dynamically adjust ventilator settings every 4 hours.
- Optimized tidal volume (Vt), PEEP, and FiO₂ based on patient-specific data.
- Clinician-Guided Ventilation:
- Standard ventilator settings chosen by ICU clinicians.
- Less frequent adjustments compared to VentAI.
Outcomes
Primary Outcomes
- 90 day mortality
- Summary statistics of this comparison are not clearly provided by the authors. The authors describe VentAI outperforming clinicians.
- In-hospital mortality
- Summary statistics of this comparison are not clearly provided by the authors. The authors describe VentAI outperforming clinicians.
Secondary Outcomes
- Frequency of ventilator adjustments
- VentAI adjusted ventilator settings every 4 hours, while clinicians maintained more static settings
- Use of lung-protective ventilation (tidal volume 5-7.5 mL/kg)
- VentAI used protective Vt 202.9% more often than clinicians
- Avoidance of excessive tidal volume (>10 mL/kg)
- VentAI reduced use of high Vt by 50.8%
- Moderate PEEP usage (5-9 cmH₂O)
- VentAI increased moderate PEEP use by 53.6%
- Avoidance of high FiO₂ (>55%)
- VentAI reduced excessive FiO₂ use by 59.8%
Adverse Events
No direct clinical interventions were performed, so no adverse events were reported. Potential risk in real-world implementation: Over-reliance on AI could lead to errors in clinical judgment if not properly monitored.
Criticisms
- Retrospective study → No prospective, real-world testing of VentAI.
- No randomization → Risk of confounding bias since VentAI’s decisions were compared to historical clinician data.
- Data limitations → ICU data from 2001-2015 may not represent modern ventilator management strategies.
- Limited to volume-controlled ventilation → Findings do not apply to pressure-controlled or other ventilation modes.
- Ethical concerns → The use of AI in critical care raises questions about autonomy, liability, and physician oversight.
Funding
European Institute of Innovation & Technology (EIT Health)