Evaluation of AI Models in Determining the Optimal PEEP

Not yet recruiting

• Conditions: PEEP, Occult

• Registration Number: NCT06844916
• Lead Sponsor: Kanuni Sultan Suleyman Training and Research Hospital

Brief Summary

his study is designed as a prospective observational clinical trial. Patients over 18 years old who are hemodynamically stable and require mechanical ventilation in the Intensive Care Unit (ICU) will be included. The inclusion criteria ensure that participants require individualized ventilatory optimization.

The study will involve a comparison between Artificial Intelligence (AI)-generated Positive End-Expiratory Pressure (PEEP) recommendations and expert-determined PEEP levels. ICU specialists will perform PEEP titration manually based on standardized protocols, identifying the lower inflection point (LIP) and upper inflection point (UIP) to optimize ventilation. The pressure-volume (P-V) curve will be analyzed to ensure optimal alveolar recruitment and prevent overdistension.

Study Procedures

Participants will:

Undergo systematic mechanical ventilation assessments, including inspiratory hold and expiratory hold maneuvers, compliance, elastance, auto-PEEP, and time constant evaluations.

Have ventilation data collected and analyzed using three AI models: ChatGPT, DeepSeek, and Gemini.

Receive AI-generated recommendations regarding optimal PEEP levels, abnormal ventilation parameters, and potential treatment suggestions.

Have their AI-based PEEP recommendations compared with those determined by ICU specialists with at least five years of experience.

Outcome Measures

The study will compare AI and expert assessments based on the following primary and secondary measures:

Primary Outcome: Agreement between AI-generated PEEP levels and expert-determined PEEP levels.

Secondary Outcomes:

AI sensitivity and specificity in detecting abnormal ventilation parameters. Accuracy of AI-generated diagnoses. Clinical applicability of AI-recommended treatment strategies. This study aims to determine whether AI models can serve as reliable clinical decision support tools for ventilator management in ICU patients.

Detailed Description

Patients receiving mechanical ventilation support in the intensive care unit (ICU), who are hemodynamically stable and over 18 years old, will be considered eligible for the study. These criteria have been set to ensure that the patients require adequate respiratory support while being suitable for individualized ventilator management optimization.

During the patient selection process, the medical records of ICU patients will be reviewed, and those meeting the criteria will be included in the study. To ensure the accuracy of clinical data, the patient selection process will be carried out by intensive care specialists.

The mechanical ventilation parameters of the included patients will be systematically recorded. In this context, inspiratory hold and expiratory hold maneuvers will be applied to evaluate the respiratory mechanics of the patients. The recorded measurements will include compliance (lung elasticity), elastance (lung stiffness), auto-PEEP (presence of air trapping), and time constant (gas exchange duration).

In addition to these fundamental ventilation parameters, the pressure-volume (P-V) curve will be analyzed to collect critical data for ventilation optimization. These procedures will be performed using advanced analysis modules integrated into ICU ventilators and patient monitoring systems. The collected data will be analyzed both manually by experts and by artificial intelligence (AI) models throughout the study.

In the study, PEEP titration will be performed manually by intensive care specialists. This process will be applied immediately after intubation for intubated patients and will continue daily. The procedure will be standardized according to a predefined protocol and applied consistently across all patients.

The lower inflection point (LIP) will be identified, and PEEP will be applied at the point where alveolar recruitment begins. This step is critical to preventing lung collapse and ensuring optimal oxygenation. Additionally, the upper inflection point (UIP) will be determined to prevent overdistension, ensuring that tidal volume remains below this pressure threshold. If the lower inflection point is not clearly identified, a PEEP level of approximately 10 cmH₂O will be applied as a standard approach.

The pressure-volume curve of the ventilator will be analyzed to assess whether the optimal PEEP level has been achieved. If the curve shifts upward, it will be considered an indicator of successful alveolar recruitment. If the curve shifts to the right, it will indicate a risk of overdistension, prompting a reduction in PEEP. By this approach, the most suitable PEEP level for each patient will be determined, and ventilator settings will be optimized accordingly. These procedures are routinely applied in our hospital and globally and are among the most commonly preferred methods for PEEP titration.

All collected ventilation parameters will be input into three different artificial intelligence models: ChatGPT, DeepSeek, and Gemini. These systems will be asked to provide:

An optimal PEEP level recommendation for the patient The ability to detect abnormal ventilation parameters Three possible diagnostic suggestions Three proposed treatment approaches The data provided by the AI models will be recorded and compared with manual assessments performed by experienced ICU specialists. This comparison aims to determine the extent to which AI models can serve as clinical decision support tools. The reliability, accuracy, and clinical applicability of AI-generated recommendations will be thoroughly analyzed.

In the final phase of the study, the PEEP levels determined by AI models will be compared with those identified by ICU specialists with at least five years of experience. The following criteria will be evaluated:

Agreement between AI models and expert opinions on PEEP levels (primary outcome measure) Sensitivity and specificity of AI in detecting abnormal ventilation parameters Accuracy of AI-generated diagnoses Clinical applicability of AI-recommended treatment approaches

Study Timeline

• Status Verified: 2025-02
• Study First Submitted: 2025-02-20
• Study First Submitted QC: 2025-02-20
• Study Start: 2025-02-24
• Study First Posted: 2025-02-25
• Last Update Submitted: 2025-02-26
• Last Update Posted: 2025-02-28
• Primary Completion: 2026-02-01
• Study Completion: 2026-02-02

Recruitment & Eligibility

• Status: NOT_YET_RECRUITING
• Sex: All
• Target Recruitment: 145

Inclusion Criteria

• Patients aged 18 years or older (adult patient group).
• Patients requiring mechanical ventilation in the intensive care unit (ICU).
• Hemodynamically stable patients with stable blood pressure and heart rate.
• Patients with complete medical records, including arterial blood gas values and ventilation parameters.
• Patients whose legal representatives (if applicable) have provided informed consent for study participation.

Exclusion Criteria

• Patients receiving extracorporeal membrane oxygenation (ECMO).
• Patients with severe hemodynamic instability, such as refractory hypotension or arrhythmias requiring continuous vasopressor support.
• Patients with incomplete medical records, particularly those missing critical data on ventilation parameters or arterial blood gas analysis.
• Patients or their legal representatives who decline participation in the study.

Study & Design

• Study Type: OBSERVATIONAL
• Study Design: Not specified

Primary Outcome Measures

Name	Time	Method
Agreement Between AI-Generated and Expert-Determined PEEP Levels	7 day	The primary outcome will be the agreement between the PEEP levels recommended by AI models (ChatGPT, DeepSeek, Gemini) and those determined manually by ICU specialists. Agreement will be assessed using Bland-Altman analysis and intraclass correlation coefficient (ICC) to measure consistency

Secondary Outcome Measures

Name	Time	Method
Sensitivity and Specificity of AI Models in Detecting Abnormal Ventilation Parameters	7 days	AI models will be evaluated for their ability to correctly identify abnormal ventilation parameters (e.g., auto-PEEP, poor compliance, overdistension) compared to expert evaluations. Sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) will be calculated.
Accuracy of AI-Generated Diagnostic Predictions	7 days	The accuracy of diagnostic suggestions made by AI models will be compared with ICU specialists' clinical assessments. Agreement will be measured using Cohen's kappa statistic.
Clinical Applicability of AI-Generated Treatment Recommendations	Within the first 24 hours of mechanical ventilation.	The feasibility and clinical relevance of treatment recommendations provided by AI models will be assessed by ICU specialists using a 5-point Likert scale (1 = Not useful, 5 = Highly useful).

Trial Locations

Locations (1)

Health Science University İstanbul Kanuni Sultan Süleyman Education and Training Hospital; 🇹🇷 Istanbul, Turkey