Using Artificial Intelligence Models for Predicting The Need for Intubation and Successful Weaning From Mechanical Ventilation in ICU Patients

Intubation is a medical procedure that involves inserting a breathing tube into a patient's airway to prevent suffocation and provide assistance with breathing. This intervention is commonly performed in the Intensive Care Unit (ICU) when patients are unable to clear secretions or breathe independently, Intubation is particularly beneficial for patients with conditions such as emphysema, collapsed lungs, pneumonia and pulmonary edema, allowing them to breathe effectively, However, accurate prediction of the timing of intubation remains an unsolved challenge due to the noisy, heterogeneous, and unbalanced nature of ICU data. So, determining the optimal timing for intubation is a stressful task for medical personnel, as delays can worsen patient outcomes and increase morbidity and mortality.

Also, estimating an appropriate weaning time from mechanical ventilation is an essential clinical decision in critical care. Premature attempts to extubate patients increase the risk of ventilator-associated pneumonia, prolonged intensive care unit (ICU) stay, and mortality. Meanwhile, an unnecessarily prolonged duration of mechanical ventilation causes an enormous economic health burden and is associated with deteriorated clinical outcomes .

Previously proposed weaning indices have shown conflicting results, as over one-quarter of patients require reintubation despite meeting the criteria for such indices, such as the rapid shallow breathing index (RSBI) . Therefore, an accurate prediction tool for deciding when patients are ready for extubation is critical for managing patients with respiratory failure.

Artificial intelligence models have emerged as promising solutions for tackling these challenges. These technologies have the potential to learn from data and predict the optimal intubation timing. Unlike medical professionals, machine learning models are not affected by personal emotions.

Regarding intubation prediction; AI models incorporating time-series analysis of respiratory rate variability, oxygen saturation trends, and evolving PaO2/FiO2 ratios can predict impending respiratory failure with greater accuracy than conventional clinical assessment. Also, regarding weaning decision support; Machine learning algorithms analyzing integrated parameters during spontaneous breathing trials demonstrate superior prediction of extubation success compared to traditional weaning indices. Recent studies show AI-guided protocols can reduce ventilator days by 1.8 days (95% CI 0.5-3.1) without increasing reintubation rates.

These systems function as cognitive aids, processing multidimensional data at a scale beyond human capacity while preserving clinician judgment. Once the model has done enough learning, the trained model can be easily deployed with low costs at various ICUs.

This study aims to using artificial intelligence (AI) approach to build a model to determine the optimal timing of intubation and optimal timing of weaning from MV for ICU patients, and compare outcomes with these which depend on clinicians' decision. Then assess whether AI-assisted decisions improve patient outcomes (e.g., reduced intubation delays, shorter ventilation duration).

Study tools include collecting patients data and AI model development;

• Data collection:

All participates will be subjected to

• Baseline data: Demographics ( age, sex, weight, length) , comorbidities (DM, HTNs ,chronic kidney disease… ), APACHE II score, reason for mechanical ventilation.
• Daily data:

• Resolution of the underlying condition: The primary reason for mechanical ventilation should be resolved or significantly improved.
• Hemodynamic stability:

blood pressure and heart rate without significant vasopressor support. c- oxygenation: - PaO₂ ≥ 60 mmHg on FiO₂ ≤ 0.4. - PEEP ≤ 5-8 cm H₂O. - Oxygen saturation (SpO₂) ≥ 90%. d- ventilation: - PaCO₂ within normal limits or at the patient's baseline. - pH ≥ 7.25.

• Normal electrolyte levels: K, Mg, phosphate

• fluid balance.
• Normal metabolic status: No significant acidosis or alkalosis.
• hemoglobin levels: > 7 g/dL

e. Respiratory Mechanics

• (RR): < 35 breaths/min.
• Tidal volume : > 5 mL/kg ideal body weight.
• Minute ventilation (VE): < 10-15 L/min.
• Rapid shallow breathing index : < 105 breaths/min/L (measured during a spontaneous breathing trial).
• Endotracheal cuff leak test. f. Neurological Status
• mental status: Patient should be awake, alert, and able to protect their airway.
• Cough and gag reflex
• Minimal sedation g. Others
• Infection control: No active infection or fever that could impair weaning.
• Pain management 3- Outcome data: Duration of ventilation, ICU length of stay, complications, reintubation rates.

b.AI model development :

• model choice
• model training and validation; Split Data 70% training, 15% validation, 15% testing
• performance metrics , AUC-ROC, sensitivity , specificity, F1-score, calibration (Brier score).

Inclusion criteria

• Adult patients (≥18 years) in ICUs at risk of respiratory failure.
• Adults (≥18 years) receiving mechanical ventilation for >24 hours and deemed ready for weaning.

Exclusion criteria

• - Patients with a pre-existing "Do Not Intubate" (DNI) order.
• - Patients with irreversible neurological conditions affecting weaning