Prediction of Hospital Discharge, ICU Discharge and Death Event Using AI: A Pilot Study

Background: Machine learning has shown promise in predicting clinical outcomes in critically ill patients. Respiratory data obtained during mechanical ventilation (SpO₂, respiratory rate, PaCO₂) may play a critical role in these predictive models. Methods: Data from 100,000 patients were analyzed from the MIMIC-IV dataset (Moukheiber,et al., 2025. A Temporal Dataset for Respiratory Support in Critically Ill Patients [version 1.1.0]. PhysioNet. RRID:SCR_007345. https://doi.org/10.13026/wewp-sj67 ). Decision Tree and Random Forest algorithms were applied together with Particle Swarm Optimization. In total, 53 parameters were evaluated. All machine learning codes were developed and implemented in Python v3.10. Results: The Random Forest model achieved very high accuracy in predicting hospital discharge (RF = .97), while Decision Tree performed more modestly (DT = .76). For ICU discharge, prediction accuracy was lower overall, though Random Forest (RF = .64) outperformed Decision Tree (DT = .50). Mortality prediction showed stronger results, with Random Forest performing best (RF = .98; DT = .75). Feature importance analysis revealed that mechanical ventilation parameters such as SpO₂ (15–17%) and respiratory rate (8%) were key predictors, alongside ICU length of stay (21–24%) and vasopressor use (8–18%). In both models, race also emerged as a notable factor (DT 12%; RF 7%). Conclusion: Machine learning analyses, including mechanical ventilation data, enable respiratory monitoring and early decision support for clinicians and physiotherapists. This approach may enhance the safety and effectiveness of respiratory rehabilitation as a strategic tool in clinical management.