Rationale: Machine learning has been illustrated in various medical fields, such as fluid management in sepsis, prediction of renal failure and others. The weaning period is the key to the management of a patient on mechanical ventilation. Weaning can take up to half the time spent in intensive care. Up to 20% of patients do not pass their first withdrawal test. However, mortality can reach 38% in patients with the most difficult weaning. [1] Only a small number of studies have looked at the application of machine learning to the weaning process. Being able to predict the success of spontaneous breathing test and extubation is essential to improve morbi-mortality and reduce length of stay. We propose to develop a predictive algorithm for the success of a weaning test and identify the different factors that determine this success. Patients and methods/materials and methods: It is a critical care, single-centre and retrospective study. Most of the variables are taken from the literature and extracted from the computerized patient record. We designed several machine learning algorithms: Logistic Regression, Random Forest Classifier, Support Vector Classifier (SVC), K-Nearest Neighbors algorithm (KNN), XGBoost, and Light Gradient Boosting Method (LGBM). In order to maximize the accuracy of the prediction and to reduce the risk of overfitting, we computed different methods: multiple imputation with K-Nearest Neighbors for missing data, the SMOTE method (Synthetic Minority Oversampling technique) and K-fold cross validation. The results are expressed in terms of AUC, and the factors’ importance is determined by the SHAP (Shapley added explanations) method. Results: Our cohort included 80 patients (60 successful, 20 unsuccessful). The best algorithm was LGBM with AUC 0.873 (F1score: 0.87, AUC-pr: 0.95, accuracy: 0.813), Figure 1. The other algorithms had a worse predictive performance (Random Forest: 0.727, Logistic Regression: 0.8, XGBoost: 0.836). Weight gain since admission, presence of VAP, protidemia and fibrinogenemia at the time of testing, and duration of invasive ventilation were considered to have the greatest impact on the model. Discussion: Despite the small number of patients, all the techniques in place allow good prediction and reproducibility. However, the cohort will continue to grow until the congress (to a minimum of 250 patients). The main determinants affecting the success of the test are similar to those found in other studies in this field and are factors that can be directly influenced. Conclusion: Machine learning will be useful in predicting weaning success and will be a real help in daily clinical practice.