The intensive care unit (ICU) generates a vast and complex array of patient data, posing challenges for clinicians in identifying meaningful patterns and making timely decisions. This study investigates the impact of different feature selection methods on the performance of machine learning predictive models for patient outcome prediction in the ICU. Five distinct experiments were conducted, each employing a different feature selection strategy: (1) a baseline model with all features, (2) SHAP-based feature selection with a threshold of 0, (3) SHAP-based feature selection with a threshold of 0.5, (4) feature importance-based selection using XGBoost, and (5) feature importance-based selection using a Decision Tree classifier. Model performance was assessed using the F1 score. Our findings reveal that feature selection significantly influences model performance. Our results highlight the potential of both SHAP-based and feature importance-based methods for identifying the most impactful features. Further research with larger datasets and diverse clinical scenarios is warranted to validate these findings and refine feature selection strategies for optimal patient care.