Classification algorithms with unbalanced datasets tend to produce high predictive accuracy over the majority class, but poor predictive accuracy over the minority class. This problem is very common in biomedical data mining. This paper introduces a Support Vector Machine (SVM)-based optimized feature selection method, to select the most relevant features and maintain an accurate and well-balanced sensitivity-specificity result between unbalanced groups. A new metric called the balance index (B) is defined to implement this optimization. The balance index measures the difference between the misclassified data within each class. The proposed optimized feature selection is applied to the classification of patients' weaning trials from mechanical ventilation: patients with successful trials who were able to maintain spontaneous breathing after 48 h and patients who failed to maintain spontaneous breathing and were reconnected to mechanical ventilation after 30 min. Patients are characterized through cardiac and respiratory signals, applying joint symbolic dynamic (JSD) analysis to cardiac interbeat and breath durations. First, the most suitable parameters (C+,C-,σ) are selected to define the appropriate SVM. Then, the feature selection process is carried out with this SVM, to maintain B lower than 40%. The best result is obtained using 6 features with an accuracy of 80%, a B of 18.64%, a sensitivity of 74.36% and a specificity of 82.42%.