The large number of hospitalizations and surgeries caused by the high incidence of cardiac disease makes patient monitoring essential to ensure a safe recovery and reduce the risk of readmission. Smartwatches can be powerful, non-invasive allies in monitoring cardiac patients due to their potential for early intervention, thereby reducing the need for emergency hospitalizations. Using physiological signals including blood pressure (SBP and DBP), heart rate (HR), and oxygen saturation (SpO2), collected by the smartwatch, along with each patient's medical history, a machine learning model was built for the binary classification of patients based on Emergency Room (ER) readmission. We used the Uniform Manifold Approximation and Projection (UMAP) algorithm for feature extraction and a Random Forest (RF) classifier for this task. The model combining SBP, DBP, and SpO2 signals yielded the best performance, achieving a mean balanced accuracy of 76.89 ± 7.77% and a mean AUC of 77 ± 7.77%. Despite this, the model struggled to correctly identify the positive class and showed significant performance variance between cross-validation folds, with F1-scores ranging from 40 to 80%. While the findings demonstrate the potential of this approach, the significant performance variability among folds shows the need for larger datasets to improve model robustness for clinical application.