Telemonitoring of cardiac patients, particularly in the postoperative period, can be improved through Machine Learning architectures that enable remote clinical supervision. Dimensionality reduction models such as Autoencoders (AEs) are suitable for processing cardiological data in this context. In this study, vital signals collected from smartwatches, such as Heart Rate, Blood Pressure (SBP and DBP), and Peripheral Oxygen Saturation (SpO2), along with patient history, were used as input for a supervised AE-based binary classification model to predict hospital readmission. Data from 49 postoperative cardiac patients were collected over 30 ± 3 days, with a 9/49 readmission rate. After preprocessing, the data were passed through an AE architecture with dense layers, batch normalization, dropout, and a latent space. A total of 63 input combinations were evaluated across latent space dimensions of 8, 12, 16, 20, and 24. The model classified patients as readmitted or not. Results show that including patient history significantly improves prediction. Cross-validation revealed the best performance for the SBP + History input, with an average F1-score of 83.41 ± 3.22%. These findings highlight the model's potential, although further architectural optimization and larger datasets are needed to ensure robustness and clinical applicability.