Data quality is crucial for building reliable machine learning models. To ensure the robustness of machine learning models, it is essential to assess the fidelity of generated tabular data, that is, their ability to replicate the distributions, correlations, and predictive patterns of the original data, but unified indexes for this comprehensive assessment are currently lacking. This work proposes “TabDSFidelity,” a new unified quality index for assessing the fidelity of such data in classification tasks. TabDSFidelity integrates classifier accuracy, distribution similarity, and correlation preservation into a single, adaptive, weighted score. Validation used ten real-world datasets from an Intensive Care Unit. Balanced sets were generated through oversampling and the synthetic large-scale datasets using SMOTE RSB* Adapted with Gaussian Noise and CTGAN, with optimal datasets selected at each stage using TabDSFidelity. Experimental results using various classifiers consistently demonstrate that models trained on the data selected by TabDSFidelity significantly outperform those trained on the original data. Furthermore, a significant positive correlation was found between TabDSFidelity scores and the actual accuracy of the models. In conclusion, TabDSFidelity presents itself as an effective and adaptable tool for objectively assessing the fidelity of generated tabular data, facilitating the informed selection of datasets that improve the predictive effectiveness of classification models.