Goal: Machine learning (ML) technologies that leverage large-scale patient data are promising tools predicting disease evolution in individual patients. However, the limited generalizability of ML models developed on single-center datasets, and their unproven performance in real-world settings, remain significant constraints to their widespread adoption in clinical practice. One approach to tackle this issue is to base learning on large multi-center datasets. However, such heterogeneous datasets can introduce further biases driven by data origin, as data structures and patient cohorts may differ between hospitals. Methods: In this paper, we demonstrate how mechanistic virtual patient (VP) modeling can be used to capture specific features of patients' states and dynamics, while reducing biases introduced by heterogeneous datasets. We show how VP modeling can be used for data augmentation through identification of individualized model parameters approximating disease states of patients with suspected acute respiratory distress syndrome (ARDS) from observational data of mixed origin. We compare the results of an unsupervised learning method (clustering) in two cases: where the learning is based on original patient data and on data derived in the matching procedure of the VP model to real patient data. Results: More robust cluster configurations were observed in clustering using the model-derived data. VP model-based clustering also reduced biases introduced by the inclusion of data from different hospitals and was able to discover an additional cluster with significant ARDS enrichment. Conclusions: Our results indicate that mechanistic VP modeling can be used to significantly reduce biases introduced by learning from heterogeneous datasets and to allow improved discovery of patient cohorts driven exclusively by medical conditions.