Nowadays, ontologies and machine learning constitute two major technologies for domainspecific knowledge extraction. While the aim of these two technologies is the same -- the extraction of useful knowledge -- little is known about how the two sources of knowledge can be integrated. This problem is especially important for biomedicine where relevant data are often naturally complex having large dimensionality and including heterogeneous features. In this paper we propose an approach for improving the performance of machine learning by integrating the knowledge provided by ontologies for large-dimensional disease-specific heterogeneous data. The basic idea is to redefine the concept of similarity by incorporating available ontological knowledge. Benefits and difficulties of this integration are discussed and an example from the field of paediatric cardiology is described.