Hi João,
You can still use fitcdiscr even if your features do not follow a Gaussian distribution. It is assumed that the data for each class are drawn from a multivariate Gaussian distribution and not a requirement of the data. Though it is important to note that fitcdiscr implements Linear Discriminant Analysis (LDA) and Quadratic Discriminant Analysis (QDA), both of which assume that the data for each class are drawn from a multivariate Gaussian distribution.
However, fitcdiscr can still be used. The classifier might not perform optimally if this assumption is violated, but in practice, LDA and QDA can still work reasonably well even if the data are not perfectly Gaussian.
If the data significantly deviates from a Gaussian distribution, you might want to transform your data to make it more Gaussian-like. You can also consider using a different classification algorithm that does not make this assumption like Support Vector Machines (fitcsvm and fitcecoc), Decision Trees (fitctree), ensemble methods (fitcensemble) and other such algorithms.
The choice of classification algorithm depends on your specific use case and the nature of your data. It is often a good idea to try several different algorithms and see which one works best for your problem.
Hope this helps!
