Checking convergence based on the probability of predictions

I'm trying to implement Bing Liu's I-EM algorithm for PU learning (binary classifier learning from positive and unlabeled examples only) in scikit-learn. As usual with EM instances, I need a stopping criterion. I would, however, like to implement this algorithm in a very general way, without needing to inspect classifier parameters.

Instead, I would like to check convergence by inspecting the probability outputs from my classifiers (from the scikit-learn predict_proba interface) on my training set, but I don't know if this is valid. Can someone who has more knowledge of EM algorithms shed some light on this?