Anomaly detection for time series data with only positive samples?

If you want to model it as a regression problem, using only the normal ECG data sounds enough to me. In that setting, abnormal data would be only used for validation of your model, but even those few negative samples that you have could be valid for that.

You have a variety of options for getting that divergence, e.g. the Mahalanobis-Taguchi System. You can also model the error coming from your regression model or go fully Bayesian with something like Gaussian Processes, so that you already have a distribution for your posterior. That way, you can cross check new negative samples.