# autoencoding spiky time series - better loss function?

I am experimenting with convolutional autoencoders for time series. My first network architectures work quite well. However, the autoencoder has a tendency to soften the spikes in the time series. And on this electrocardigram data the spikes are essential features.

(blue: original time series, orange: reconstruction, red: error)

I have used mean squared error as a loss function so far, and perhaps another loss function could be the solution. Any recommendations?