One add-on comment.

You might want to define your protocol such that every 32 or so readings, it transfers a full reading to re-center the subsequent relative readings. This will keep your system from creeping off center if line noise drops a few bits now and then.

I'd recommend that you use a spreadsheet to figure the best balance of full readings to relative readings and the span of relative readings to get the best mix of compression. The latter will have to do with the dynamics of your data and the sample rate.

aka j