Using averaging in firmware has made the readings less noisy (see graph after 19th Sept) , so I am going to switch to the 'Averaging of N-X ADC samples' technique described in the application note, to see if it results in a further improvement.