My understanding is that Dual ISO works by combining information coming from one row of pixels @low ISO and one @high ISO.
So in the midtones, both rows of pixels carry useful information. More specifically, on cameras equipped with ISO-invariant sensors (like all Sony ILC's and a very few Canons like 80D AFAIK) they do carry the *same* information actually (i.e. you can pump up low ISO image exposure and obtain same image as out-of-camera high ISO image, without additional noise). So no loss of details in this case. On the contrary with 5D3 / M etc (which are very far from ISO-invariant cameras), higher ISO always has lower SNR (Signal to Noise Ratio) in the midtones: so this means a *slight* loss of detail in the midtones when using dual ISO.
In the deep shadows, low ISO is almost signal-less, and pumping it up will result in extreme noise. High ISO has noise of course, but a way higher SNR. So here dual ISO beats single ISO by far (as long as we are considering the case where single ISO = low ISO of course), as it will exploit way higher SNR for half of pixel rows (the ones @high ISO).
Exactly the opposite happens in the extreme highlights, where high ISO rows will be clipped (completely useless) and only rows @low ISO will be used to reconstruct the image => so in this case Dual ISO mode uses *half pixels* with respect to single (low) ISO.
Of course the 'de-interlacing' algorithm applied to reconstruct the image from Dual ISO mode is smart enough to try and exploit all available information capture on sensor. However, I think about Dual ISO as a small compromise on resolution to obtain a huge boost in dynamic range.
Sergio