I would expect it to be a straightforward signal processing problem if all that's needed is 20dB of suppression across a relatively narrow frequency range. If

I would expect it to be a straightforward signal processing problem if all that's needed is 20dB of suppression across a relatively narrow frequency range. If the frequency response is consistent from MEMS to MEMS, it should be straightforward to simply create a model, invert it, and apply it to the incoming waveform. DSPs these days can be pretty tiny and power efficient so I bet it wouldn't even take up too much power and area to do something like that.