So, some kind of optoisolator driven, or feedback, amp?
Rest assured they've tried everything already -- one of the more peculiar examples that comes to mind was an HP electrometer, I think: they used varactor diodes as a parametric sense element, excited by RF, coupled by capacitors, so the DC leakage is extremely low. When the diodes are biased by a small applied (DC) voltage, the capacitance changes, sensed by a bridge circuit or whatever. The AC balance (or even harmonic distortion, I forget) is amplified and detected, as you would with any chop amp -- giving extremely high loop gains as well as extremely low leakage. This would've been early 60s I think, before JFETs arrived, or at least were very good?
Anything you can transmit and amplify, has been used for amplification or control. Modern laser apparatus I suppose is an excellent example, some things working down to quantum levels of energy, or particle counts.
It sounds like you'd be interested in analog isolation techniques?
There are two major problems with a traditional LED + photodiode or phototransistor pair:
1. The CTR is unreliable, over ranges of manufacture (device parameters, alignment, packaging), aging (LEDs lose some efficiency with use), and the usual current and temperature (giving nonlinear, temp- and history-dependent responses). As lone or feed-forward devices, they're terrible. As feedback devices, they can be acceptable.
A typical use case is feedback from a secondary-side error amplifier, to a primary side PWM modulator: the isolator gain and offset is corrected by the loop.
Another case is the IL300 and relatives, where the emitter gain isn't well known (see above), nor the ratio between two receivers (due to manufacturing variation), but that ratio is stable enough to be useful. The emitter's output is servoed to get a desired current in one receiver, and the other receiver will deliver a proportional current. These require calibration to use accurately, but are stable otherwise.
2. Bandwidth. Optos are slow. The photodiode/transistor junction is large (high capacitance), and in the case of the transistor, has slow recombination too. The capacitance can be reduced somewhat by reverse-biasing it (diode capacitance varies with voltage, being highest at zero bias), and indeed such circuits are available, 6N136 for instance. Still better performance can be had with a smaller photodiode, integrated amplifier (probably some transresistance amp thing -- this forces the AC voltage on the diode to be as small as possible, minimizing the effect of capacitance), stronger drive, maybe some kind of lens in the optical path... LEDs themselves roll off in the low 10s MHz, and to go much higher you need lasers.
So then we get into fiber optics: lasers, photodiodes, extremely high radiance; RF circuits and controlled impedances, etc. Gain is all over the place, the optical path might vary 20dB due to alignment, cable length and so on. Don't have to worry about that so much in a short path isolator of course, though there are applications where this is important -- optically triggered gate drive circuits for HVDC power inverters, for example.
Anyway, all the gain and offset stuff can be solved with a modulation scheme. Typical cases are AM with AGC, FM with ALC, or various more complex RF modulations. It's still not easy to get raw baseband analog through these: AGC needs a known baseline signal level, so the signal needs to be momentarily keyed to 0/100% to set that level (analog TV did it this way!), or AC coupled so the average is consistent, etc. FM is a bit better, but your receiver's frequency reference isn't going to be perfectly matched; it might be nice to track the signal with a PLL, but then you once again have to AC couple it, or detect coded levels or something.
And meanwhile, all the propagation delays and filtering and detection and decoding, either delays your signal, or reduces its bandwidth or noise floor. Which puts you at a disadvantage, when used in a feedback loop.
This is just one example, concerning a few common types of optoisolator components; but it's illustrative of other methods. Anything that has a wave mechanism and a transducer, applies!
Electric-electric transducers (i.e., electromagnetic coupling) work from DC (sensing electric or magnetic field) to AC to beyond-light: at low frequencies, we have cored transformers; at high frequency, air core transformers and antennas; at optical frequencies, photodiodes, mixer crystals, and weird quantum stuff; at high energies, scintillators (downconverting the problem to an already solved optical problem, heh) and various particle detectors. (Subatomic particles in turn being another wave we can work with, though with much less ease outside the vacuum chamber!)
- Acoustic transducers ranging from motors ("DC") to loudspeakers, piezos, and weird physics (phonon) stuff. Self-contained cases include the "piezo transformer", more of an acoustic resonant network with electrical terminals.
- You could even do uh, like, gravity waves on the surface of water... not that you'd have any meaningful bandwidth.
- Or gravitational waves, which we can receive right now, but not that we'll have any meaningful way to generate them until we're a few levels up the Kardashev scale...

So -- hopefully this has given you some ideas as far as directions to continue researching. No idea if all the possibilities have been made into patents, or any kind of publication even. There are surely some ideas just too bad to bother enumerating, but there are a lot of combinations; who knows if your idea is among them or not.
As for commercialization, I suspect you'll have a hard time beating plain old monolithic or hybrid circuits; and photonic circuits are, well they've been in experiments for years, but I don't think anyone's yet commercialized a CPU with optical paths onboard or whatever? The main reason, I suspect, APEX amps generally have poor specs, is they just don't need to be much better? They're pretty specialty, and mainly used for piezo drivers. There are, or have been, a few specialty applications for higher voltages, and high bandwidth -- CRT drivers were a great example, using several 1GHz+ 100V transistors, or ICs to the same effect. But with CRTs obsoleted, you don't see much of them anymore; for that matter, the high current pulsed types that plasma TVs used, either.
There are relatively few places a discrete circuit has any advantage left over modern components, even when the fit is relatively poor (i.e. requires several chips to implement). Not that they can't be found; I claim to
have one myself, not that it's a very important thing, and it's certainly not compact at that. It could be beaten by an IC performing the same function -- handily, no doubt about it. It just so happens that nothing available (at least, last I checked) can implement the same function while consuming less overall supply current.
Tim