I have a particular problem and I am wondering how to frame it and which NN architectures to try.
I have a biological sensor which can detect amino acid (glutamine) concentrations in animal plasma. There is a gold standard for measuring plasma glutamine, and so the aim is work out how best to calibrate this sensor so that the difference between the gold standard and the sensor value is small.
I have additional data for the sensor. This is ex vivo data (in the lab) whereby the sensor is exposed to increasing concentrations of glutamine in a step-wise fashion, and the sensor response is measured. Normally, we have used a simple linear relationship between steady state sensor response and glutamine concentration during this ex vivo testing to calibrate the sensor in vivo. However, the raw sensor data from the ex vivo testing is quite rich (e.g. time constants of response times - data which is not utilized in sensor calibration). There is a lot of sensor-to-sensor variability and the ex vivo data for sensor testing contains 3600 data points (a time series for 1 hour at 1/s).
So my question is: how best to formulate this problem as a ML task? I am not even sure what the target should be? My feeling is that it should somehow be an objective function that minimises the difference between the gold standard and the sensor current. But what should I put as the input data? What neural structure should I use?
Any thoughts on this would be really appreciated!