Hello,
nice topic!!
as I see it, there are two points here: a. That the result of a scan is not the same as the label of the problem and b. There is no consistency on the measurement of the same sample.
The first is easier:
– Generally the labels of the products give “average values”. That means that the took some samples, measured them and then produced the label. Especially on cheese, I know that the number of samples is quite small, and also the label is printed once a year (or even for more years). So, even if you took the sample to a laboratory, the value probably wont be the same as the label. It is just to give you an idea. This creates also a problem with our sampling. When we collect the samples and give the value that the label says (that probably isn’t a exact value), the model we make isn’t that good. On the other hand, if we collect many many samples, the model will improve (because one label will be higher than the real value, the other will be lower, and they will “correct each other”).
About the inconsistency:
– I have also noticed what you mention (different scans with the scio and the sample not moved), but it was with the spectrum. I scanned a liquid (I think it was milk), and the spectrum was different on the right end. After that I noticed that generally the far right end of the spectrum is not so consistence. I can’t imagine what it may be, but because it is toward the end, I think it is (more or less) ok.
– Now, about the different values that the test gave you… The only thing I can think is the difference of temperature. The collection we have has only one sample at 2oC and then all the other samples 20-24oC. What was the temperature of your sample? I have noticed that the spectrum of the same sample changes in different temperatures. I don’t know if the “Preprocessing Methods (Performs normalization of the signal. This is meant to compensate for changing measurement conditions)” can take that into consideration, but I believe that a correct model should have samples with different temperatures, and the one we are given, doesn’t.
But also, I don’t understand exactly the results of the measurements. I suppose you don’t mean that between the measurements there was a 1% difference (the first was 20% fat and the second was increased by 1%, that means went to 20.2%). Because this is a very small and normal difference. If you mean that the first measurement was 20% and the second was 21%. Again it depends on the first value, if it was 3% and went to 4%, it is a very big deal, because it is an increase of 33%. If from 25% went to 26% I believe it is more acceptable 4% increase (taking into consideration also that the model is not the best given).
I believe that in order to “solve” these questions, we need two things. First, the percentage of confidence for a measurement that the team is working to add to the software. It is necessary for any work. The second is for us to perform tests of the model in lab conditions, with known standards of high purity. When we receive the “SCiO liquid accessory”, I am going to create standards of glucose and BSA protein, and perform detailed measurements.
Also I would like to mention that many laboratory equipment, designed for specific measurements, also perform (automatically) more than one measurements, and then give as a result a mean value. We can not expect from anything a 100% “correct” result.