Is this a proper way to find the derivate of a 12-bit ADC?

Others talked about the timing issues, so I will focus on removing the noise from the measurements as it could be large. Basic principle is to approximate a derivative with more points than two. One of the simplest technique is called five point stencil which can be easily used on most MCUs. It uses 4 points to calculate the first derivative as: $$f'(x)approxfrac{-f(x+2h)+8f(x+h)-8f(x-h)+f(x-2h)}{12h}$$ Where f(x) are the ADC readings and h is the time difference between two readings. This technique can be generalized with finite difference coefficients. Another possibility would be to filter your ADC samples before calculating the derivative, for example with Gaussian filter.

The RTOS and timing catches having been discussed by others, the numerical differentiation greatly amplifies quantization noise for medium- and low-changing signals. Depending on a nature of signals under consideration, there exist well established numerical techniques to deal with this amplification: oversampling, interpolation, Kalman filtering etc.

An article on an oversampling technique application compares a dependence of quantization noise amplification on signal change rate in Figure 2 (see below).

(a) Piecewise quadratic acceleration profile. (b) Theoretical and finite-difference derivatives.]1

You may see patterns relevant to your observation. If this is the case, you can try oversampling. But don't forget about other techniques!

Your understanding of the concept looks basically correct but there are a couple of issues.

Unless your sampling time is variable, you should not need to read the time. You should already know the sampling interval beforehand

If the sampling time is variable, you need to record the time associated with when the actual sample was taken. Your current code doesn't do that. It records the time completely independently of the sample some time after the sample has already been taken. If you are behind an RTOS there is probably a much better way to do it like setting the ADC to automatically sample and automatically writie the samples into a buffer so you can read the samples and differentiate whenever you want knowing that the interval time is constant.

Yes, noise will be an issue. You want to filter the ADC readings before differentiating since it is very sensitive to noise. Even a small magnitude change over a small time interval (such as when sampling at a very high frequency) results in enormous slopes that may completely swamp your graph. I've found the easiest way to deal with this is to actually reduce the sampling rate which reduces the slope of the line between adjacent samples.