Doing statistical tests on populations with 3 values is not a very sensible approach, mathematically you can obtain these metrics, but you should have more points to make these really representative. The same will happen with Machine Learning.
Can you obtain more samples? That would be the best and then repeat your t-tests or whatever you are using.
Alternatively, better to display your measurements and try to interpret than to summarise with a metric that may not be representative.
