It is unclear what you are looking for when you ask "which model is best". All three models are telling you the same thing.
Saying "the lower the IDe value, the best the model is" or "the higher the R^2 value is, the best the model is", is not accurate.
- The IDe within a formulation represents a bias point on the graph for a given subject.
- A higher regression value within a formulation represents the criterion. In the case of Fitts', this is MT.
A particular model may be more predictive for your situation. The way to decide that is to run more studies and compare the results to your regression.
You can't just point to one formulation and say "that's best".
Fitts' model is the original and has been shown to apply under a number of conditions.
The Welford model is a 2-factor variation, which takes target distance and target width as influencers, that can not be directly compared to Fitts' 1-factor equation. You can't plot Fitts' and Welford on the same graph and simply compare the two lines. If you really did calculate width as a seperate influence, looking at your data I'm not sure this was done, Welford's model generally provide greater predictive power.
The Shannon model is another variation. It has arguably become more popular when working in human-computer interaction. At least in part that ISO 9241 recommends its use. So, if you're looking for the "offical" internation standard variation, use Shannon's.