Splitting Subject Data in train, validation and test set for 3D Human Pose Estimation for better accuracy

This is a 3D Human Pose Estimation problem.

There are totally 15 normal subjects in train set, 7 normal subjects in validation set and 7 normal subjects in test set.

There are 7 impaired subjects totally. This is a density plot without adding those impaired subjects where train, validation and test split density, count and proportion plots can be observed. The density plots show the distribution of subjects inside train, validation and test sets, while the count shows the number of samples given inside each set and the third graph is an ECDF plot which shows the proportion of each of these sets.

How can the splitting of these impaired subjects be done to get a higher accuracy by adding these in train, validation and test set respectively?