We present an approach using a combination of coupled channel scattering calculations with a machine- learning technique based on Gaussian Process regression to determine the sensitivity of the rate constants for non-adiabatic transitions in inelastic atomic collisions to variations of the underlying adiabatic interaction potentials. Using this approach, we improve the previous computations of the rate constants for the fine-structure transitions in collisions of O(3Pj) with atomic H. We compute the error bars of the rate constants corresponding to 20 % variations of the ab initio potentials and show that this method can be used to determine which of the individual adiabatic potentials are more or less important for the outcome of different fine-structure changing collisions.