This paper proposes a novel message embedding solution based on modifying the split decisions of HEVC videos. The encoder starts by computing a mapping between the split decisions of a Coding Unit (CU) and its features variables. This results in model weights that can be used to predict the split decisions. Message embedding is then carried out as a function of the predicted and true split decisions per CU. If the message bit to embed is '1' and the predicted and the true split decisions are different then the true split decision is modified to be identical to the predicted flag. Otherwise if the message bit is '0' and the predicted and the true split decisions are identical, then the true split flag is modified to become different than the predicted flag. We apply this embedding concept at two CU coding levels; 32x32 and 16x16. To extract a message, the decoder starts by regenerating the model weights which are then used to predict split decisions and compare them against the true decisions received in the bit stream. Identical decisions indicate a message bit of '1' and nonidentical split decisions indicate a message bit of '0'. In the experimental results we examine the prediction accuracy, the effect of the proposed solutions on message payload, the number of modified split decisions and the corresponding impact on video quality. Comparison with an existing solution reveals that the proposed solution is superior in terms in message payload whilst resulting in reduced video distortions.