This paper addresses gait deviation change predic- tion for patients with gait disorders. Two cases are predicted and referred to as two classes: the patient’s gait improves (gait deviation from normative data is reduced) within the next clinical gait analysis (CGA) session (class 1) or not (class 0). To cope with available kinematic data obtained from CGA, a two- dimensional fast Fourier transform (2D FFT) is used to build 2D frequency images. After that, fine-tuned VGG16, ResNet34 and EfficientNet_b0 are applied to these images separately. These pre-trained models showed an over-fitting behaviour. To overcome this issue, a 2D convolutional neural network (2D CNN), trained from strach and with low computational footprint, is proposed. Ourmodel attains an area under the curve (AUC) above 0.69, which is promising for such a complex task. The achieved results demonstrate the efficiency of deep learning models and are the first of their kind regarding the state-of-the-art.