In tokamaks, rotating magneto-hydro-dynamic modes frequently decelerate as their amplitude increases. Once a critical threshold in amplitude is reached, these modes stop rotating into a specific toroidal and poloidal position and are commonly named Locked Modes (LMs). The presence of LMs, especially with low toroidal mode numbers, causes degradation of plasma performance, i.e. particle and energy losses, and can lead to a plasma disruption. Several strategies can be adopted when designing the plasma scenario to avoid the onset of these modes, which foresee the use of plasma heating, current drive methods, error field correction and density rising either via gas puffing or pellet injection. Despite the efforts to avoid the onset of LMs, disruption mitigation systems are considered essential during the International Thermonuclear Experimental Reactor (ITER) operations and in the view of the next fusion reactors such as the DEMOnstration Power Plant (DEMO), which can withstand a limited number of unmitigated plasma disruptions. The necessary condition towards the realization of an effective mitigation system, for safe and steady-state operations, is the development of robust and reliable metrics which are capable of predicting with a sufficient time margin the proximity to a LM. Thanks to the large availability of data, from lots of experimental campaigns performed in different experimental fusion devices, Machine Learning (ML) methods show to be a promising tool towards the achievement of this task. In this context, a wide database of MAST-U discharges has been analyzed considering data from multiple diagnostics with the scope of identifying recurrent paths which lead to LM onset. The data gathered have been the starting point for both the training and testing of two ML models, namely, K-Nearest-Neighbor (KNN) and Classification Tree (CT), developed for mode locking prediction. Both algorithms showed to be very reliable in predicting the proximity to a LM, with low percentages of missed and tardy detections. The methodology adopted for data selection, model training, as well as, the assessment of model performance are described in this work.