Fault-induced delayed voltage recovery (FIDVR) events have become a critical threat to modern power systems with high-level inverter-interfaced renewable power generation. Aiming at the real-time assessment on FIDVR, this paper proposes a data-driven method using real-time bus voltage trajectory measurements. Based on ensemble learning and probabilistic prediction techniques, a self-adaptive decision-making model is developed to rapidly predict the FIDVR severity index following a disturbance in the system. The salient feature of the proposed method is that the FIDVR assessment result can be delivered as early as possible without impairing the assessment accuracy, thereby more time is available for emergency controls. The proposed method is tested on New England 39-bus system, and the results demonstrate its high accuracy and exceptionally faster speed over existing methods.