Ball screw remaining useful life (RUL) prediction is of great interest to industry and academia. However, the lack of a reliable prediction model limits accuracy. To address this, a hybrid method that combines physical-based and data-driven methods is proposed. A novel integrated index is developed to capture wear degradation by integrating the preload and precision parameters, and the optimum partitioning method is used for wear stage categorization. A physical-based method of a two-stage empirical model is constructed to characterize the randomness and nonlinearity of the degradation process. Model parameters are initialized and updated using particle filtering (PF) through a data-driven method for RUL prediction. To address discontinuous predictions in the empirical model, the random forest with PF (RF-PF) method is employed. The effectiveness of this approach is evaluated through experiments and comparisons with other methods.