Forecasting X-ray tube filament degradation requires capturing changes in the current-decline rate while preserving a consistent long-horizon trend. This paper proposes a Degradation-Rate-Guided Mamba model (DR-Mamba) using Internet of Medical Things time-series observations. A causal rate estimator extracts local decline information from historical filament current. The estimated rate enriches the input representation and modulates the encoder's input-dependent discretization through a bounded pathway. A jointly trained multistep head accumulates nonnegative current decrements, preserving a nonincreasing trajectory while allowing step-dependent decline rates. Successive prediction blocks extend the forecast from a fixed origin without incorporating later measurements, and a filtered-difference criterion provides a rapid-degradation onset readout. Evaluation on clinical records from four X-ray tubes yields an average Pearson correlation coefficient of 0.9983, a root mean square error of 0.0421, and an asymmetric Score of 0.0022. Relative to LSTM, the latter two metrics decrease by 20.86% and 33.33%, respectively. Ablation results show that rate-guided state modulation improves prediction when coupled with the trend-consistent head, supporting their joint use for filament degradation forecasting.

filament degradation,selective state-space model,multistep prediction