Mitigating Exposure Bias in Risk-Aware Time Series Forecasting with Soft Tokens
This work addresses safety-critical predictive control in clinical settings like diabetes and hemodynamic management, offering incremental improvements over standard models.
The paper tackled exposure bias in autoregressive forecasting for clinical risk-aware time series, introducing Soft-Token Trajectory Forecasting (SoTra) to propagate continuous probability distributions and a risk-aware decoding module, resulting in an 18% reduction in average zone-based risk for glucose forecasting and approximately 15% lower effective clinical risk for blood-pressure forecasting.
Autoregressive forecasting is central to predictive control in diabetes and hemodynamic management, where different operating zones carry different clinical risks. Standard models trained with teacher forcing suffer from exposure bias, yielding unstable multi-step forecasts for closed-loop use. We introduce Soft-Token Trajectory Forecasting (SoTra), which propagates continuous probability distributions (``soft tokens'') to mitigate exposure bias and learn calibrated, uncertainty-aware trajectories. A risk-aware decoding module then minimizes expected clinical harm. In glucose forecasting, SoTra reduces average zone-based risk by 18\%; in blood-pressure forecasting, it lowers effective clinical risk by approximately 15\%. These improvements support its use in safety-critical predictive control.