Mitigating Exposure Bias in Risk-Aware Time Series Forecasting with Soft Tokens

TL;DR

This paper introduces SoTra, a novel forecasting method that uses soft tokens to reduce exposure bias and improve risk calibration in clinical time series predictions, enhancing safety in predictive control.

Contribution

The paper presents Soft-Token Trajectory Forecasting (SoTra), a new approach that propagates continuous distributions to mitigate exposure bias and incorporate risk-aware decoding in clinical forecasting.

Findings

Reduces zone-based risk by 18% in glucose forecasting

Lowers clinical risk by approximately 15% in blood-pressure forecasting

Supports safer predictive control in critical health applications

Abstract

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.