MR-SIB: Imputation-Free Forecasting of Sparse Multivariate ICU Time Series with a Set-Token Inter-Feature Bridge

Forecasting the near-future evolution of ICU vital signs and laboratory measurements is central to clinical decision support, enabling earlier detection of patient deterioration and more timely interventions. However, time series are challenging to model as observations are irregularly sampled, occur at multiple cadences (e.g., vital signs every few minutes versus laboratory tests every several hours), and are largely missing on a regular time grid. Most existing approaches first impute missing values onto a dense grid and then forecast, requiring the model to learn patterns of structural missingness and potentially introducing imputation bias. We present MR-SIB (Multi-Rate Set-token Inter-feature Bridge), an imputation-free forecasting architecture that represents each observation as a (value, feature, time, cadence) token and predicts future measurements through explicit query tokens for each target feature and horizon. Context and query tokens interact through double-stream attention with patient-level conditioning, while a residual forecasting head predicts deviations from each feature's most recent observation. On MIMIC-IV ICU, forecasting 63 variables over a 12-hour horizon from 24 hours of context, MR-SIB reduces masked MAE by 12.3\% and masked MSE by 31.0\% relative to a strong forward-fill baseline. It also outperforms the state-of-the-art irregular-series forecaster TPGN, the high-capacity clinical forecaster TFTM, GRU-D, and the token-based STraTS across forecast horizons, vital signs, and laboratory measurements, using only 2.4M parameters. Beyond predictive performance, we find a reproducible architecture -- physiology complementarity: set-token attention performs best on core cardiorespiratory vitals, whereas graph-based structure learning remains advantageous for device- and ventilator-derived vitals, and sparse, pathophysiologically coupled laboratory variables.