Paper Details

Sepsis is a leading cause of death in intensive care units (ICUs), and every hour of treatment delay increases mortality, so hourly early-warning models built from routine bedside data have clear practical value. Bedside records are irregularly sampled, with most laboratory values observed in fewer than 10% of hours. Recurrent neural networks remain the dominant sequence model for this task, and the recently proposed Mamba selective state-space model has not been evaluated on hourly sepsis prediction. This paper presents a missingness-aware Mamba model for the PhysioNet/Computing in Cardiology Challenge 2019 dataset (40,336 ICU stays). Each hour is represented by forward-filled, standardised measurements concatenated with a binary observation mask and processed causally by residual Mamba blocks. Under identical preprocessing and an equal hyper-parameter search per model, Mamba is compared with GRU and LSTM baselines over five random seeds using the area under the ROC curve (AUROC), the area under the precision–recall curve (AUPRC) and the official utility score. Mamba reached an AUROC of 0.827 ± 0.005, an AUPRC of 0.112 ± 0.010 and a utility score of 0.386 ± 0.018, matching the tuned recurrent baselines within seed-to-seed noise and with the lowest variance in AUROC. Removing the observation mask reduced AUROC by 0.030 and utility by 0.054, several times the seed-to-seed standard deviation. Missingness encoding therefore matters more than the choice of backbone on this task.

Yuze Li, Tohgoroh Matsui, “Missingness-Aware Mamba for Early Sepsis Prediction from Irregularly Sampled ICU Time Series”, International Journal of Advance Engineering and Research Development (IJAERD), Vol. 13, Issue 10, pp. 6-13, October 2026, DOI: 10.5281/zenodo.23601054.