Аңдатпа
Intensive care unit (ICU) time series are irregular, incomplete, and computationally demanding to model at high temporal resolution. Dense Transformer attention captures long-range dependencies but evaluates all pairwise interactions, including many stable or clinically weak measurements. This study presents Sparse Physio-Attention, a physiology-guided Transformer that retains critical-range violations, patient-relative deviations, informative missingness patterns, and task-relevant variables before sparse attention is computed. Dynamic routing subsequently removes weak attention edges, and a late-fusion adapter incorporates static electronic health record context. The analysis included 25,368 eligible MIMIC-IV ICU stays, of which 2740 were sepsis positive. On the held-out test set, the proposed model achieved an AUROC of 0.892 and an AUPRC of 0.835 at the 24-h prediction horizon, with an expected calibration error of 0.031. Relative to dense attention, it reduced floating-point operations by approximately 72%, peak GPU memory from 3.2 to 1.8 GB, and 24-h-window latency from 670±24 to 235±11 ms. Attention-guideline agreement reached 0.826 at 24 h, indicating alignment with prespecified deterioration rules; this result reflects clinical plausibility rather than causal explanation. The evidence is limited to internal MIMIC-IV validation, and the rule-based retention thresholds may require recalibration across hospitals and patient subgroups. External and prospective validation is therefore required before clinical deployment.