Interpretability Limits of Attention-Based Sepsis Alerting Under Sparse Charting in Low-Connectivity Wards: A Computer Robustness Evaluation

Attention-based sepsis alerting models are increasingly proposed for low-connectivity hospital wards, where charting is sparse and intermittent. Their interpretability is often assumed to be a property of the model alone, yet sparse charting may systematically distort the attention weights on which clinical trust depends. the present analysis characterises the interpretability limits of attention-based sepsis alerting under sparse charting by formalising attention as a function of observed charting density and deriving a bound on attention-weight distortion. The analysis is a computer robustness evaluation: no observational data are used or claimed. A before-after-control-impact design is specified with a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), and power is estimated across 2000 design cells for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The design requires 54 clusters for 5 percentage points at 80% power. The results show that attention weights are not stable identifiers of clinical evidence when charting density falls below a threshold that depends on the intraclass correlation. Interpretability claims for attention-based sepsis alerting in low-connectivity wards therefore require charting-density reporting alongside model performance, and the paper specifies the conditions under which such claims are defensible.