Interpretability of Shapley Value Attributions for Hypertension Risk Models Under Sparse Vital Sign Charting: A Computer Stability Evaluation

Sparse vital sign charting in hypertension risk models creates a specific interpretability problem: Shapley value attributions computed on incomplete feature vectors may not be stable under clinically plausible patterns of missingness. The analysis treats the attribution map as a function of the missingness pattern and defines stability as the boundedness of attribution change under clinically admissible perturbations. A variance decomposition separates between-cluster and within-cluster components of attribution variance, with an intraclass correlation of 0.64 indicating that most attribution variability is attributable to cluster-level charting practice rather than patient-level signal. The evaluation spans cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points, with 2000 replications per design cell. Power to detect a stable attribution exceeded 80% only at 56 clusters for a 5 percentage point effect, and the minimum detectable effect at 80% power declined steeply as cluster count rose. The results imply that Shapley-based explanations for hypertension risk models should be reported with explicit stability bounds tied to charting density, and that attribution stability is a design property rather than a property of the attribution method alone.