Sepsis alert thresholds tuned once and never revisited

1. System failure identification

   A sepsis detection rule is configured at go-live and left unchanged for years, while case mix, documentation habits, and the underlying population all move away from the conditions it was tuned against. A detector tuned to a population that no longer exists degrades silently.

2. Data pipeline diagnostics

   There is no scheduled revalidation. Performance is measured at procurement and then never again, so degradation produces no signal anywhere in the system. Procurement validation answers whether it worked once, not whether it works now.

3. Predictive computational model

   Re-estimate sensitivity and false-positive rate on a rolling recent window and chart them over time. The thing under surveillance is the detector, not the patient. A rolling window turns a one-off validation into continuous surveillance.

4. Workflow integration

   Give the figures an owner and a review cadence. An unowned performance chart is a dashboard nobody opens, which is indistinguishable from not measuring at all. Measurement without ownership decays into a chart nobody reads.

Back to the case study database