Sepsis Bundle Compliance vs. the Alert-Fatigued Floor Nurse
A 300-bed hospital's sepsis bundle compliance (3-hour) sits at 61%, below its payer quality threshold of 75%, which puts an incentive payment of roughly $600,000 at risk. The sepsis alert fires an average of 34 times per day per unit; nurses report clicking through most of them. Chart audit of 40 missed cases shows: in 22, the alert fired but was dismissed; in 11, the patient met criteria but no alert fired; in 7, the bundle was started but the lactate was drawn late. The CMIO proposes replacing the rule-based alert with a machine-learning sepsis prediction model. The nursing quality council is skeptical and wants more education instead. The medical director of quality notes that two of the missed cases resulted in unplanned ICU transfers and one in a prolonged stay.
1. Using the audit data, decide between the ML model, more education, and a third option, and justify the choice.
Read the worked answer
The audit data rules out education as the primary fix and undercuts the ML model as a first move.
Break down the 40 misses:
- 22 (55%) - alert fired, dismissed. This is not a knowledge gap; it is a signal-to-noise and workflow problem. 34 alerts/day/unit is far beyond what a human can meaningfully process. Education does not fix a dismiss-reflex; it may worsen it by adding another mandatory module.
- 11 (27.5%) - criteria met, no alert. This is a detection gap. An ML model could plausibly help here - but it addresses the smaller bucket.
- 7 (17.5%) - bundle started, lactate late. This is a process/turnaround gap (lab draw, transport, result routing), not a prediction gap.
Recommendation: Sequence a process fix first, because it attacks the largest buckets at low cost:
- Reduce alert burden - tier alerts (silent for low-risk, interruptive only for high-risk), which directly targets the 22 dismissals.
- Fix the lactate loop - standing order for lactate at trigger, point-of-care or priority lab routing, targeting the 7.
- Then evaluate the ML model as a detection intervention for the 11, with a prospective validation and a plan for how its output enters the workflow (a prediction that no one acts on is just a better-looking alert).
Education should be a small, targeted component (e.g., how to escalate a dismissed alert), not the strategy.
2. The payer threshold is 75% and the hospital is at 61% with the incentive at risk. How should leadership frame the improvement target and timeline to the board?
Read the worked answer
Resist the temptation to promise 75% by the next measurement period. A 14-point jump in one cycle is not credible from a 61% baseline and sets the organization up to miss publicly.
Frame it as a trajectory with leading indicators:
- Split the outcome metric into its drivers. Report alert burden (alerts/unit/day), dismissal rate, detection rate (met-criteria-no-alert), and lactate turnaround separately. These move faster than the composite and give the board early evidence.
- Set an interim milestone (e.g., 68% at 90 days) tied to the specific interventions, with a clear statement of which interventions are expected to produce which points.
- Be honest about the incentive. Model the partial-payment scenarios so the board understands the financial exposure is graduated, not binary - and state plainly that full recovery may not be achievable this cycle.
- Name the safety stakes. The two ICU transfers and one prolonged stay are the reason to act, independent of the $600K. Leading with the incentive invites gaming; leading with harm invites engagement.
- Assign a single accountable owner with authority across nursing, lab, and informatics, since the failure modes span all three.