Hospital-Onset Sepsis Systematic Detection and Structured Interventions

BACKGROUND AND RATIONALE Hospital-onset sepsis carries higher mortality than sepsis present on admission, and recognition on general wards depends on individual clinician vigilance. At the study institution, approximately 300 adult sepsis events occur annually on the target units. Observed outcomes in this population include \>20% inpatient mortality and \>50% unplanned transfer to intensive care within 72 hours. The median interval from the sepsis reference time to a new antibiotic order is 15 hours (interquartile range 6 to 23 hours), and fewer than 5 percent of patients receive a new antibiotic within 3 hours. Approximately 55 percent of these patients are already receiving antibiotics at the reference time, so the potential failure for that majority is delayed reassessment and escalation rather than delayed initiation. The institution already runs Epic Sepsis Model version 2 (ESM2) passively. Scores are visible on configurable patient lists but are not wired to any structured response. The trial therefore does not evaluate the algorithm, whose discrimination has been characterized externally and locally. It evaluates whether linking an existing detection signal to a structured response pathway with independent ordering authority changes patient outcomes. Prior evaluations of machine-learning early-warning systems have generally demonstrated process improvement without consistent outcome benefit, which establishes equipoise for the pathway rather than for the algorithm. SETTING AND POPULATION Single-center pragmatic trial at an academic medical center. Target units are non-cancer medicine (hospitalist, resident, and family medicine services), general surgery, and emergency department observation. These units account for approximately 60% of hospital-onset adult sepsis events at the institution. At an ESM2 threshold of 30, approximately 2,500 threshold crossings occur per year across these units, corresponding to roughly 8 alerts per day with an approximately flat distribution across hours of the day. The positive predictive value for sepsis at this threshold is approximately 5 percent, which is anticipated and is addressed analytically rather than by threshold restriction. INTERVENTION PATHWAY For patients assigned to the intervention condition, crossing the ESM2 threshold triggers a Best Practice Advisory delivered to the assigned bedside nurse. The advisory presents a small set of structured response options, including declining activation with a captured reason. The architecture is default-to-action: if the nurse does not respond within a pre-specified timeout window, the Critical Care Activation Team is paged automatically. Inaction therefore produces help rather than silence, and the nurse's discretion is exercised by declining rather than by initiating. The Critical Care Activation Team is an existing 24-hour rapid response service staffed by critical care clinicians credentialed for independent ordering. On activation, the team performs a bedside evaluation, may place diagnostic and therapeutic orders directly, and documents the encounter in a structured note. Each activation is classified into one of five categories: sepsis unrecognized before evaluation; sepsis recognized with bundle incomplete or delayed; sepsis recognized with care already appropriate; not sepsis but acute illness with an alternative diagnosis pursued; and not sepsis with the patient clinically stable. This classification replaces binary true-positive and false-positive framing and is used for fidelity and threshold-refinement analyses. The primary team is notified simultaneously with activation and retains authority to modify or discontinue any team-initiated order at any point. Overrides occurring before the four-hour mark are documented in the structured note with a reason, and override rate is analyzed as an implementation outcome. A four-hour reassessment by the responding team is required for activations classified into the first two categories and constitutes the formal handoff back to the primary team. CONTROL CONDITION For patients assigned to the control condition, the study advisory is suppressed silently. No advisory is displayed, no automatic orders are generated, and no page is sent. Suppression is limited to the study's own advisory and its two downstream actions. No pre-existing alert, alarm, order set, score display, or item of clinical information is altered or withheld in either condition. ESM2 scores remain visible on patient lists exactly as they are today, and the separately deployed deterioration index continues to trigger the existing rapid response pathway at its established thresholds. The care a control-condition patient receives is therefore identical to current standard practice at the institution. RANDOMIZATION, CONCEALMENT, AND MASKING Assignment is computed deterministically within the electronic health record at the first qualifying threshold crossing of an admission, with 1:1 random allocation (with block size 2). Subsequent threshold crossings within the same admission inherit the original assignment. The trial is unstratified; unit-level and temporal subgroup effects are addressed analytically. Concealment holds until the threshold crossing by the determinism of the rule and the silence of the suppression mechanism. At crossing, assignment resolves operationally on the intervention side only. On the control side no human observes the suppressed advisory, and the assignment never becomes operationally visible during care. Patients are unmasked by design. Clinicians are unmasked on the intervention side and effectively masked on the control side, having never been alerted. The analytic dataset is prepared with condition replaced by a study-side code, and the analyst conducting the pre-specified primary analysis remains masked until the locked analysis. RESPONSE FIDELITY INSTRUMENTATION An ordered chain of timestamped checkpoints is captured natively in the electronic health record: advisory fired, nurse acknowledgment, team page, team bedside evaluation, per-element order placement with ordering-clinician attribution, and override capture. These support a completion funnel across the pathway and a decomposition of time-to-antibiotic into a detection-and-decision interval and a system-execution interval. An interpretive rule is pre-specified in advance of analysis to separate failure of the concept from failure of implementation on the basis of observed fidelity, so that a null result can be attributed rather than left ambiguous. IMPLEMENTATION STAGING The advisory, automatic orders, and paging workflow are first deployed in a non-randomized configuration on pilot units for a minimum of 1 week. Randomization begins only after that period completes without major workflow incident. Nurse education is delivered identically to staff caring for patients in both conditions and is treated as a study artifact with pre-specified scripts, materials, and a post-training assessment. Because education is necessarily uniform across conditions, a secular-trend analysis of time-to-antibiotic among non-alerted threshold-crossing patients is pre-specified as a probe for broader attention effects attributable to education rather than to the pathway. SAFETY MONITORING An independent safety monitor who is not a member of the study team reviews unmasked safety data at pre-specified intervals. Pre-specified operational triggers prompt immediate pause and review: a death within 24 hours of an activation classified as clinically stable; a death of a control-condition patient within 24 hours of a suppressed advisory; a cluster of adverse events associated with team-initiated antibiotics; team bedside response exceeding 30 minutes for more than 20 percent of activations in a week; nurse decline rate above 60 percent sustained over two weeks; and a stable-patient classification rate above 50 percent sustained over two weeks. STATISTICAL APPROACH The primary analysis is by intention to treat and estimates a common odds ratio on the day-5 clinical-state ordinal using proportional odds logistic regression, with all threshold-crossing patients analyzed by assigned condition regardless of advisory response or team arrival. Sample size was determined by Monte Carlo simulation against the empirically observed control-condition ordinal distribution rather than by analytic approximation, which was found to be optimistic in this setting. Eighteen months of randomization at the observed accrual rate yields approximately 1,840 patients per condition and approximately 86 percent power to detect a common odds ratio of 0.80. Extension to 24 months is reserved as a contingency if realized accrual undershoots projection by more than 10 percent. A per-protocol sensitivity analysis restricted to patients who received a team evaluation is pre-specified alongside the primary analysis, as is an effect-modification analysis by actionable lead-time window, defined from the distribution of threshold-crossing time relative to first antibiotic order in a retrospective cohort. REGULATORY FRAMING The trial is structured as a single integrated protocol with layered determinations. The algorithm and advisory workflow are implemented through routine clinical decision support governance. Patient-level randomization and outcome ascertainment proceed as minimal-risk research under a waiver of informed consent and waiver of documentation of consent. Accompanying clinician interviews and surveys proceed as exempt research with verbal consent.