# Research decision and analysis protocol

Prepared 2026-11-03 (America/Los_Angeles), before extracting trial-level mortality data or calculating an updated synthesis. This is a dated analysis plan, not a prospective registration: the investigator has read the 2025 review abstract/full text and SCREEN's abstract.

## Question and importance

Does randomized evidence support a mortality benefit from automated sepsis screening, and does that evidence specifically establish benefit from machine learning?

The strongest plausible contribution is an updated, reproducible synthesis that incorporates newer randomized evidence, separates rules-based alerts from machine learning, and uses cluster-adjusted estimates and consistent populations. This can inform consequential hospital procurement and screening decisions. Anticipated importance, if complete and convincing: 60–75/100. A repeat of the existing seven-trial analysis alone would not meet the threshold. If no meaningful evidence or methodological advance is possible, do not publish filler.

## Eligibility

Published randomized individual, cluster, and stepped-wedge trials assigning adult inpatient or emergency populations to automated sepsis alerts/screening versus a contemporaneous control without the intervention. Include any algorithm class, classify as rules-based, fitted statistical, machine-learning, or uncertain using original methods. Exclude retrospective validation, nonrandom before-after studies, protocols without results, treatment optimization after diagnosed sepsis, pediatric-only trials, and monitoring without a sepsis alert unless sepsis outcomes are separately defined. Record excluded borderline trials.

## Search

Search Europe PMC/PubMed bibliographic records from inception through 2026-11-03 with sepsis AND (alert OR screening OR warning OR detection OR monitoring) AND (random* OR trial). Supplement original references in the 2025 review (DOI 10.1007/s11606-025-09569-5), forward citation searches, paper-index searches, and ClinicalTrials.gov trial records. Record exact queries and all screened identifiers; no claim of duplicate independent human screening.

## Outcomes and estimands

Primary: intention-to-treat all-cause mortality in the randomized population, at each trial's prespecified main reported horizon (in-hospital, 28/30/90 days). Separate whole-randomized-population effects from effects restricted to post-randomization sepsis or alert subsets. Record outcome horizon, unit of randomization, model adjustment, and denominators. Prefer cluster/time-adjusted reported RR/OR/HR for cluster trials; never treat 60,000 patients as independent observations in a cluster trial. Distinguish effect measures and avoid indiscriminate OR/RR/HR pooling. Unadjusted 2x2 effects are sensitivity/descriptive analyses only for cluster designs.

## Synthesis and robustness

First present an evidence map. Pool only clinically and statistically compatible estimands. Use generic inverse variance on log relative effects with REML and Hartung–Knapp uncertainty for random effects; show fixed-effect, alternative heterogeneity estimator, leave-one-out, outcome-horizon, rules-vs-ML, and risk-of-bias sensitivities where feasible. If pooling would conflate distinct populations or effect measures, present stratified estimates rather than a misleading single number. Quantify uncertainty without equating a non-significant effect with no benefit. Report whether any conclusion depends on one trial. No causal comparison of algorithm classes from across-trial differences.

## Bias and transparency

Assess randomization, deviations, missingness, mortality measurement, selective reporting, and cluster-specific identification/recruitment/time effects from protocols and reports. Extract intervention bundles/cointerventions rather than attribute their effects solely to an algorithm. Check retractions/corrections. Maintain source URLs and table/section locations for every extracted number. Report single-agent screening/extraction as a limitation and invite independent checking if eligible. Publish only aggregate already-public data and original prose/code; no private participant-level or user data.
