Strategies for Handling Missing Serum Lactate When Identifying Septic Shock in ICU Databases: A Plasmode Simulation Study in Alberta ICUs and MIMIC-IV

Abstract

Septic shock is defined by a serum lactate of 2 mmol/L or more measured around the time of suspected infection, but ICU databases often lack a lactate in that window. Studies exclude such patients, use a value measured at another time, predict it or impute it, and the choice changes who is classified as having septic shock. This project compares ten such strategies in a plasmode simulation, in which lactate values are deleted from patients whose values are known under four missingness mechanisms and three proportions missing, and in an application to patients whose lactate is genuinely missing. Performance is judged for the classification of septic shock and for the bias, precision and confidence-interval coverage of the adjusted hazard ratio for ICU death. The primary cohort is the operational Sepsis-3 cohort of Alberta ICUs, 2012-2021 (two extracts of eCritical Alberta); the replication cohort is MIMIC-IV v3.1. Registration. This registration contains the study design and data analysis plan (version 1.2, 16 October 2026; attached) of a plasmode simulation study of ten strategies for handling a missing serum lactate when septic shock is identified in ICU databases: complete case; the closest-in-time value; the maximum value within 3, 5 or 7 days; linear and logistic regression single imputation; and multiple imputation with a linear model, a logistic model or predictive mean matching. Cohorts. Primary: adults in Alberta ICUs, 2012-2021, who met every operational Sepsis-3 criterion for septic shock except lactate (a vasopressor infusion; an antibiotic and a body-fluid culture within the Seymour time epoch; a SOFA score of 2 or more), from two extracts of eCritical Alberta built with one code. Replication: the same definition in MIMIC-IV v3.1 (Sepsis-3 table of the MIMIC code repository). Part 1 (simulation). In participants with an in-window lactate (48 hours before to 24 hours after time zero), lactate is deleted under MCAR, MAR (odds ratios from the observed missingness model), MAR depending on ICU death, and MNAR, at 10%, 20% and 40%. Each replicate resamples 5,000 participants; 5,000 replicates per scenario in the primary cohort and 2,000 in the replication cohort. Estimands: the adjusted log hazard ratio for ICU death (lactate of 2 mmol/L or more versus less; Cox model with follow-up from a landmark 24 hours after time zero to 30 days), against its full-data value, and the true septic shock status of participants whose lactate was deleted. Performance measures follow Morris et al.: bias, empirical and model standard errors, coverage, bias-eliminated coverage, mean squared error and relative precision; accuracy, sensitivity, specificity and positive predictive value of the classification. Part 2 (application). The strategies are applied to participants whose lactate is genuinely missing: the number each adds to the septic shock cohort, their characteristics and outcomes, and the hazard ratio. Subgroups: ICU type in both cohorts; period of admission (2012-2018, 2019-2021) in the primary cohort. Sensitivity analyses: follow-up from the lactate draw; general ICUs only; first vasopressor within 24 hours of time zero; a lactate above 2 mmol/L; Alberta 2012-2018 only; a full-size bootstrap (MIMIC-IV). Prior data use. Known before registration: the 2025 descriptive comparison and pilot simulations in the 2012-2018 Alberta data; descriptive checks and code test runs in a 2012-2025 data cut that is no longer used; an aggregate report of the MIMIC-IV cohort (counts, lactate availability, crude mortality by lactate availability); an inventory of the 2019-2021 extract (file and column names, row counts). No simulation and no application of the strategies with the definitions of this plan had been run in either cohort. The analyses will be run as registered; any later change will be registered as a dated amendment and reported alongside the pre-specified results. Ethics and funding. University of Calgary Conjoint Health Research Ethics Board, REB19-0959 (both Alberta extracts); MIMIC-IV under the PhysioNet Credentialed Health Data Use Agreement. Funding: none. Note on the registry. The plan states that the study will be registered with the Open Science Framework (OSF). OSF was unavailable from 16 to 19 October 2026 for a scheduled cloud infrastructure migration, so version 1.2 of the plan was deposited here instead as a public, time-stamped record. The attached file is the plan as written, unchanged.