Long-term Host Immune Response Trajectories Among Hospitalized Patients With Sepsis

This multicenter cohort study of adults who have survived hospitalization for sepsis assesses whether abnormalities in the host immune response during hospitalization for sepsis persist after discharge.

1. Cardioascular: systolic blood pressure was < 90 mmHg, mean arterial pressure was ≤ 70mm Hg, the lactate level is >4 mMol/L, or subject received vasopressors 2. Renal: Increase in serum creatinine by >0.3 mg/dl (>26.5 umol/l)within 48 hours, increase in serum creatinine >1.5 times baseline, which is known or presumed to be normal, or urine volume <0.5 ml/kg/hour for 6 hours.
Subjects were excluded if they have any of the following criteria: 6. Unable to conduct follow-up home visits because subject resides in a long-term care facility (e.g. nursing home or skilled nursing facility), was homeless, or lacked reliable contact information to schedule follow-up visits

C. Readmissions
We asked patients about all overnight hospital planned and unplanned admissions at 3, 6, and 12 months. Once an admission was reported, we obtained medical records, including a history and physical examination, discharge summary, and relevant radiologic and laboratory findings.
Additionally, we queried the health records within the healthcare systems where the patients were enrolled. We centrally adjudicated a primary and secondary cause for each readmission using a structured approach. 1 Two reviewers determined the cause of readmission for a subset (n=126), and Cohen's kappa showed good agreement between raters (K=0.79, p<0.0001). All cause-specific readmission analyses included only unplanned readmissions.

D. Long-term mortality
We determined if patients had died using telephone follow-up, review of hospitalization records, and National Death Index search. Death was confirmed by obtaining death certificates whenever possible. We adjudicated cause of death by review of hospital records, next-of-kin interviews, and National Death Index records.

E. Laboratory procedures
We assessed the host immune response by measuring biomarkers at 5 time points during the index hospitalization (0-72 hours and 7-11 days) and at home (3, 6, and 12 months laboratory, aliquoted, stored at -4C, and shipped in dry ice from each site, and were centrifuged immediately using portable centrifuges and shipped overnight using cold packs (~4C) from the patient's home. The samples were stored at -80C at the University of Pittsburgh until biomarker analyses were performed.

F. Statistical analysis
We calculated descriptive statistics for clinical characteristics during the index sepsis admission and for readmissions and mortality up to 1 year. For each biomarker, we generated box plots of observed levels at the 5 time points and calculated the proportion that exceeded the reference value. Samples could not be obtained in many patients after hospital discharge, resulting in instances where there were early measures but none subsequently (dropout) or instances where there were missing observations between successful collections (e.g., measures at 3 and 12 months, but none at 6 months).
In order to identify biomarker trajectory groups in the presence of missing data, we used joint latent-class mixture models (JLCMM). The JLCMM simultaneously models biomarker trajectories and time to dropout, eliminating bias induced by the correlation between biomarker values and occurrence of dropout. It also handles intermittent missing observations through the inclusion of random effects and the use of empirical Bayes estimates. 2 Because the biomarker values were not normally distributed, we transformed the data using either log or spline functions. We modeled dropout using a Weibull model with group-specific baseline hazards, and included the following covariates: age, infection site, dialysis, vasopressor use, mechanical ventilation, acute physiology and chronic health evaluation (APACHE) II score, and total sequential organ failure assessment (SOFA) score because these factors may explain time to dropout. We did not include covariates in the longitudinal model of biomarkers to limit the influence of clinical characteristics on biomarker trajectory classes. We determined the optimal number of latent classes (trajectories) using the Bayesian information criterion (BIC).
Following identification of trajectories for individual biomarkers, we selected biomarkers with persistent immune dysregulation if the mean values of at least one trajectory for a biomarker were consistently higher than the reference range. Since more than one biomarker met this criterion, we explored patterns of trajectories across biomarkers. We determined the predicted probability of membership in each pattern by estimating the product of the probabilities of individual biomarker trajectory. These patterns were referred to as phenotypes and each patient was assigned to a phenotype.
We compared the frequency of these phenotypes in the overall sample and stratified by age and by presence of chronic diseases to ensure that the distribution of the phenotypes was similar across different age groups and in those with and without chronic diseases. We also compared the clinical characteristics of these phenotypes to determine whether baseline or hospital course affected phenotype assignment.
We chose models to determine the association between phenotypes and outcomes based on frequency of outcomes, proportionality assumptions being met, and need to incorporate competing risk. We used logistic regression to estimate odds ratios (ORs) to compare all-cause mortality, Cox models to estimate hazard ratios (HRs) to compare all-cause readmission or mortality, and the Fine-Gray model to estimate subdistribution HRs to compare cause-specific readmission or mortality in the presence of competing risk of death due to other causes. Because the proportional hazards assumption was violated for the Cox and Fine-Gray models, we estimated different hazard ratios (HRs) and their 95% CIs at prespecified intervals (0-6 months and 6 months-1 year) based on prior studies. 3,4 All models included the following covariates: age, sex, race, chronic disease, illness severity, organ support, and site of infection.

eFigure 8. Kaplan-Meier Plot of Mortality by Phenotype
This and other diagnostics assessing the proportionality of hazards between the two phenotypes suggested strongly that proportionality was violated. Because mortality was sparse in the normal phenotype, hazard ratios could not be estimated for the time period encompasing days 0 to 180 since hospitalization for sepsis. Therefore, the probability of mortality at 1 year was modeled using a logistic regression instead of hazard ratios using a Cox models. eFigure 9. Cumulative Incidence of Readmission and Mortality Stratified by Phenotype for 2 Hypothetical Patient Scenarios The hyperinflammation/ immunosuppressed phenotype is represented by a solid line and red shading, and the normal phenotype by a dashed line and blue shading. Three event types are considered: all-cause readmission or mortality (panels A and B), readmission or mortality due to cardiovascular disease (panels C and D), and readmission or mortality due to cancer (panels E and F). Hypothetical patient #1 is a 50-year white male, without chronic diseases, hospitalized for sepsis due to pneumonia, an APACHE-II score of 8, and received mechanical ventilation. Patient #2 is a 70-year old female, with three chronic diseases, hospitalized for sepsis due to pneumonia, an APACHE II score of 16, and received vasopressors and dialysis. Cumulative incidence curves were estimated using Cox or Fine-Gray models and 95% pointwise confidence intervals were estimated via bootstrapping with 5000 resamples.