Sepsis is defined by life-threatening organ dysfunction resulting from a dysregulated host response to infection (1–4). While short-term mortality has declined due to improved hospital care, sepsis exacts a considerable toll on posthospital morbidity and mortality (5–12). Yet, sepsis is a heterogeneous syndrome and patients manifest widely variable types of organ dysfunction with variable short-term mortality (13). The specific impact that each type of acute sepsis-related organ dysfunction plays in long-term mortality is unknown, but this knowledge is critical to personalize postdischarge prognosis as well as uncover potential mechanisms through which sepsis affects long-term survival (14 , 15).
Traditional approaches for estimating the association between acute organ dysfunction and long-term mortality typically fail to account for patients’ baseline predisposition for organ dysfunction (14). For example, patients with chronic kidney disease are much more likely to experience acute kidney injury (16 , 17), whereas patients with underlying neurocognitive disorders are much more likely to experience acute brain dysfunction (18 , 19). Failure to adequately account for these predisposing risk factors could result in biased estimates of the independent effects of acute, rather than chronic, organ dysfunction (14 , 19). However, there are limited existing data that can be used to quantify the risk of sepsis-related organ dysfunction based on presepsis patient characteristics.
In this study, we assessed the association between acute organ dysfunction for each of six organ systems in a large, unselected cohort from a community-based healthcare system with rich data on both their acute organ dysfunction and their longitudinal presepsis medical history. To adjust for presepsis characteristics, we applied organ-specific propensity score–matched cohorts to estimate the independent effects of each organ dysfunction on long-term mortality among those who survived hospitalization.
This study was approved by the Kaiser Permanente Northern California (KPNC) Institutional Review Board for the Protection of Human Subjects.
Our starting sample included a total of 35,000 randomly selected inpatient admissions at 21 KPNC hospitals admitted for sepsis through the emergency department (ED) between 2010 and 2013 (20). For patients with multiple sepsis hospitalizations, we kept only their first admission. We used International Classification of Disease, Ninth Edition (ICD-9), Clinical Modification present at admission diagnosis codes to identify sepsis, including 0.38 and subtypes, 995.91, 995.92, and 785.52, because these were the prevalent codes in clinical and operational use at the time in an ongoing regional sepsis quality improvement program (3 , 7 , 21).
Organ Dysfunction Scores
We quantified acute organ dysfunction with a modified Sepsis Organ Failure Assessment (SOFA) score (22) calculated in 6-hour increments starting from the time of ED arrival. Standard SOFA scores range from 0 to a theoretical maximum of 24 and include six organ subscores (hepatic, renal, coagulation, neurologic, cardiac, and respiratory) which each range from 0 to. For selected organ systems, we broadened the SOFA subscore criteria to include clinically relevant organ dysfunction variables (e.g., liver transaminase values > 200 for hepatic dysfunction, clinical documentation of agitation or coma within nursing flowsheets for neurologic dysfunction, oxygen saturation-inspired oxygen ratios for respiratory dysfunction; Appendix Table 1, Supplemental Digital Content 1, http://links.lww.com/CCM/D284) (23–26). We classified patients as having each type of acute organ dysfunction if they had a SOFA subscore greater than or equal to 1 during the first 48 hours of hospitalization. We also recorded maximum SOFA subscores within the first 48 hours (theoretically identifying acute organ dysfunction resulting from present at admission sepsis) and over the entire hospitalization (theoretically attributable to either sepsis or secondary inpatient events). We were not able to assess prehospital SOFA scores to specifically differentiate preexisting and acute organ dysfunction (e.g., patients with altered neurologic status even prior to hospitalization).
Organ Dysfunction and Hospital Mortality
To assess the association between each type of acute sepsis-related organ dysfunction and short-term mortality, we determined hospital mortality within each SOFA organ-specific subscore category. We also estimated the association between organ dysfunction and hospital mortality with multivariable logistic regression models adjusting for age, gender, predicted hospital mortality, acute severity of illness (Laboratory and Acute Physiology Score, version 2), composite comorbid disease burden (Comorbidity Point Score, version 2), ICU utilization, and full code status (24 , 25 , 27–29). To reduce residual confounding, we also adjusted these models for covariates quantifying illness severity including the maximum SOFA subscore values for each organ besides the specific organ of interest during the entire hospitalization (since patients could have multiple types of organ dysfunction within the same hospitalization).
Long-Term Mortality Analysis
Because we were primarily interested in long-term outcomes following sepsis among those who survived hospitalization, we then used organ-specific Cox proportional hazards models to assess the association of each organ dysfunction (as integer values ranging between 0 and 4) with postsepsis mortality including only patients discharged alive. We determined mortality based on electronic health records and state mortality data through April 2015 (24 , 25 , 27 , 28). We also estimated the impact of organ dysfunction as a binary value (e.g., dysfunction present or absent) on long-term mortality using the same covariates.
Sensitivity Analysis Using Propensity Score Model
We conducted several sensitivity analyses. Our primary sensitivity analysis used a propensity scoring method to adjust for patients’ presepsis risk factors for each specific organ dysfunction type. While traditional propensity scoring methods typically rely on a limited set of covariates in a logistic regression model as putative confounders, machine learning methods facilitate the development of propensity scores based on a much larger set of covariates (30 , 31). For each of the six organ dysfunction types, we developed a propensity score based on the unique counts of all possible ICD-9 diagnosis and procedure codes in the year “prior to sepsis hospitalization” as covariates. We eliminated ICD-9 codes if they were present in less than 15 sepsis patients, resulting in 3,265 ICD-9 codes and demographic variables as features (i.e., covariates). We used gradient boosted trees, the “one SE” rule, and out-of-fold prediction to estimate each patient’s organ-specific propensity score (for full details, see Appendix, Supplemental Digital Content 1, http://links.lww.com/CCM/D284).
We subjected propensity score models to a face validity check by examining whether the most important predictors of each type of organ dysfunction (based on variable importance scores) were clinically relevant to that organ dysfunction (32 , 33). We then used the propensity scores (one per organ dysfunction per patient) to create matched cohorts of patients who did and did not have that acute organ dysfunction using standard 1:1 caliper matches with a caliper size of 0.005 without replacement. To evaluate covariate balance, we quantified the absolute standardized difference in means across all covariates in each matched cohort. We evaluated adjusted Cox regression models using the propensity score–matched cohorts. In addition, we conducted sensitivity analyses using an adjusted logistic regression model where the outcome was 1-year mortality as well as a Cox regression model including death starting from the time of sepsis hospital admission, thereby including both hospital and long-term mortality.
Data are reported as mean ± SD, median (interquartile range [IQR]), or number (percent). Analyses were conducted in R (R Foundation for Statistical Computing, Vienna, Austria; https://www.r-project.org) and STATA/SE 14.1 (StataCorp., College Station, TX).
We evaluated 30,163 unique sepsis patients admitted through the ED with a median hospital length of stay of 3.7 days (Table 1). The majority of patients (60.8%) had total SOFA scores greater than or equal to two points during the first 48 hours of hospitalization. Overall hospital mortality was 9.4%; mortality was 31.7% at 1 year, 44.0% at 2 years, and 59.7% at 3 years. Our median follow-up time among sepsis survivors was 797 days from hospital discharge with an IQR of 384–1,219 days.
The total maximum SOFA score, either within 48 hours or during the entire hospitalization, was strongly correlated with hospital mortality (Fig. 1). For example, mortality was 1.7% among patients with less than or equal to 1 total SOFA points; it was 77.9% for patients with greater than or equal to 15 points. The most prevalent organ dysfunction was cardiac (62.4% of patients), whereas the least common was liver (16.5%) (Fig. 2). Increasing organ-specific SOFA subscores were also associated with increased hospital mortality (p value for trend < 0.01 for each) (Appendix Table 2, Supplemental Digital Content 1, http://links.lww.com/CCM/D284). The organ dysfunctions most strongly associated with hospital mortality included neurologic (adjusted odds ratio [OR], 1.86; p < 0.001) (Appendix Table 3, Supplemental Digital Content 1, http://links.lww.com/CCM/D284), respiratory (OR, 1.43; p < 0.001), and cardiac (OR, 1.31; p < 0.001), compared with sepsis patients without these organ dysfunctions.
Evaluating only sepsis survivors, adjusted regression analyses demonstrated that acute neurologic dysfunction was most strongly associated with long-term mortality after living discharge with an adjusted hazard ratio for a 1-point increase in the SOFA subscore of 1.18 (95% CI, 1.15–1.20; p < 0.001) (Table 2). Acute coagulation (1.12; 95% CI, 1.09–1.15; p < 0.001) and liver (1.08; 95% CI, 1.04–1.11; p < 0.001) dysfunction were also associated with increased hazards of long-term mortality in the primary analysis, albeit with smaller effect sizes.
In sensitivity analyses, the tree-based propensity score models identified the presepsis conditions that had the greatest influence on acute organ dysfunction with good clinical face validity (Appendix Fig. 2, Supplemental Digital Content 1, http://links.lww.com/CCM/D284). For example, the variables of greatest importance in predicting sepsis-related neurologic dysfunction included age, a prior history of dementia and other neurologic diseases, and a prior history of falls. For coagulation dysfunction, the most influential predictors included patients with a history of anticoagulation, thrombocytopenia, leukemia, and cirrhosis. In addition, even though the models were trained to predict binary outcomes (SOFA organ subscore ≥ 1), each model showed good stratification according to the numerical SOFA scores (Appendix Fig. 3, Supplemental Digital Content 1, http://links.lww.com/CCM/D284). The tree-based propensity score model resulted in good covariate balance (Appendix Fig. 4, Supplemental Digital Content 1, http://links.lww.com/CCM/D284).
Sensitivity analyses using the matched and full cohorts as well as Cox and logistic regression supported the finding that the most robust increase in long-term mortality was attributable to neurologic organ dysfunction (Table 2; Appendix Table 3, Supplemental Digital Content 1, http://links.lww.com/CCM/D284; and Appendix Fig. 4, Supplemental Digital Content 1, http://links.lww.com/CCM/D284), whereas the effect sizes of other organ dysfunction categories were attenuated and, in several cases, not statistically significant. The marginal increase in predicted 1-year mortality attributable to acute organ dysfunction was greatest for neurologic dysfunction (6.0%; 95% CI, 4.6–7.4%; p < 0.001) (Table 3).
We evaluated a large multicenter sample of sepsis patients to determine which infection-related organ dysfunction portended the greatest risk to long-term mortality among those surviving hospitalization. We found that neurologic dysfunction within the first 48 hours of hospitalization had the greatest adverse impact on mortality after living hospital discharge adjusting for other concomitant organ dysfunction and severity of illness metrics.
Our findings are consistent with a large body of literature demonstrating that acute neurologic or brain dysfunction occurs commonly in sepsis patients and increases short- and long-term adverse outcomes (14 , 34). In 1990, Young et al (35) reported that encephalopathy was present in 71% of sepsis patients. Similarly, in 1996, Eidelman et al (36) found that at least half of sepsis patients had encephalopathy. Several other studies have reported similar prevalence rates (37). In the current study, we also found that roughly half of our patients met criteria for neurologic dysfunction based on a modified SOFA subscore including both Glasgow Coma Score values as well as clinician documentation of mental status changes such as confusion or agitation.
Acute neurologic dysfunction in sepsis is strongly associated with adverse outcomes. The Sepsis-3 definitions incorporate altered mental status as one of three clinical factors comprising the quick SOFA score which identifies patients with increased risk of mortality or critical illness (1 , 23). In the Statins for Acutely Injured Lungs from Sepsis trial, 72% of patients exhibited delirium and, at 6 months, greater than 35% of patients in each arm exhibited cognitive impairment (38). The prevalence of delirium and the rates of long-term cognitive impairment were remarkably similar to those reported in the Bringing to Light the Risk Factors and Incidence of Neuropsychological Dysfunction-ICU study of patients with respiratory dysfunction or shock treated in the ICU (39). Other landmark studies have similarly established the increased rates of cognitive impairment after sepsis (6).
We were surprised to find that other types of organ dysfunction had a more modest effect on long-term mortality after living hospital discharge even while they were strongly associated with hospital mortality. For example, liver dysfunction was associated with an increased risk of long-term mortality in our primary analysis and in most of our sensitivity analyses. The rate of liver dysfunction in the current study was similar to the lower end of rates reported in various studies of organ dysfunction and sepsis epidemiology (40–44). Liver dysfunction is strongly associated with worsened hospital mortality, but relatively few studies have looked at its role in longer term outcomes. We found that increasing severity of respiratory and cardiac dysfunction was associated with increased hospital mortality but had favorable associations with long-term mortality among sepsis survivors. Further validation of these findings is needed.
In sensitivity analyses, we used a propensity score–matching approach to adjust for the considerable baseline differences in health that could predispose patients to increased risks of specific organ dysfunctions. To account for the multitude of potential preexisting conditions, we included greater than 3,000 counts of diagnosis codes and were able to identify presepsis clinical conditions that had clinical face validity. We also conducted additional sensitivity analyses, which confirmed that our findings were robust to a variety modeling scenarios.
Our study has several other strengths. We included a large cohort of patients drawn from a contemporary multicenter sample with granular clinical data allowing us to characterize organ dysfunction both in the early phase of hospitalization and throughout the hospital stay. These data allowed us to adjust for concomitant organ dysfunction in patients who might be experiencing multiple organ dysfunction. We also had a median follow-up time of greater than 2 years allowing us to confidently assess the role of organ dysfunction in longer-term outcomes after sepsis.
The study also has important limitations. First, we expanded our SOFA score variables in order to identify a larger spectrum of patients meeting clinical organ dysfunction criteria. For example, instead of only relying on Glasgow Coma Scale recordings, we also included patients with clinician documentation of confusion or agitation as having neurologic dysfunction. While this hampers direct comparisons with other studies which rely exclusively on SOFA criteria, we believe this enhances the clinical validity of our findings. Second, we were not able to assess prehospital SOFA scores and determine the incremental increase in SOFA scores resulting acutely from sepsis hospitalization, potentially resulting in our effect estimates including contributions of chronic organ dysfunction. However, accounting for organ-specific prehospital diagnosis data in our propensity score model resulted in largely consistent findings. Third, although we used a standard regression analysis and propensity scoring sensitivity analysis in order to establish the impact of specific organ dysfunctions on long-term mortality, there may be residual confounding that weakens the casual interpretability of our results. Finally, we were not able to ascertain the rates of residual neurocognitive impairment in our patients after hospitalization which may be a key pathway by which acute neurologic dysfunction impacts long-term mortality. We also were not able to identify long-term goals of care for patients which may differ for patients with significant neurologic dysfunction after sepsis.
In summary, we studied a large community-based cohort of sepsis patients using granular clinical data to characterize acute organ dysfunction and its impact on long-term mortality following living hospital discharge. We found that acute neurologic dysfunction during sepsis hospitalization most strongly increased the risk of long-term mortality, whereas other types of organ dysfunction had a relatively modest impact on long-term outcomes.
1. Singer M, Deutschman CS, Seymour CW, et al. The third international consensus definitions for sepsis and septic shock (Sepsis-3). JAMA 2016; 315:801–810
2. Shankar-Hari M, Phillips GS, Levy ML, et al; Sepsis Definitions Task Force: Developing a new definition and assessing new clinical criteria for septic shock: For the third international consensus definitions for sepsis and septic shock (Sepsis-3). JAMA 2016; 315:775–787
3. Liu V, Escobar GJ, Greene JD, et al. Hospital deaths in patients with sepsis from 2 independent cohorts. JAMA 2014; 312:90–92
4. Pfuntner A, Wier LM, Steiner C. Costs for hospital stays in the United States, 2010. HCUP Statistical Brief #16. January 2013. Rockville, MD, Agency for Healthcare Research and Quality. 2013. Available at: https://www.hcup-us.ahrq.gov/reports/statbriefs/sb146.pdf
. Accessed January 10, 2017
5. Iwashyna TJ, Cooke CR, Wunsch H, et al. Population burden of long-term survivorship after severe sepsis in older Americans. J Am Geriatr Soc 2012; 60:1070–1077
6. Iwashyna TJ, Ely EW, Smith DM, et al. Long-term cognitive impairment and functional disability among survivors of severe sepsis. JAMA 2010; 304:1787–1794
7. Liu V, Lei X, Prescott HC, et al. Hospital readmission and healthcare utilization following sepsis in community settings. J Hosp Med 2014; 9:502–507
8. Prescott HC, Langa KM, Iwashyna TJ. Readmission diagnoses after hospitalization for severe sepsis and other acute medical conditions. JAMA 2015; 313:1055–1057
9. Prescott HC, Osterholzer JJ, Langa KM, et al. Late mortality after sepsis: Propensity matched cohort study. BMJ 2016; 353:i2375
10. Prescott HC, Langa KM, Liu V, et al. Increased 1-year healthcare use in survivors of severe sepsis. Am J Respir Crit Care Med 2014; 190:62–69
11. Maley JH, Mikkelsen ME. Short-term gains with long-term consequences: The evolving story of sepsis survivorship. Clin Chest Med 2016; 37:367–380
12. Cohen J, Vincent JL, Adhikari NK, et al. Sepsis: A roadmap for future research. Lancet Infect Dis 2015; 15:581–614
13. Shankar-Hari M, Harrison DA, Rowan KM. Differences in impact of definitional elements on mortality precludes international comparisons of sepsis epidemiology-a cohort study illustrating the need for standardized reporting. Crit Care Med 2016; 44:2223–2230
14. Shankar-Hari M, Rubenfeld GD. Understanding long-term outcomes following sepsis: Implications and challenges. Curr Infect Dis Rep 2016; 18:37
15. Iwashyna TJ. Survivorship will be the defining challenge of critical care in the 21st century. Ann Intern Med 2010; 153:204–205
16. Mårtensson J, Bellomo R. Sepsis-induced acute kidney injury. Crit Care Clin 2015; 31:649–660
17. Uchino S, Kellum JA, Bellomo R, et al; Beginning and Ending Supportive Therapy for the Kidney (BEST Kidney) Investigators: Acute renal failure in critically ill patients: A multinational, multicenter study. JAMA 2005; 294:813–818
18. Shah FA, Pike F, Alvarez K, et al. Bidirectional relationship between cognitive function and pneumonia. Am J Respir Crit Care Med 2013; 188:586–592
19. Iwashyna TJ, Netzer G, Langa KM, et al. Spurious inferences about long-term outcomes: The case of severe sepsis and geriatric conditions. Am J Respir Crit Care Med 2012; 185:835–841
20. Liu VX, Fielding-Singh V, Greene JD, et al. The timing of early antibiotics and hospital mortality in sepsis. Am J Respir Crit Care Med 2017; 196:856–863
21. Liu VX, Morehouse JW, Baker JM, et al. Data that drive: Closing the loop in the learning hospital system. J Hosp Med 2016; 11(Suppl 1):S11–S17
22. Vincent JL, de Mendonça A, Cantraine F, et al. Use of the SOFA score to assess the incidence of organ dysfunction
/failure in intensive care units: Results of a multicenter, prospective study. Working group on “sepsis-related problems” of the European Society of Intensive Care Medicine. Crit Care Med 1998; 26:1793–1800
23. Seymour CW, Liu VX, Iwashyna TJ, et al. Assessment of clinical criteria for sepsis: For the third international consensus definitions for sepsis and septic shock (Sepsis-3). JAMA 2016; 315:762–774
24. Escobar GJ, Fireman BH, Palen TE, et al. Risk adjusting community-acquired pneumonia hospital outcomes using automated databases. Am J Manag Care 2008; 14:158–166
25. Escobar GJ, Gardner MN, Greene JD, et al. Risk-adjusting hospital mortality using a comprehensive electronic record in an integrated health care delivery system. Med Care 2013; 51:446–453
26. Adams JY, Rogers AJ, Schuler A, et al. The association between SpO2
ratio time-at-risk and hospital mortality in mechanically ventilated patients. Am J Respir Crit Care Med 2017; 195:A5029 2017
27. Escobar GJ, Greene JD, Scheirer P, et al. Risk-adjusting hospital inpatient mortality using automated inpatient, outpatient, and laboratory databases. Med Care 2008; 46:232–239
28. Liu V, Turk BJ, Ragins AI, et al. An electronic Simplified Acute Physiology Score-based risk adjustment score for critical illness in an integrated healthcare system. Crit Care Med 2013; 41:41–48
29. Kim YS, Escobar GJ, Halpern SD, et al. The natural history of changes in preferences for life-sustaining treatments and implications for inpatient mortality in younger and older hospitalized adults. J Am Geriatr Soc 2016; 64:981–989
30. McCaffrey DF, Ridgeway G, Morral AR. Propensity score estimation with boosted regression for evaluating causal effects in observational studies. Psychol Methods 2004; 9:403–425
31. Schneeweiss S, Eddings W, Glynn RJ, et al. Variable selection for confounding adjustment in high-dimensional covariate spaces when analyzing healthcare databases. Epidemiology 2017; 28:237–248
32. Elith J, Leathwick JR, Hastie T. A working guide to boosted regression trees. J Anim Ecol 2008; 77:802–813
33. Friedman JH, Meulman JJ. Multiple additive regression trees with application in epidemiology. Stat Med 2003; 22:1365–1381
34. Gofton TE, Young GB. Sepsis-associated encephalopathy. Nat Rev Neurol 2012; 8:557–566
35. Young GB, Bolton CF, Austin TW, et al. The encephalopathy associated with septic illness. Clin Invest Med 1990; 13:297–304
36. Eidelman LA, Putterman D, Putterman C, et al. The spectrum of septic encephalopathy. Definitions, etiologies, and mortalities. JAMA 1996; 275:470–473
37. Ziaja M. Septic encephalopathy. Curr Neurol Neurosci Rep 2013; 13:383
38. Needham DM, Colantuoni E, Dinglas VD, et al. Rosuvastatin versus placebo for delirium in intensive care and subsequent cognitive impairment in patients with sepsis-associated acute respiratory distress syndrome: An ancillary study to a randomised controlled trial. Lancet Respir Med 2016; 4:203–212
39. Pandharipande PP, Girard TD, Jackson JC, et al; BRAIN-ICU Study Investigators: Long-term cognitive impairment after critical illness. N Engl J Med 2013; 369:1306–1316
40. Sands KE, Bates DW, Lanken PN, et al; Academic Medical Center Consortium Sepsis Project Working Group: Epidemiology of sepsis syndrome in 8 academic medical centers. JAMA 1997; 278:234–240
41. Bakker J, Grover R, McLuckie A, et al; Glaxo Wellcome International Septic Shock Study Group: Administration of the nitric oxide synthase inhibitor NG-methyl-L-arginine hydrochloride (546C88) by intravenous infusion for up to 72 hours can promote the resolution of shock in patients with severe sepsis: Results of a randomized, double-blind, placebo-controlled multicenter study (study no. 144-002). Crit Care Med 2004; 32:1–12
42. Brun-Buisson C, Meshaka P, Pinton P, et al; EPISEPSIS Study Group: EPISEPSIS: A reappraisal of the epidemiology and outcome of severe sepsis in French intensive care units. Intensive Care Med 2004; 30:580–588
43. Vincent JL, Angus DC, Artigas A, et al; Recombinant Human Activated Protein C Worldwide Evaluation in Severe Sepsis (PROWESS) Study Group: Effects of drotrecogin alfa (activated) on organ dysfunction
in the PROWESS trial. Crit Care Med 2003; 31:834–840
44. Nesseler N, Launey Y, Aninat C, et al. Clinical review: The liver in sepsis. Crit Care 2012; 16:235
brain dysfunction; long-term mortality; organ dysfunction; outcomes research; sepsis
Supplemental Digital Content
Copyright © by 2018 by the Society of Critical Care Medicine and Wolters Kluwer Health, Inc. All Rights Reserved.