SEARCH

SEARCH BY CITATION

Keywords:

  • sepsis;
  • severity score;
  • severity of illness;
  • mortality prediction

ABSTRACT

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. METHODS
  5. RESULTS
  6. DISCUSSION
  7. ACKNOWLEDGEMENTS
  8. CONFLICT OF INTEREST
  9. REFERENCES
  10. APPENDIX: SEPSIS CODES
  11. APPENDIX: MIXED-EFFECTS LOGISTIC REGRESSION MORTALITY PREDICTION MODEL*

Purpose

Mortality prediction models can be used to adjust for presenting severity of illness in observational studies of treatment effectiveness. We aimed to determine the incremental benefit of adding information about critical care services to a sepsis mortality prediction model.

Methods

In a retrospective cohort of 166 931 eligible sepsis patients at 309 hospitals, we developed nested logistic regression models to predict mortality at the patient level. Our initial model included only demographic information. We then added progressively more detailed information such as comorbidities and initial treatments. We calculated each model's area under the receiver operating characteristic curve (AUROC) and also used a sheaf coefficient analysis to determine the relative effect of each additional group of variables.

Results

Model discrimination increased as more detailed patient information was added. With demographics alone, the AUROC was 0.59; adding comorbidities increased the AUROC to 0.67. The final model, which took into account mixed (hierarchical) effects at the hospital level as well as initial treatments administered within the first two hospital days, resulted in an AUROC of 0.78. The standardized sheaf coefficient for the initial treatments was approximately 30% greater than that for demographics or infection source.

Conclusions

A sepsis disease risk score that incorporates information about the use of mechanical ventilation and vasopressors is superior to models that rely only on demographic information and comorbidities. Until administrative datasets include clinical information (such as vital signs and laboratory results), models such as this one could allow researchers to conduct observational studies of treatment effectiveness in sepsis patients. Copyright © 2012 John Wiley & Sons, Ltd.


INTRODUCTION

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. METHODS
  5. RESULTS
  6. DISCUSSION
  7. ACKNOWLEDGEMENTS
  8. CONFLICT OF INTEREST
  9. REFERENCES
  10. APPENDIX: SEPSIS CODES
  11. APPENDIX: MIXED-EFFECTS LOGISTIC REGRESSION MORTALITY PREDICTION MODEL*

The ability to adjust for the severity of illness of patients receiving different treatments is critical to the conduct of high-quality observational comparative effectiveness research. Severity adjustment methods range in their complexity and predictive ability, from basic approaches that include only demographic information and comorbidity counts to methods that use hierarchical regression models and diagnoses present at admission to calculate risk-standardized mortality estimates at the hospital level.[1-4] Although actual measures of physiologic function and results of laboratory testing are powerful tools for severity adjustment in observational studies,[5-7] they are rarely available in administrative data. Thus, severity adjustment often depends upon information derived from billing data or claims.

Observational studies in the critical care setting have been a notable exception to this rule. Traditionally, these studies have adjusted for presenting severity using validated mortality prediction models that require clinical data obtained from chart review.[8-12] For example, the Acute Physiology and Chronic Health Evaluation (APACHE) score includes items such as vital signs, white blood cell count, serum albumin, and Glasgow Coma Score.[8, 12] This requirement has greatly limited the size of studies that use clinically derived severity adjustment methods, because the data are cumbersome and costly to collect.

A recent innovation in the development of claims-based multi-institutional datasets has been the incorporation of detailed information derived directly from the cost accounting systems of participating institutions, enriching these datasets with a time- or date-stamped log of all treatments and services provided to individual patients.[13, 14] Such data have the potential to improve the predictive capability of administrative mortality models, because treatments administered early in the hospitalization might be a useful proxy for physiologic measures of initial severity in patients with critical illness. For example, mechanical ventilation suggests the presence of respiratory failure, and the use of vasopressors suggests that a patient was in shock.[15] In a prior study, our team described a model that used hospital billing data to calculate a probability of mortality based on patient demographics, comorbidities, and selected treatments (mechanical ventilation, vasopressors, and intensive care unit [ICU] admission) administered on Hospital Days 1 and 2. In a development cohort of 166 931 adults with sepsis at 309 hospitals between 2004 and 2006, predicted mortality ranged from 0.002 to 0.938, with a mean of 0.199 and a Hosmer–Lemeshow c statistic[16] of 0.77, similar to the performance of models that depend on chart review, such as the Mortality Prediction Model (MPM)III. In a validation of this model on 357 adult sepsis patients, our model had discriminatory ability similar to that of mortality prediction models that contain data obtained from chart review (area under the receiver operating characteristic curve [AUROC] of all models was statistically similar: APACHE II = 0.71, SAPS II = 0.74, MPM III = 0.69, administrative model = 0.69, p= 0.35). Calibration of our administrative model was superior.[15]

Although this work examined the overall performance characteristics of our model derived from billing data, it did not describe the contribution of different groups of variables to mortality prediction. Here, we illustrate the incremental value of including demographics, comorbidities, and selected initial critical care treatments in a mortality model for patients with sepsis.

METHODS

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. METHODS
  5. RESULTS
  6. DISCUSSION
  7. ACKNOWLEDGEMENTS
  8. CONFLICT OF INTEREST
  9. REFERENCES
  10. APPENDIX: SEPSIS CODES
  11. APPENDIX: MIXED-EFFECTS LOGISTIC REGRESSION MORTALITY PREDICTION MODEL*

Data source

We used the voluntary, fee-supported perspective database (Premier Healthcare Informatics, Charlotte, NC) to gather data from hospitals that cared for patients with sepsis between 1 June 2004 and 30 June 2006. Perspective contains the elements found in hospital claims derived from the uniform billing 04 (UB-04) form as well as an itemized, date-stamped log of all items and services charged to the patient or insurer. These include medications, laboratory tests, diagnostic tests, and therapeutic services. Participating hospitals represent approximately 15% of all US hospitalizations and were drawn from all regions of the USA. Permission to conduct the study was obtained from the institutional review board at Baystate Medical Center.

Population

We have previously described a cohort that included patients who were 18 years or older, had a principal or secondary diagnosis of sepsis (APPENDIX),[17] were treated with antibiotics, and underwent blood culture.[15, 18] A diagnostic code of sepsis is assigned based on clinician documentation of sepsis, defined as infection plus host inflammatory response.[19] Because there is some controversy about the use of diagnostic codes for identification of sepsis cases,[20, 21] we attempted to increase the specificity of the definition of sepsis by limiting the population to patients who also had evidence of diagnosis and treatment for infection, as indicated by the presence of blood culture and antibiotics. Additional inclusion and exclusion criteria have been described elsewhere.[15] The population included 309 hospitals, contributing a total of 166 931 patients (Table 1).

Table 1. Characteristics, treatments, and outcomes of patients with sepsis at 309 US hospitals
 Median or n[Interquartile range] or (%)
Totaln = 166 931 
Demographics  
Age (in years)70[56, 81]
Female86,259(51.7)
Race/Ethnicity  
White103,337(61.9)
Black30,957(18.5)
Hispanic7,981(4.8)
Other/Unknown24,656(14.8)
Marital status  
Married62,853(37.7)
Widowed34,331(20.6)
Single31,820(19.1)
Separated/Divorced15,270(9.2)
Other6,543(3.9)
Not recorded16,114(9.7)
Insurance  
Medicare traditional105,448(63.2)
Managed care23,731(14.2)
Medicaid traditional12,230(7.3)
Self-pay/Other9,311(5.6)
Medicare managed8,289(5.0)
Commercial/Private5,689(3.4)
Medicaid managed2,233(1.3)
Clinical characteristics  
Comorbidities  
Hypertension59,307(35.5)
Diabetes (with or without complications)54,473(32.6)
Anemia (chronic and acute)52,335(31.4)
Chronic pulmonary disease42,012(25.2)
Congestive heart failure41,492(24.9)
Neurological disorders32,060(19.2)
Renal failure29,347(17.6)
Weight loss21,603(12.9)
Solid tumor without metastasis17,463(10.5)
Hypothyroidism16,852(10.1)
Depression14,067(8.4)
Peripheral vascular disease14,070(8.4)
Valvular disease12,854(7.7)
Paralysis10,390(6.2)
Obesity9,664(5.8)
Metastatic cancer8,671(5.2)
Liver disease7,223(4.3)
Psychoses6,461(3.9)
Alcohol abuse6,454(3.9)
Rheumatoid arthritis/Collagen disease5,728(3.4)
Lymphoma4,529(2.7)
Drug abuse4,397(2.6)
Pulmonary circulation disease3,794(2.3)
Peptic ulcer disease2,875(1.7)
AIDS503(0.3)
Site of infection  
Urinary (primary or secondary)63,648(38.1)
Lung (primary or secondary)52,753(31.6)
Abdominal (primary or secondary)23,728(14.2)
Other (e.g., skin, bone) or unknown43,208(25.9)
Infection type  
Gram-positive29,152(17.5)
Gram-negative28,062(16.8)
Other (fungal, viral, anaerobic, mixed)2,491(1.5)
Unknown107,226(64.2)
Medical (vs. surgical)139,139(83.4)
Surgical type  
General or gastrointestinal10,433(6.3)
Cardiac or vascular2,601(1.6)
Orthopedic4,439(2.7)
Other5,527(3.3)
Treatments (by Day 2)  
Intensive care unit (at least 1 day)59,871(35.9)
Mechanical ventilation24,936(14.9)
Vasopressors36,965(22.1)
Outcome  
In-hospital mortality33,192(19.9)

Model development

We created a series of nested logistic regression mortality prediction models with robust standard errors [22] to illustrate the incremental value of adding more detailed patient information to traditional risk adjustment methods. We assumed that patient outcomes were not independent within hospitals and thus imposed a clustered structure at the hospital level that incorporates robust standard errors. The use of robust standard errors and this model structure may result in a slight increase in the c statistic and slightly different coefficients than if we had used a more “naive” traditional logistic regression model.[23, 24]

Model 1 included only demographic variables: age, race, gender, marital status, and insurance. Marital status, insurance status, and race were recorded by admission or triage staff of participating hospitals using hospital-defined options. Model 2 included the demographic variables plus the presence of 25 comorbid conditions identified using the software provided by the Healthcare Costs and Utilization Project of the Agency for Healthcare Research and Quality.[2] Model 3 included demographic and comorbidity information as well as source of infection, type of infection, and a categorization of the case as either medical or surgical in nature. Source (lung, abdomen, urinary tract, blood, other) and type of infection (gram-positive, gram-negative, mixed, anaerobic, fungal) were identified using diagnosis codes.[20] If patients were categorized as surgical, we further classified them according to the type of operation. Model 4 included all of these characteristics as well as indicators for the presence or absence of three treatments provided within the first two hospital days: admission to the ICU, treatment with mechanical ventilation, and receipt of vasopressors. Finally, we constructed a multilevel hierarchical mixed-effects logistic regression model that included all the predictors in Model 4 (APPENDIX).[25] We used this model structure to adjust for intraclass correlation among patients within each hospital and to focus on patient-specific rather than on population average effects.[26]

Analysis

For each model, we calculated c statistics (AUROC) with confidence intervals based on 100 bootstrap samples with replacement on the entire dataset to measure discriminatory ability.[27] We constructed calibration plots to assess calibration and calculated Efron's pseudo-R2 statistic. We used Efron's pseudo-R2 because R2, the coefficient of determination, is used commonly in linear regression, but there is no equivalent for logistic regression because the outcome is dichotomous. Efron's pseudo-R2 is derived by having the model residuals squared, summed, and divided by the total variability in the dependent variable, mortality; it is most useful when evaluating different models predicting the same outcome on the same dataset.[28] Because the outcome is binary, Efron's pseudo-R2 generally has much lower values than those of a traditional R2—a value of 0.2 or higher usually indicates a well-fitted model.

Bayesian Information Criterion (BIC) statistics were also derived to compare the likelihoods of the nested models.[28, 29] The BIC depends on two factors: the log-likelihood and the number of parameters relative to the sample size. This creates a “penalty” for inclusion of new variables that do not contribute to the predictive ability of the model.[29]

We used a sheaf coefficient analysis to determine the relative effect of each additional “group” of variables (demographics, comorbidities, infection source and type, and three initial critical care treatments) on mortality.[30] Sheaf coefficients allow the use of a single scale to quantify the effects of groups or blocks of variables on the outcome. This is done by first creating latent variables based on each related group of variables used in the final mortality prediction model. The latent variables and the predicted probabilities (y-hat) are then standardized to have a mean of 0 and a standard deviation of 1 so that the latent variable coefficients may readily be compared.

Finally, to further understand the relative explanatory power of the variables in the prediction model, we calculated the log likelihoods of models when groups of variables were excluded and then normalized and compared our results with those of the log likelihood of the full model.[12] All analyses were carried out using STATA/SE 10.1 (StataCorp, College Station, TX).

RESULTS

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. METHODS
  5. RESULTS
  6. DISCUSSION
  7. ACKNOWLEDGEMENTS
  8. CONFLICT OF INTEREST
  9. REFERENCES
  10. APPENDIX: SEPSIS CODES
  11. APPENDIX: MIXED-EFFECTS LOGISTIC REGRESSION MORTALITY PREDICTION MODEL*

The discriminatory ability of the mortality prediction models increased substantially across the nested models (Table 2). Using demographics alone, the AUROC was 0.59. Adding comorbidities increased the AUROC to 0.67. Adding infection type, site, and medical or surgical variables further increased the AUROC to 0.71. Addition of ICU admission, mechanical ventilation, and receipt of vasopressors yielded an AUROC of 0.77. The final model, which took into account mixed hierarchical effects at the hospital level, resulted in an AUROC of 0.78 (Figure 1). Efron's pseudo-R2 also increased with each model (Table 2), indicating an improved fit as additional terms were added. The BIC statistic decreased substantially for each model, indicating that each additional group of variables improved predictive ability.

Table 2. Performance statistics of the nested models
Performance statisticModel 1: age, race, gender, marital status insurance statusModel 2: Model 1 + Elixhauser comorbiditiesModel 3: Model 2 + infection site, infection type, primary versus secondary sepsis, medical versus surgical, surgical typeModel 4: Model 3 + initial treatments of ICU admission, mechanical ventilation, vasopressor useFinal model: Model 5 with mixed (hierarchical) effects at the hospital level
  1. Note: AUROC, area under the receiver operating characteristic curve; CI, confidence interval; BIC, Bayesian Information Criterion.

AUROC (c statistic)0.594 (0.591. 0.598)0.669 (0.662, 0.673)0.715 (0.712, 0.718)0.772 (0.769, 0.775)0.78
95%CI predicted mortality probability range(0.062, 0.377)(0.014, 0.741)(0.004, 0.803)(0.004, 0.925)(0.002, 0.938)
BIC statistic163 594157 667151 638141 002 
Efron's R20.0180.0540.0890.163 
Sheaf coefficients     
Demographics   0.23 
95%CI (0.22, 0.24)
Comorbidities   0.21 
95%CI (0.20, 0.22)
Infection source and type   0.23 
95%CI (0.22, 0.23)
Initial treatments   0.30 
95%CI (0.29, 0.30)
image

Figure 1. Receiver operating characteristic (ROC) curves for nested models.

Download figure to PowerPoint

The predicted probabilities of death for each subgroup revealed that the final model had a much broader range of probabilities than that of the demographics-only model (Figure 2). The latter tended to overpredict mortality for many patients in the dataset. The final model-predicted mortality rate, as well as the range of predicted mortality, increased with age. The range of predicted probabilities was similar for medical and surgical patients. We examined the calibration of the final model among subgroups (e.g., age groups or medical vs. surgical) of patients by plotting observed mortality and predicted mortality by deciles. Calibration was very good for all subgroups, except surgical patients (Figure 3).

image

Figure 2. Box plots for selected models and patient subsets

Download figure to PowerPoint

image

Figure 3. Calibration plot for medical and surgical patients

Download figure to PowerPoint

The standardized sheaf coefficients were statistically similar (p = 0.76) for demographics and infection type and source (Table 2). Age was the most important contributor when creating the latent variable of demographics. The standardized sheaf coefficient for the initial treatments was approximately 30% greater than that of the demographics or infection source (both of which had comparable but slightly larger standardized sheaf coefficients than that of comorbidities). Analyzing the difference in log likelihoods of models when groups of variables were excluded, we also found that the initial treatments had substantially higher relative explanatory power than that of the other three variable groups (Figure 4).

image

Figure 4. Unique relative contribution of each risk factor group to hospital mortality prediction calculated using difference in log likelihoods

Download figure to PowerPoint

DISCUSSION

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. METHODS
  5. RESULTS
  6. DISCUSSION
  7. ACKNOWLEDGEMENTS
  8. CONFLICT OF INTEREST
  9. REFERENCES
  10. APPENDIX: SEPSIS CODES
  11. APPENDIX: MIXED-EFFECTS LOGISTIC REGRESSION MORTALITY PREDICTION MODEL*

In a large, nationally representative sample of patients with sepsis in both critical-care and non-critical-care settings, we found that a severity adjustment model that included ICU admission and use of both mechanical ventilation and vasopressors during the first two hospital days showed discrimination superior to models that included only demographics, comorbidities, and disease characteristics. Although including each additional group of variables produced a statistically significant improvement in model performance, initial treatments represented the most important addition in terms of predictive ability, and only by including these terms did we achieve an AUROC comparable with that of validated models traditionally used in the observational studies of critical-care patients.

There is a well-developed literature describing different approaches for carrying out risk adjustment for hospitalized patients. After adjustment for age, gender, and race, the most common focus for risk adjustment is the identification and control of comorbidity, such as the well-established methods described by Charlson and Elixhauser.[1, 2, 31] Newer severity adjustment models have taken advantage of automated clinical data from electronic medical records such as laboratory results and vital signs.[5-7]

Although models using a combination of administrative and automated clinical data have been used previously to adjust for severity of illness in patients with critical illness, the gold standard for risk adjustment in these patients remains mortality prediction models such as APACHE and MPM.[8, 9, 11, 32] Systematic evaluation of such clinically derived severity adjustment methods, however, has shown wide variation in the ability of these models to predict mortality.[33] When applied to real-world populations (such as the large, multihospital population in our study), c statistics for these models tend to be very similar to the values we found with our administrative models. APACHE II, for example, had a median c statistic of 0.77 using pooled data, and MPM II had a c statistic of 0.73.[33] Recently, both APACHE and MPM have introduced newer models, the MPM III and the APACHE IV, which have both shown good calibration and discrimination in contemporary populations, with c statistics above 0.8.[10, 12] However, these models have had limited validation in populations outside those in which they were developed. Moreover, they include even more variables than previous iterations (particularly APACHE IV), with a corresponding increased burden of chart review.

Information regarding the timing of treatments and services provided to patients offers an alternative approach to risk adjustment for patients with critical illness, because this information can serve as a proxy for physiologic measures, allowing estimation of the presenting severity of illness without the burden of clinical data collection. This approach was previously validated and found to have performance similar to contemporary clinical mortality prediction models such as MPM III.[15] In the current study, we explored this topic in greater depth by examining the marginal benefit of adding initial critical care treatments to simpler models derived from demographic characteristics and comorbidities. In addition to its superior discriminatory performance, we found that the final model had good calibration in medical patients but both overpredicted and underpredicted mortality in surgical patients. This finding may indicate that surgical sepsis is a very different clinical entity than medical sepsis and may therefore require its own mortality prediction model. Other contemporary ICU mortality prediction models have also performed poorly in selected subpopulations.[34, 35] In future work, one possible method for improving calibration in surgical populations would be to use a Cox regression model with stratum-specific hazard functions, as was done in the Framingham Heart Study, where the authors stratified 10-year predictions for heart attacks and coronary death by sex.[36]

Using both a sheaf coefficient analysis, which focuses on latent variables,[30] and an observed variable analysis based on difference in log likelihoods of models when groups of variables were excluded, we clearly showed the relative importance of early critical-care treatments in the model. A larger sheaf coefficient indicates greater relative importance in the model, and inclusion of critical care treatments administered early in the hospitalization resulted in a standardized coefficient that was approximately 30% greater than that of demographics, comorbidities, or infection source and type. Figure 4, which illustrates the uses of observed rather than latent variables, showed similar results

Although adjusting for treatments received at the time of admission is a novel approach to overcoming the limitations of administrative data, it has several shortcomings. It seems reasonable to assume that patients treated with vasopressors had severe hypotension or shock, and patients treated with mechanical ventilation were experiencing respiratory failure. Such assumptions are supported by the fact that our final model shows that the presence of mechanical ventilation or vasopressors increases the odds of mortality at least twofold. However, variation across physicians and institutions regarding the threshold for initiation of these treatments is still likely. Hospitals with high rates of mechanical ventilation may have sicker patients or may merely have a lower threshold for defining respiratory failure than that of hospitals with lower rates of mechanical ventilation. For this reason, models that use initial treatments to control for presenting severity may have limited use for profiling of hospital performance, because they may inadvertently reward hospitals with a more intensive practice style. When analyses are focused at the patient level, as is the case in studies of treatment effectiveness, the issue of variation in treatment thresholds is of less concern.

A second limitation is that the model we have presented requires detailed billing from the specific hospital day on which the service was provided, so it cannot be used with traditional administrative datasets such as Medicare claims data or the Nationwide Inpatient Sample. However, highly detailed billing data are now being shared by hundreds of hospitals in the Premier network and the University Health System Consortium. These data are widely available to investigators, meaning that the generalizability of this method will only improve. Third, the method we describe has been validated only in patients with sepsis, although the approach may be generalizable to other conditions, especially those that involve treatments provided in intensive-care settings. Fourth, the performance of the model is susceptible to secular changes in treatment. If mechanical ventilation and vasopressors are replaced by other therapies, the predictive value of these factors may change. However, existing severity-of-illness models have undergone periodic revisions to reflect changes in clinical practice as well. Fifth, we described this model's performance in only a development set, not a validation population. However, in a previous study, we validated this model against several clinical severity adjustment methods (e.g., MPM III, APACHE II) using data from Project IMPACT—a critical care database developed by the Society of Critical Care Medicine—and found that its performance was acceptable.[15]

In conclusion, we report that adding information about critical care treatments during the first days of hospitalization elicited substantial improvement in models that predict mortality in hospitalized sepsis patients. Until clinical data obtained from electronic medical records are more widely available, a method that uses only administrative data could act as a critical bridge, allowing for much-needed observational studies of treatment effectiveness in patients with critical illness.

ACKNOWLEDGEMENTS

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. METHODS
  5. RESULTS
  6. DISCUSSION
  7. ACKNOWLEDGEMENTS
  8. CONFLICT OF INTEREST
  9. REFERENCES
  10. APPENDIX: SEPSIS CODES
  11. APPENDIX: MIXED-EFFECTS LOGISTIC REGRESSION MORTALITY PREDICTION MODEL*

This study was conducted with funding from the Division of Critical Care and the Center for Quality of Care Research at Baystate Medical Center. Premier Healthcare Informatics, Charlotte, NC, provided the data used to conduct this study but had no role in its design, conduct, analysis, interpretation of data, or the preparation, review, or approval of the manuscript. The authors also thank Nicholas Hannon for his help in formatting the tables and the Reference section.

CONFLICT OF INTEREST

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. METHODS
  5. RESULTS
  6. DISCUSSION
  7. ACKNOWLEDGEMENTS
  8. CONFLICT OF INTEREST
  9. REFERENCES
  10. APPENDIX: SEPSIS CODES
  11. APPENDIX: MIXED-EFFECTS LOGISTIC REGRESSION MORTALITY PREDICTION MODEL*

Drs Lagu, Lindenauer, Rothberg, and Nathanson have no potential conflicts of interest. Dr Steingrub has previously received research grant support from and participates in the lecture bureau of Eli Lilly & Company. Dr Nathanson, through his company OptiStatim LLC, was paid by the investigators with funding from the Department of Medicine at Baystate Medical Center to assist in conducting the statistical analyses in this study. Dr Nathanson's company has a consulting agreement with the Cerner Corporation. Drs Lagu and Lindenauer had full access to all of the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. Drs Lagu, Lindenauer, Steingrub, and Rothberg conceived of the study. Dr Lindenauer acquired the data. Drs Lagu, Lindenauer, Rothberg, Steingrub, and Nathanson analyzed and interpreted the data. Dr Lagu drafted the manuscript. Drs Lindenauer, Rothberg, Nathanson, and Steingrub critically reviewed the manuscript for important intellectual content. Dr Nathanson carried out the statistical analyses.

KEY POINT

  • A sepsis mortality prediction model that incorporates information about the use of mechanical ventilation and vasopressors is superior to models that rely on demographic information and comorbidities.

REFERENCES

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. METHODS
  5. RESULTS
  6. DISCUSSION
  7. ACKNOWLEDGEMENTS
  8. CONFLICT OF INTEREST
  9. REFERENCES
  10. APPENDIX: SEPSIS CODES
  11. APPENDIX: MIXED-EFFECTS LOGISTIC REGRESSION MORTALITY PREDICTION MODEL*

APPENDIX: SEPSIS CODES

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. METHODS
  5. RESULTS
  6. DISCUSSION
  7. ACKNOWLEDGEMENTS
  8. CONFLICT OF INTEREST
  9. REFERENCES
  10. APPENDIX: SEPSIS CODES
  11. APPENDIX: MIXED-EFFECTS LOGISTIC REGRESSION MORTALITY PREDICTION MODEL*
  • 038 Septicemia

    • Excludes bacteremia (790.7)

    • 038.0 Streptococcal septicemia

    • 038.1 Staphylococcal septicemia

    • 038.2 Pneumococcal septicemia (Streptococcus pneumoniae septicemia)

    • 038.3 Septicemia due to anaerobes

      • Septicemia due to bacteroides

      • Excludes gas gangrene (040.0)

        • That due to anaerobic streptococci (038.0)

    • 038.4 Septicemia due to other gram-negative organisms

      • 038.40 Gram-negative organism, unspecified

        • Gram-negative septicemia NOS

      • 038.41 Hemophilus influenzae (H. influenzae)

      • 038.42 Escherichia coli (E. coli)

      • 038.43 Pseudomonas

      • 038.44 Serratia

      • 038.49 Other

    • 038.8 Other specified septicemias

      • Excludes septicemia (due to)

        • Anthrax (022.3)

        • Gonococcal (098.89)

        • Herpetic (054.5)

        • Meningococcal (036.2)

        • Septicemic plague (020.2)

    • 038.9 Unspecified septicemia

      • Septicemia NOS

      • Excludes bacteremia NOS (790.7)

    • 995.92 Severe sepsis

    • 790.7 Bacteremia

117.9 Disseminated fungal infection

112.5 Dissemintaed candidal infection

112.81 Disseminated fungal endocarditis

APPENDIX: MIXED-EFFECTS LOGISTIC REGRESSION MORTALITY PREDICTION MODEL*

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. METHODS
  5. RESULTS
  6. DISCUSSION
  7. ACKNOWLEDGEMENTS
  8. CONFLICT OF INTEREST
  9. REFERENCES
  10. APPENDIX: SEPSIS CODES
  11. APPENDIX: MIXED-EFFECTS LOGISTIC REGRESSION MORTALITY PREDICTION MODEL*
CovariateOdds ratiop > |z|
  • *

    Because of the high mortality rate (>10%), odds ratios may slightly overestimate the relative risk.

Demographics  
Age (in years)1.03 (1.03, 1.03)<0.001
Female1.08 (1.05, 1.12)<0.001
Race/Ethnicity  
White = referent1.00 
Black1.00 (0.95, 1.04)0.81
Hispanic0.97 (0.90, 1.05)0.42
Other/Unknown1.00 (0.95, 1.06)0.94
Marital status  
Married = referent1.00 
Single1.19 (1.14, 1.24)<0.001
Widowed1.11 (1.06, 1.15)<0.001
Separated/Divorced1.06 (1.01, 1.12)0.02
Other1.17 (1.08, 1.27)<0.001
Not recorded1.13 (0.98, 1.29)0.09
Insurance  
Medicare traditional = referent1.00 
Managed care0.95 (0.91, 1.00)0.04
Medicaid traditional1.10 (1.04, 1.17)0.002
Self-pay/Other1.17 (1.09, 1.26)<0.001
Medicare managed1.04 (0.97, 1.10)0.27
Commercial/Private1.01 (0.93, 1.10)0.74
Medicaid managed1.17 (1.02, 1.34)0.02
Clinical characteristics  
Comorbidities  
Hypertension0.72 (0.70, 0.75)<0.001
Diabetes (with or without complications)0.85 (0.82, 0.88)<0.001
Anemia (chronic and acute)0.67 (0.65, 0.69)<0.001
Chronic pulmonary disease0.93 (0.90, 0.96)<0.001
Congestive heart failure1.26 (1.22, 1.30)<0.001
Neurological disorders1.26 (1.22, 1.30)<0.001
Renal failure1.24 (1.19, 1.28)<0.001
Weight loss1.33 (1.28, 1.38)<0.001
Solid tumor without metastasis1.08 (1.04, 1.13)<0.001
Hypothyroidism0.88 (0.84, 0.93)<0.001
Depression0.72 (0.69, 0.76)<0.001
Peripheral vascular disease1.23 (1.17, 1.29)<0.001
Valvular disease0.95 (0.91, 1.00)0.07
Obesity0.81 (0.76, 0.87)<0.001
Metastatic cancer2.23 (2.11, 2.35)<0.001
Liver disease1.98 (1.86, 2.11)<0.001
Psychoses0.70 (0.65, 0.75)<0.001
Alcohol abuse1.26 (1.18, 1.35)<0.001
Rheumatoid arthritis/Collagen disease1.03 (0.95, 1.11)0.50
Lymphoma1.50 (1.39, 1.62)<0.001
Drug abuse0.74 (0.67, 0.82)<0.001
Pulmonary circulation disease1.05 (0.96, 1.14)0.30
Paralysis0.93 (0.88, 0.99)0.02
Peptic ulcer disease0.82 (0.73, 0.91)<0.001
AIDS1.34 (1.04, 1.73)0.03
Site of infection  
Urinary tract (primary or secondary)0.72 (0.69, 0.75)<0.001
Lung (primary or secondary)1.60 (1.53, 1.67)<0.001
Abdominal (primary or secondary)1.36 (1.29, 1.42)<0.001
Blood and not urinary, lung, or abdominal0.33 (0.30, 0.37)<0.001
Other (e.g., skin/bone) or unknown1.56 (1.49, 1.65)<0.001
Infection type  
Unknown = referent1.00 
Gram-positive0.71 (0.68, 0.74)<0.001
Gram-negative0.67 (0.64, 0.69)<0.001
Other (fungal, viral, mixed)0.97 (0.88, 1.08)0.61
Medical (vs. surgical)1.31 (1.25, 1.39)<0.001
Surgical type  
General or gastrointestinal1.03 (0.96, 1.10)0.44
Cardiac or vascular1.08 (0.96, 1.22)0.19
Orthopedic0.80 (0.72, 0.89)<0.001
Other0.87 (0.80, 0.95)0.002
Primary versus secondary diagnosis of sepsis0.95 (0.93, 0.98)0.002
Treatments (by Day 2)  
Intensive care unit (at least 1 day)1.28 (1.24, 1.33)<0.001
Mechanical ventilation2.55 (2.45, 2.65)<0.001
Vasopressor2.15 (2.07, 2.23)<0.001