INTRODUCTION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Pneumocystis carinii pneumonia (PCP) is a common reason for hospitalization of patients with acquired immunodeficiency syndrome (AIDS). Hospital mortality rates for patients with human immunodeficiency virus (HIV)-associated PCP have ranged from 60% in the late 1980s to 10% in recent studies (1- 4). There appears to be large variations in the quality of care for patients with PCP, with the implication that hospitals with worse quality of care and outcomes may be providing poorer quality of care (5). To explain differences in hospital mortality rates, it is necessary to develop an effective means of controlling for severity of illness at admission when attempting to assess correlations between quality of care and in-hospital mortality rates. While many studies have focused on assessments of quality of care, severity of illness, and outcomes for HIV-associated PCP, these studies do not include individuals who receive newer antiretrovirals (8).

In the early highly active antiretroviral therapy (HAART) era, among patients with CD4 lymphocyte counts less than 100 cells/mm³, the incidence of major opportunistic infections including PCP, Mycobacterium avium complex (MAC) disease, and cytomegalovirus retinitis declined markedly (14). As the rates of prophylaxis remained constant over this period, the decrease in opportunistic infections is believed to be secondary to the increased use of protease inhibitors and combination antiretroviral therapy (14). The staging system reported here is based on a review of patients admitted with HIV-associated PCP between 1995 and 1997, a time period that included the early adoption of non-nucleoside reverse transcriptase inhibitors and protease inhibitors as HIV therapy. The staging system provides a model for evaluating severity of illness for individuals with this opportunistic infection in an era where clinical manifestations of HIV infection and outcomes differ dramatically from those seen the late 1980s. As such, it may serve as a useful adjunct in subsequent studies that evaluate quality of care and in-hospital mortality rates.

	METHODS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Sampling of Patients and Hospitals

Medical records were abstracted for HIV-infected patients with confirmed or probable PCP who received a portion of their medical care in a study institution between January 1, 1995 and December 31, 1997. Study institutions included 78 public and private hospitals in seven metropolitan areas of the countryNew York, Los Angeles, Miami, Chicago, Seattle, Nashville/Memphis, and Phoenix/Tucson. For 1996 to 1997, New York had the largest number of reported AIDS cases in the country, with Los Angeles second, Miami fifth, and Chicago eighth (15). New York and Miami have large numbers of HIV-infected drug users while the majority of HIV-infected individuals in Los Angeles and Chicago are males who have sex with males.

We have previously described our sampling methodology that employed two hierarchical levels: hospitals within cities and patients within hospitals (16, 17). In each hospital, a random sample of patients was selected such that the proportion of patients from each hospital relative to the overall number of charts reviewed in a given city was roughly proportional to the square root of the individual hospital caseload divided by the total hospital caseload for the given city. Cases were also stratified by whether patients were discharged alive or dead in order to approximate the actual in-hospital mortality for each hospital.

Retrospective chart reviews of hospitalized patients identified by medical information system analysts at study hospitals were performed by trained registered nurses who were experienced with AIDS patients and utilization review. All discharges including International Classification of Diseases, Ninth Revision (ICD-9) codes for PCP (136.3) and HIV-related disease (042-044) were screened. Patients were included in the study if they had microbiological confirmation of PCP or physician notes indicating that PCP rather than other causes accounted for their pulmonary process, and if their age was at least 18 yr. Patients were excluded if they received medical care for the current episode of PCP at another hospital, left the hospital against medical advice, or were admitted for other conditions or diagnostic bronchoscopy only.

Data abstracted included: patient sociodemographics; insurance status; HIV and non-HIV-related comorbid conditions; cigarette, alcohol, and drug use history; preadmission use of antiretroviral and prophylactic medications; T-cell count and HIV viral load titer; initial vital signs, arterial blood gas, and laboratory data; treatment medications received; principal and secondary diagnostic and procedure codes; length of stay; discharge and 30-d postadmission status. Patients were classified as wasting if on Day 1 or 2 of the admission a physician noted that the patient exhibited wasting, cachexia, a greater than 20% weight loss, and/or a greater than 20 pound weight loss. Data quality was maintained by overreading abstraction forms by two physicians trained in quality assurance for the project. Less than 1% of entries were categorized as possibly inaccurate by the physician overreaders.

Statistical Analysis

Univariate associations between measured attributes and mortality status were evaluated using optimal discriminant analysis, and a multivariate nonlinear model was obtained via hierarchically optimal classification tree analysis (CTA) (18). For all analyses, jackknife validity analysis was conducted to assess the potential generalizability of the findings. For characteristics that were statistically significant in jackknife validity analysis, the optimal cut-points identified are expected to cross-generalize at the jackknife estimate if they are used to classify independent random patient samples. In contrast, if statistically significant characteristics were unstable in jackknife validity analysis, these characteristics are expected to be predictive of mortality, but using model cut-points that are different than those identified. A sequentially rejective Sidak Bonferroni-type multiple comparisons procedure was used to ensure an experimentwise Type I error rate of p  0.05 (18).

	RESULTS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

There were 1,660 cases with confirmed or suspected HIV-associated PCP that met our eligibility criteria and were used in our analyses. Sociodemographic and mean laboratory values for the sample are given in Table 1. Forty-two percent of the cases were white, with a mean age of 38.2 yr (± 8.1 SD) and an overall in-hospital mortality rate of 11.3%.

Univariate associations between preadmission characteristics and in-hospital mortality are displayed in Table 2 for the total sample. Nine characteristics were found to have significant univariate associations with mortality status in both training (total sample) and jackknife validity analysis. These characteristics included illicit drug use, alveolar-arterial oxygen gradient (AaPO₂), respiratory rate, prior MAC prophylaxis, prior AIDS diagnosis, white blood cell count, creatinine, wasting, and neurological symptoms. Two characteristics, age and prior PCP prophylaxis, had statistically marginal univariate associations with mortality. There were eight characteristics that had significant univariate associations with mortality in training analysis, but in jackknife validity analysis the associations were not significant. These eight characteristics included PO₂, O₂ saturation, CD4 lymphocyte count, systolic blood pressure, total lymphocyte count, hemoglobin, hematocrit, and albumin.

Among the preadmission characteristics analyzed (Table 2), wasting had the greatest associated effect strength that was stable in jackknife validity analysis and was selected as the initial node in the hierarchically optimal CTA model (Figure 1). The subsequent nodes in the CTA model are the AaPO₂ and albumin level. The CTA model classifies patients with a given attribute profile as either dead or alive. For example, consider a patient who had wasting, initial AaPO₂ of 60 mm Hg, and albumin level of 3.0 g/dl on admission. Because the patient had wasting, following the right branch at the first node is appropriate. Because the patient had an AaPO₂ > 52.6 mm Hg, following the right branch at the second node is appropriate. Finally, because the patient had an albumin level > 2.55 g/dl, following the right branch at the third node is appropriate. Thus, the patient is classified by the CTA model as alive. Forty-six of the 60 actual patients who were classified into this same endpoint via this particular branch of the classification tree were classified correctly. Therefore, the probability of being alive for a patient in this particular group is 76.7%. The empirical probability of death is 1.0 minus 0.767, or 0.233, resulting in a predicted mortality rate of 23.3% for patients classified into this group. If the patient's albumin level had been  2.55 g/dl, then the patient would have been classified by the CTA model as "dead." Twenty-eight of 57 actual patients classified into this latter endpoint were correctly classified as "dead," so the predicted mortality rate is 49.1% for patients in this group.

View larger version (23K): [in this window] [in a new window]   Figure 1.   Schematic illustration of the CTA model for predicting mortality. Circles represent nodes, arrows indicate branches, and rectangles represent prediction endpoints (as is indicated, the model classifies patients having a given attribute profile as either Dead or Alive). Numbers/words adjacent to arrows indicate the value of the cut-point or category for each node. Fractions located under each prediction endpoint give both the number of correct classifications at the endpoint (numerator), and the total number of observations classified at the endpoint (denominator). Numbers in parentheses, adjacent to the fractions, give the predictive value for the endpoint. Numbers in brackets indicate the mortality rate at the node, and numbers beneath the brackets indicate the mean and standard deviation of the bootstrap predictive value for the model endpoint. Numbers underneath nodes give the generalized Type I error for the node.

The CTA model classified 1,194 out of the 1,660 patients (71.9%) as dead or alive. Among those patients who were classified in training analysis, 862 (72.2%) were classified correctly (Table 3). Nonclassified patients were most commonly missing data on AaPO₂ or albumin level. The CTA model's classification accuracy, sensitivity, and predictive value were stable in bootstrap validity analysis. For sensitivity, this CTA model represents a 33.1% improvement over the theoretical classification possible by chance alone, which is a moderate effect.

The CTA model was used to develop a staging system for predicting in-hospital mortality for HIV-associated PCP (Table 4). Because the prediction endpoints from the CTA model (Figure 1) are ordered from lowest to highest mortality rate, stage is an ordinal scale of severity of illness on which increasing integers reflect an increasing likelihood of in-hospital mortality. Among patients without wasting, those with an AaPO₂ greater than 53.4 mm Hg (Stage 3) have a fourfold higher likelihood of in-hospital mortality compared with patients with an AaPO₂ of 53.4 mm Hg or less (Stage 1), with associated mortality rates of 16.1% versus 3.7%. Stages 4 and 5 include patients with both wasting and an AaPO₂ greater than 52.6 mm Hg and are distinguished by albumin level. Patients with an albumin level greater than 2.55 g/dl (Stage 4) have a twofold lower likelihood of in-hospital mortality compared with patients with an albumin level of 2.55 g/dl or less (Stage 5), with associated mortality rates of 23.3 versus 49.1%. Patients classified in Stage 5 have a 13-fold higher likelihood of in-hospital mortality compared with patients in Stage 1, with associated mortality rates of 49.1 versus 3.7%.

	DISCUSSION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

There were 1,660 cases with confirmed or suspected HIV-associated PCP from the early HAART era and immediately preceding widespread use of highly active antiretroviral therapy included in our analysis. The overall in-hospital mortality in our cohort was 11.3%, in contrast to the twofold higher mortality rates reported from the cohorts of PCP patients evaluated in the 1980s (16, 21). A staging system predicting in-hospital mortality was developed using clinical and laboratory data that were readily available at the time of admission. This staging system stratifies patients with PCP into five risk classes associated with in-hospital mortality rates ranging from 3.7 to 49.1%.

The variables used in our staging systemwasting, AaPO₂, and serum albumin levelreflect both the severity of pulmonary illness and the severity of comorbid medical illnesses. The AaPO₂ has been included in all prior severity models for PCP and is a direct measure of the severity of the acute pulmonary process. Other clinical factors included in prior models were body mass index, a measure of nutritional status, and total lymphocyte count, a measure of chronic immune compromise. Data on total lymphocyte counts were obtained but did not add to the predictive ability of the current staging system. In the current study, we chose to use wasting rather than calculating body mass index (defined as height divided by weight squared) because it was readily available in the patient records whereas heights were not recorded in most patient records. Serum albumin level, like body mass index in the prior model, represents a measure of a patient's nutritional status. With the advent of combination antiretroviral therapy and protease inhibitors, patients infected with HIV are living longer and often develop end-stage complications such as wasting (22). Opportunistic infections such as PCP occur late in the course of HIV infection and it seems reasonable that factors related to comorbidities that have an impact on general health, nutritional status, and immune function (as reflected by lower albumin levels and wasting) would be important factors in predicting survival from these opportunistic infections.

Prior studies have described prognostic schemes for predicting mortality for patients with HIV-related PCP in the time period when zidovudine (AZT) was the only antiretroviral medication. Based on review of 576 patients at 56 hospitals between 1987 and 1990, a four-stage severity system predicting in-hospital death was developed (16). This four-stage system was based on AaPO₂, total lymphocyte count, and body mass index. A model based on 159 PCP patients treated in the 1980s at Saint Mary's Hospital in London found AaPO₂, history of prior AIDS, and serum hemoglobin level to be independent predictors of mortality (21). A third multivariate prediction model was developed prior to AZT and included the AaPO₂, hemoglobin level, and patient age (2). Other studies in the pre-AZT era included smaller numbers of patients and reported univariate associations of clinical factors with mortality. One study of seriously ill patients admitted to intensive care units found that high lactate dehydrogenase (LDH) levels and modified multisystem organ failure scores predicted poor outcome (23). In a retrospective analysis of 58 first episodes of HIV-associated PCP without prophylaxis, hemoglobin less than 10 g/dl, albumin less than 3 g/dl, and gamma-globulin level less than 1.2 g/dl were found to be associated with a 14-fold increased risk of in-hospital mortality (4). For predicting in-hospital mortality resulting from AIDS-related PCP, the current model from data for the period 1995 to 1997 represents a 4% to 56% improvement in sensitivity effect strength over prior CTA-based models that evaluated PCP during 1987 to 1990 (20), and a 231% or better improvement in sensitivity effect strength versus previous traditional linear and nonlinear models for the years 1987 to 1990 (16).

The predictive ability of our staging system compares favorably with previously published systems from the 1980s and early 1990s (2, 4, 8, 16, 20). This new model performed well in bootstrap validity analysis implying that the model is expected to cross-generalize if used to classify independent random patient samples. The CTA model returned a 52% greater effect size than the best prior published model predicting in-hospital mortality for HIV-associated PCP. The information needed to apply our staging system is easily obtained from most medical records of patients with HIV-associated PCP. The cut-points reported are exact and were designed to maximize the predictive ability of the model. The clinical use of our model may require the use of less stringent cut-points that will diminish the classification accuracy of the model. The staging system should be useful as a means to adjust for the severity of illness at hospital admission when comparing in-hospital mortality rates between hospitals and cities. The ability to adequately adjust for the severity of illness of patients is essential to the utilization of in-hospital mortality rates as a valid measure of hospital quality of care for PCP patients.

There are several limitations to our study. Because we used chart reviews from 78 hospitals to obtain our data, information on important clinical factors such as LDH level was not collected primarily because normal values varied between hospitals and over time. During the study period 1995-1997, the normal range for serum LDH differed between individual study hospitals and the normal range changed within individual hospitals as well (11). Therefore, serum LDH, a laboratory value commonly used in severity of illness measures from single hospital studies, was not included in our final model because of concerns over adequate standardization of LDH levels. We were not able to reliably evaluate the presence or absence of additional pulmonary infections from patient records. We were able to evaluate bronchoscopy use and in contrast to our findings from 1987-1990, bronchoscopy use was not a significant predictor of mortality in the current study. Becasue our staging system is based on inpatients, it has limited applicability to outpatients with PCP. All of our patients are from major metropolitan areas and it may be necessary to validate this staging system on a group of patients from rural areas as well. Most importantly, wasting was noted based on physician records using this term. As noted by others, the clinical manifestations of HIV wasting are extremely variable (22).

In conclusion, we developed a new staging system for predicting in-hospital death for patients with HIV-associated PCP in the early HAART era. The system was based on easily obtained clinical information for PCP patients treated at 78 private and public hospitals in seven major geographic areas. The staging system provides a simple method by which to classify patients into five risk classes that correlate to in-hospital mortality and provides a useful tool in evaluating variation in hospital mortality rates. This new staging system may assist with severity of illness adjustment investigation of variations in quality of care and outcomes for hospitalized patients with HIV-associated PCP.