INTRODUCTION
Highly active antiretroviral therapy (HAART) has been one of the most
successful medical innovations in treating AIDS. It has led to a decrease
in the incidence of opportunistic infections, a marked decrease in AIDS
mortality, and an increase in AIDS prevalence due to longer survival
(Palella et al., 1998; Moore
& Chaisson, 1999; Carrasco & Tyring,
2001).
Along with these benefits, HAART regimens can also produce both
short-term adverse effects and long-term complications. These include
response failure due to drug resistance, systemic symptoms,
hepatotoxicity, metabolic and lipodystrophic disorders, and insulin
resistance associated with coinfection with hepatitis C virus. Poor
adherence or simply advanced disease also can lead to HAART response
failure (Detels et al., 1998; d'Arminio Monforte et al., 1998; Gallant, 2000; Max & Sherer,
2000; Jain et al., 2001; Reisler, 2001; Duong et al.,
2001). In a recent study of patients with advanced AIDS who were
admitted to a long-term care facility, we reported a 15% mortality within
3 months of admission (Brechtl et al., 2001).
Patients who died had been maintained on HAART regimens until death. Of
the patients who survived longer than 3 months, 25% could not tolerate a
HAART regimen.
These results highlight the challenge for patients and physicians who
must make difficult treatment decisions about initiating or continuing
HAART in advanced AIDS. Identifying predictors of death in this population
could provide valuable guidance. A reliable predictive model could help
optimize patient care by alleviating burdensome, perhaps ineffective,
medication loads while providing a more palliative, quality-of-life
approach.
Studies of prognostic indicators of disease progression and death are
abundant for the pre-HAART era and earlier stages of infection (Friedland et al., 1992; Mocroft et
al., 1995; Hanson et al., 2002).
Specifically, plasma HIV-1 RNA viral load levels and CD4+ cell counts have
demonstrated high predictive value (Mellors et al., 1996, 1997; O'Brien et al., 1996). Unfortunately, these
markers are less predictive in patients with advanced AIDS, irrespective
of antiretroviral regimens (Spino et al., 1997;
Cozzi Lepri et al., 1998; HIV
Surrogate Marker Collaborative Group, 2000). The purpose of the
present study, therefore, was to identify the best predictors of death
within 6 months in patients with advanced AIDS and to develop a prognostic
model that could be used to guide patient management decisions.
METHODS
Setting and Subjects
This retrospective cohort study examined a consecutive series of AIDS
patients admitted to Terence Cardinal Cooke Health Care Center (the
Center) between January 1 and December 31, 1999. The Center, which is
located in Manhattan, is a 729-bed skilled nursing facility that includes
a 156-bed AIDS Discrete Unit (the Unit). Patients in the Unit are provided
with full-time medical supervision; their health is too poor or unstable
for them to be discharged. The goal of the Unit is for patients to improve
clinically receiving aggressive and supervised medical care as tolerated.
Patients are not admitted specifically for end-of-life care, although such
care is anticipated for patients who decline during their course of
treatment. All study subjects had been referred to the Center from acute
care hospitals because of advanced illness requiring ongoing, skilled
nursing care; thus investigators refer to these patients as having
“advanced AIDS.”
Data Collection
We abstracted demographic, clinical, laboratory, and outcome data from
medical records using a standardized data collection instrument.
Demographic data included age, gender, and ethnicity. Clinical data
included height, weight, comorbidites (e.g., cancer, nephropathy,
thrombocytopenia), HIV transmission group (heterosexual, homosexual, IV
drug use), and CDC AIDS Category and AIDS-defining conditions (as
specified by the 1993 revised CDC-Defined Classification for HIV
Infection; Centers for Disease Control and Prevention,
1992). Laboratory data included hemoglobin, hematocrit, white blood
count, lymphocyte count, platelet count, albumin, CD4 count, HIV-1 RNA
viral load, serum ALT, and serum AST. We noted whether a patient had died
within 6 months of admission and, for patients who were discharged before
6 months had elapsed, we requested a search from the National Death Index
(Centers for Disease Control and Prevention,
2005) to ascertain those patients' survival status at 6
months. Given that there is an estimated 1-year lag between a death and
its entry into the National Death Index, data were requested from the
National Death Index in 2001.
This study was approved by the Center's Institutional Review
Board.
Analysis
Univariate analyses utilized chi-square tests for categorical
variables and t tests for continuous variables to evaluate the
association between each independent variable (demographics, clinical
descriptors, laboratory values, and several constructed variables) and the
outcome variable (death at 6 months). Constructed variables included
number of comorbidities, number of AIDS-defining conditions, and
“weight deviation percent” (percentage difference between a
patient's actual and ideal body weight at the time of admission). For
a man, the ideal weight was calculated as 106 lbs. for the first 5 ft. of
height plus 6 lbs. for each additional inch, and for a woman, it was
calculated as 100 lbs. for the first 5 ft. plus 5 lbs. for each additional
inch (Wilkes, 1999). In addition, the
distributions of all continuous laboratory variables between survivors and
nonsurvivors were examined to determine whether using cutoff points to
create dichotomous variables would yield satisfactory predictions; such
variables produce more intuitive, easy-to-apply prognostic guidelines.
After completing the univariate analyses, we conducted a multiple
logistic regression analysis to select a combination of variables to
include in a final prognostic model. This analysis considered as potential
predictors all independent variables having a p value <0.25 in
the univariate analysis (Hosmer & Lemeshow,
2000). Once variables for the final model had been selected,
continuous variables were evaluated to confirm the assumption of linearity
in the logit. Variables were also evaluated for interaction, and any
interaction terms that rendered a p value <0.05 were added to
the final model. The Hosmer–Lemeshow chi-square was used to assess
model fit (Hosmer & Lemeshow, 2000). All
analyses were performed using STATA data analysis software (StataCorp, 1997).
RESULTS
A total of 152 patients with AIDS were admitted to the Center in 1999.
Over 82% of patients were between the ages of 30 and 60 and 76% were male.
In terms of race, 53.9% of patients were black, 28.3% Hispanic, and 17.8%
white. For those patients for whom HIV transmission data were available
(76.3%), the majority had become infected with HIV through intravenous
drug use (59.5%), a third by heterosexual contact (30.2%), and the
remainder by homosexual contact (10.3%). All patients had advanced AIDS in
that they fell into CDC AIDS Category A3 or above as follows: A3, 5.9%;
B3, 15.1%; C1–C3, 78.9%. Forty-eight percent of patients were on
HAART at the time of admission to the Center. A comparison of these
variables between those 61 patients who died within 6 months of admission
(nonsurvivors) and those 91 patients who survived are presented in Table 1. The results of the univariate analyses for
the clinical and laboratory data are presented in Tables 2 and 3. (In Table 3, median values were used for the comparison of
those variables with skewed distributions that were statistically
significant.)
Overview of study patients (N = 152), by survival status at 6
months
AIDS-defining conditions and comorbidities in study patients (N =
152), by survival status at 6 months
Laboratory values in study patients (N = 152), by survival status at 6
months
In comparing the distribution of continuous laboratory variables
between survivors and nonsurvivors, the following dichotomous variables
were created (with cutoff points shown in parentheses): hemoglobin (10
g/dl), hematocrit (30%), lymphocytes (1000/mm3),
albumin (3.0 g/dl), CD4 count (50/mm3), platelet count
(100,000/mm3), and serum AST (250 IU/L). Survivors and
nonsurvivors had similar distributions of age, serum ALT, HIV viral load,
and white blood cell count, so cutoff points were not established for
these variables. To facilitate its use in a prognostic guideline, we
converted weight deviation percent to a categorical variable with six
ranges: >0, 0 to −9%, −10% to −19%, −20% to
−29%, −30% to −39%, and −40% to −49%.
The final multivariate prognostic model included albumin (OR 2.5
[1.1, 5.6], p = 0.03), ideal body weight deviation (OR
1.7 [1.3, 2.3], p < 0.01), and number of
comorbidities (OR 2.0 [1.3, 3.0], p ≤ 0.01; see
Fig. 1). Thirty-seven percent of survivors
versus 74% of nonsurvivors had an albumin level ≤ 3.0 g/dl
(p < 0.01). Almost half of nonsurvivors (49%) fell below their
ideal body weight by 20% or more as compared to 23% of survivors
(p < 0.01). Finally, 33% of survivors versus 56% of
nonsurvivors had two or more comorbidities (p < 0.01). Of the
34 nonsurvivors with two or more comorbidities, 94% had been diagnosed
with anemia, 68% with thrombocytopenia, 35% with hepatic failure, and 24%
with nephropathy. The predicted probability of death within 6 months based
on this prognostic model is detailed in Figure
2.
Percent of patients who died within 6 months, according to albumin
level, deviation from ideal body weight, and number of comorbidities
(N = 152).
Predicted probability of death within 6 months, based on a logistic
prediction model that combines albumin level, deviation from ideal body
weight, and number of comorbidities. Degree of shading indicates the
predicted probability of death within 6 months (<0.50, 0.50–0.70,
0.70–0.90, or >0.90) associated with each combination of
predictor values.
Two variables that were significantly associated with mortality in the
univariate analyses (serum AST and hepatic failure) were not included in
the final multivariate model because their estimates were unstable, as
indicated by extremely wide confidence intervals. The
Hosmer–Lemeshow chi-square (2.6; p = 0.96) revealed good
model fit. Because no interaction terms showed statistical significance,
none were included in the final model.
Using the data that generated the model, a probability cutoff value of
0.50 for predicting death within 6 months yielded 67% sensitivity and 80%
specificity. Raising the probability cutoff to 0.70 decreased sensitivity
to 33% and increased specificity to 96%. A higher cutoff increasing
specificity (reducing false-positive classifications) at the expense of
decreasing sensitivity (increasing false-negative classifications) may
provide an acceptable trade-off in guiding decision making with such a
model, assuming patients and physicians would prefer to err on the side of
aggressive treatment (Table 4).
Performance of prognostic model using probability cutoff values 0.50
or 0.70 (N = 152)a
DISCUSSION
This study identified a combination of three descriptors—albumin
level, deviation from ideal body weight, and number of
comorbidities—that were good predictors of death within 6 months in
patients with advanced AIDS. Our findings are consistent with previous
studies that have shown decreased albumin to be an independent predictor
of death in patients with AIDS (Turner et al.,
1996; Sabin et al., 2002; Wheat et al., 2002), and under-nutrition—one of
the major causes of low albumin— to be common in these patients
(Salomon et al., 2002). Furthermore, deaths
attributable to comorbidities are increasing with the improved management
of this disease. According to a 2000 report, for example, deaths from
opportunistic infections decreased significantly between 1995 and 1999
(from 20.9% to 8.5%; p < 0.05), whereas deaths from hepatic
failure increased significantly (from 0% to 6%; p < 0.05;
Sansone & Frengley, 2000).
It is possible that such differences could alter the risk of death
predicted by our model. It is important to emphasize, therefore, that the
model in Figure 2, which is based on data for a
single long-term care facility, is intended for patients with advanced
disease based on CDC AIDS category and the need for continual skilled
nursing care—that is, patients similar to the population from which
the model was derived. The demographic profile of our patients resembles
that of AIDS populations in other large metropolitan areas in the
southern, north-central, and northeastern regions of the United States
(Centers for Disease Control and Prevention,
2000; Welch et al., 2002); however, the
clinical profile of our patients differs in some respects from other
series of patient with AIDS (Welch et al.,
2002). Therefore, until the model has been independently validated
in other groups of patients with advanced AIDS, its predictions must be
considered with caution.
Based upon input from patients and family, other research has
identified certain key components that result in what is perceived as
high-quality end-of-life care. These include interactions with health care
professions to provide physical comfort and emotional support to patients
and their families as well as information to facilitate shared decision
making (Teno et al., 2004). To meet these
expectations and provide such care, it is necessary to equip patients and
families with the tools to share in such delicate and yet crucial matters.
This study has taken an early step to develop such a tool for patients
with advanced AIDS. Numerous studies have developed end-of-life decision
aids for terminally ill cancer patients, but few have looked to provide
similar help to those living with AIDS.
Early treatment decisions can have a profound impact on the ultimate
course of HIV disease and a patient's quality of life. We are now
capable of turning this disease into a manageable condition for many,
though not all, patients through informed and judicious use of drug
therapy. The prognostic model presented here may be useful as a building
block in predicting short-term survival in patients with advanced AIDS. It
should, at minimum, heighten sensitivity to the appropriateness of
withholding or withdrawing HAART in certain patients.
ACKNOWLEDGMENT
This study was supported in part by Award 2D54HP00022 from the Health
Resources and Services Administration.
