David R. Mehr, Ellen F. Binder, Robin L. Kruse, Steven C. Zweig, Richard Madsen, Lori Popejoy, Ralph B. D'Agostino. Predicting Mortality in Nursing Home Residents With Lower Respiratory Tract InfectionThe Missouri LRI Study. JAMA. 2001;286(19):2427–2436. doi:10.1001/jama.286.19.2427
Author Affiliations: Center for Family Medicine Science, Department of Family and Community Medicine, School of Medicine (Drs Mehr, Kruse, and Zweig), Department of Statistics (Dr Madsen), and Sinclair School of Nursing (Ms Popejoy), University of Missouri, Columbia; Department of Internal Medicine, Division of Geriatrics and Gerontology, Washington University School of Medicine, St Louis, Mo (Dr Binder); and Department of Mathematics and Statistics, Boston University, Boston, Mass (Dr D'Agostino).
Context Lower respiratory tract infection (LRI) is a leading cause of mortality
and hospitalization in nursing home residents. Treatment decisions may be
aided by a clinical prediction rule that identifies residents at low and high
risk of mortality.
Objective To identify patient characteristics predictive of 30-day mortality in
nursing home residents with an LRI.
Design, Setting, and Patients Prospective cohort study of 1406 episodes of LRI in 1044 residents of
36 nursing homes in central Missouri and the St Louis, Mo, area between August
15, 1995, and September 30, 1998.
Main Outcome Measure Thirty-day all-cause mortality.
Results Thirty-day mortality was 14.7% (n = 207). In a logistic analysis, using
generalized estimating equations to adjust for clustering, we developed an
8-variable model to predict 30-day mortality, including serum urea nitrogen,
white blood cell count, body mass index, pulse rate, activities of daily living
status, absolute lymphocyte count of less than 800/µL (0.8 × 109/L), male sex, and deterioration in mood over 90 days. In validation
testing, the model exhibited reasonable discrimination (c = .76) and calibration (nonsignificant Hosmer-Lemeshow goodness-of-fit
statistic, P = .54). A point score based on this
model's variables fit to the entire data set closely matched observed mortality.
Fifty-two percent of residents had low (score of 0-4) or relatively low (score
of 5-6) predicted 30-day mortality, with 2.2% and 6.2% actual mortality, respectively.
Conclusions Our model distinguishes nursing home residents at relatively low risk
for mortality due to LRI. If independently validated, our findings could help
physicians identify nursing home residents in need of different therapeutic
approaches for LRI.
Pneumonia and the somewhat broader category of lower respiratory tract
infection (LRI), which includes pneumonia, bronchitis, and tracheobronchitis,
are the leading causes of mortality and hospitalization among nursing home
In recent outcome studies, 30-day mortality from pneumonia or LRI ranged between
11.4% and 30%.5- 9
Clinical findings consistently associated with mortality have included functional
dependence (defined by activities of daily living [ADLs]) and elevated respiratory
Because many nursing home residents are chronically ill and near the
end of life, the first step in making treatment decisions should be determining
appropriate therapeutic measures (eg, aggressive care, limited curative treatment,
or strictly palliative care). For strictly palliative care, maximizing comfort
should guide treatment. Otherwise, clinicians need to determine illness severity
to decide on specific treatment and whether residents should be treated in
the nursing home or hospital. Residents at low risk for mortality may be excellent
candidates for nursing home management, which may prevent complications from
hospitalization.12,13 For community-acquired
pneumonia, the validated Pneumonia Severity Index (PSI) provides guidance.14 However, because of its broad scope, most nursing
home residents would be classified in its 2 highest-risk categories.
In a large prospective sample of nursing home residents with LRI, we
derived and validated a new predictive model to better distinguish residents
at low risk of dying. We chose 30-day mortality as the most useful outcome
for considering treatment decisions, which is consistent with the approach
of other researchers.14
We identified participants from 36 nursing homes in central Missouri
and the St Louis, Mo, area. Facility characteristics were similar to 1997
national averages.15 For example, 64% were
for profit vs 67% nationally. Forty-seven percent of the facilities had fewer
than 100 beds, 44% had between 100 and 199, and 8% had 200 or more, whereas
facilities nationwide had 50%, 42%, and 8%, respectively. Thirty-one percent
of the facilities were in nonmetropolitan areas vs 38% nationally.
We chose to study mortality from LRI rather than pneumonia to make our
findings most relevant to nursing home practice. Physicians caring for nursing
home patients frequently do not obtain chest radiographs, and the clinical
distinction between bronchitis or tracheobronchitis and pneumonia is difficult,
even though the conditions are pathologically distinct. Thus, LRI includes
pneumonia and other LRI (BOX 1).
The definition is a modification of a surveillance
definition for long-term care facilities.16
An LRI was defined to include either pneumonia or other LRI. Both of
the following criteria were required for pneumonia:
Interpretation of a chest radiograph as demonstrating pneumonia or probable pneumonia. If a previous radiograph exists for comparison, the infiltrate should be new.
At least 2 of the LRI symptoms and signs below are present.
All 3 of the following criteria were required for other LRI (bronchitis, tracheobronchitis):
Pneumonia as defined above is absent or no chest radiograph is available.
At least 3 of the LRI signs and symptoms below are present.
In the presence of chronic obstructive pulmonary disease or congestive heart failure, additionally, the resident must have a temperature of ≥38°C for the illness to qualify as an LRI.
LRI symptoms and signs used in the definition:
New or increased cough
New or increased sputum production
Pleuritic chest pain
New or increased physical findings on chest examination (rales, rhonchi, wheezes, bronchial breathing)
One of the following indications of change in status or breathing difficulty: new/increased shortness of breath, or respiratory rate greater than 25/min, or worsening mental or functional status (significant deterioration in the resident's cognitive status or in the resident's ability to carry out the activities of daily living, respectively).
*Based on the statement of a consensus development conference concerning
infection-surveillance definitions for long-term care facilities.16 We modified the definition to explicitly exclude residents with chronic obstructive pulmonary disease or congestive heart failure who lacked either a fever or probable pneumonia on chest radiograph to avoid
including congestive heart failure or chronic obstructive pulmonary disease
exacerbations as an LRI.
Project nurses called or visited facilities at least 6 days per week
to identify residents who had respiratory (eg, cough, sputum production) or
nonspecific (eg, fever, acute confusion) symptoms compatible with an LRI.
The nurses were also available by pager, and facility staff and physicians
were encouraged to report ill residents at other times. Under a physician-authorized
protocol, residents with such symptoms received a focused history and physical
examination by a trained project nurse within 24 hours and usually on the
same day. Most evaluations included a chest radiograph, complete blood count,
and a chemistry panel. Project nurses predominantly had advanced-practice
education or extensive clinical experience and training in physical assessment.
Since evaluations were authorized by attending physicians, who also received
clinical information regarding each case, they were considered part of appropriate
care. Therefore, institutional review boards at each institution approved
a simplified consent process using a simple acceptance or refusal of the evaluation
as part of medical care. Potential cases were identified from August 15, 1995,
through September 30, 1998. However, all facilities were not involved until
December 1997. Additional details of resident identification and evaluation
are described elsewhere.17
Of the 4959 illness episodes reported by nursing homes, project nurses
performed 2592 evaluations to determine whether to include the episode in
the study. We did not evaluate (hereafter excluded) residents who accounted
for a total of 1191 episodes because they did not have lower respiratory or
systemic symptoms or signs except for cough (Figure 1). We also excluded 1176 episodes in which residents were
(1) ineligible because they were younger than 60 years old, not in the facility
at least 14 days, or had taken an antibiotic in the last 7 days for a previous
LRI; (2) not appropriate for an outcomes study because they had a "no antibiotics"
order, were not expected to live more than 30 days, were enrolled in hospice,
or had acquired immunodeficiency syndrome; (3) cared for by a physician not
participating in the protocol or the resident, family, or physician declined
a specific evaluation; or (4) identified too late for a timely evaluation
(>48 hours after treatment was initiated). Some episodes were excluded for
more than 1 reason. We compared age and vital signs between the 2592 evaluations
and the 724 episodes that would have qualified for an evaluation but were
excluded because of lack of permission for evaluation or because of late notification
(categories 3 and 4 above). Age, pulse, and respiratory rate were not significantly
different, but average temperature was slightly higher in those not evaluated
(37.4°C vs 37.2°C; P = .002).
Clinical evaluations of nursing home residents were recorded on standardized
forms and placed in the medical record. When an LRI seemed likely, project
nurses collected additional data using the nursing home Minimum Data Set (MDS),18,19 which is a reliable instrument when
used by trained nurse assessors.20 From hospital
(for residents who were hospitalized) and nursing home records, we obtained
the following: active diagnoses and studies pertaining to diagnosis (for example,
urinalysis and cultures of blood, urine, or sputum); oxygen therapy; immunization
information; medications, including antibiotics, psychotropic drugs, and respiratory
drugs; prior diagnoses; and prior hospital use. In 9.2% of evaluations, the
resident was transferred to the hospital before project nurses could complete
a physical assessment. In these instances, we obtained vital sign and clinical
examination data from hospital records. Vital sign data used in the analysis
were those obtained by the project nurse or, if not available, those first
obtained at the hospital (usually the emergency department record). We chose
the first available laboratory data after the resident qualified for evaluation.
From the MDS, we obtained data on depression and delirium; height and
weight; other diagnoses and conditions (including pressure ulcers); use of
devices, such as restraints; and the Cognitive Performance Scale (CPS),21 which measures cognitive impairment. We measured
ADL dependency, by summing self-performance scores for 4 ADL items (grooming,
using the toilet, locomotion, and eating) from the MDS (MDS ADL [Short Form],
scale range of 0-16).22 In the final multivariable
analyses, we simplified this to a 0 to 4 scale by counting the number of these
4 ADL items in which the individual was rated as either dependent or required
extensive assistance. Consistent with MDS instructions, we evaluated ADL and
cognitive status for the week prior to evaluation; delirium symptoms (MDS
section B5)19 include an indication of new
onset or worsening.
We ascertained survival or mortality from all causes at 30 days for
all residents. Project nurses returned to nursing homes at 30 days to reassess
functional status in living residents. In the few instances in which residents
had moved, we followed up on their status at their new location. In the 3
instances in which this was not possible, we performed a death certificate
Chest radiographs were obtained in 2337 of the 2592 evaluations. We
chose to evaluate radiology reports rather than reviewing all radiographs
because only the report is typically available to clinicians. Based on defined
criteria, 2 clinicians independently classified radiology reports into 3 categories:
negative, possible, or probable (this group includes definite pneumonia).
For example, according to these criteria, a report describing "new left lower
lobe infiltrate suggestive of pneumonia" is probable pneumonia, while a report
indicating "possible infiltrate" or "infiltrate suggestive of pneumonia or
congestive heart failure" is possible pneumonia. In St Louis, 2 clinicians
evaluated the reports, and in central Missouri 2 of 4 clinicians considered
each report. When disagreement occurred, all 6 raters at the 2 sites independently
reviewed the reports and attempted to reach consensus. In 11.7% of cases,
consensus either could not be achieved, or was for possible pneumonia when
only probable pneumonia would have qualified the episode for inclusion as
an LRI under the study definition. In those instances, an additional radiologist
independently interpreted the actual radiographs.
Following abstraction of all clinical information and final radiographic
classification, project geriatricians (D.R.M., E.F.B., and S.C.Z.) reviewed
clinical information from all evaluations to make a final determination of
whether an episode met our case definition. In addition to 1117 episodes that
did not meet the LRI definition (Figure 1), we found an additional 43 that technically met our definition
but were not included as LRI cases because another illness or combination
of illnesses was more likely (including 36 in which there was a documented
urinary tract infection).
An additional 26 episodes were dropped from our analytic sample; in
23 there were inadequate data on predictor variables and 3 residents had 2
episodes of LRI in a 30-day interval during which they died. Since death should
only be attributed to 1 episode, we excluded the second episode in these 3
Data imputation was used for missing data since in developing multivariable
models, data imputation is recommended as less biased than dropping cases.23 In this study, imputing mean values for missing continuous
data and the largest category value for missing dichotomous variables was
as efficient as more complicated procedures for imputation. Episodes were
then randomly assigned to a 70% development (n = 975) and 30% validation (n
= 431) sample. Selecting candidate variables and model building were restricted
to the development data until a final variable reduction step.
The initial step in variable selection was based on the literature and
clinical relevance, as judged by the 3 geriatrician investigators. A list
of 25 categories of variables that might be related to mortality was constructed,
including demographic factors (age, sex, race), vital signs (pulse, respiratory
rate, temperature, blood pressure), findings of delirium (eg, acute confusion,
decreased alertness), cognitive status, nutritional status (weight, body mass
index [BMI], total lymphocyte count), physical function (ADL status and other
mobility indicators), indicators of depression, comorbid conditions (eg, congestive
heart failure, chronic obstructive pulmonary disease, stroke), and other laboratory
findings (eg, white blood cell count, serum urea nitrogen, serum sodium).
We then considered descriptive and bivariable statistics describing
the relationship of specific symptoms and examination findings to 30-day mortality.
Using S-Plus software,24 continuous variables
were examined with smoothed plots showing the shape of the relationship between
the variable and mortality. Based on clinical relevance and statistical considerations,
we then took the best representatives from these 25 categories of variables
for consideration in building our multivariable model.25
We excluded 2 indicators of nutritional status, albumin and cholesterol, because
of excessive missing data (35% and 48%, respectively). Changes in Medicare
regulations during the study precluded physicians from ordering a comprehensive
chemistry panel in nursing home residents with a possible LRI. Consistent
with contemporary standards of care, most subjects did not receive an arterial
blood gas or pulse oximetry.
We used forward and backward stepwise logistic regression to consider
combinations of variables for inclusion in our final model (using P = .10 as an initial criterion for statistical significance). We used
generalized estimating equations to adjust logistic regression estimates for
2 kinds of correlation within our data: individuals nested within facilities
and participants represented by more than 1 episode.26
As few individuals had more than 4 episodes of LRI, we restricted the generalized
estimating equations analysis to 4 or fewer episodes to avoid unstable estimates.
In testing continuous variables in these models, we considered the shape
of the variable's relationship to mortality. For example, temperature exhibits
a minimum mortality with a slight elevation of temperature and higher mortality
with both high and low temperatures. Therefore, we tested linear and quadratic
terms as well as using dummy variables to represent low, midrange, and high
temperatures. We also limited the range of continuous variables to avoid undue
influence of outliers. For example, serum urea nitrogen was set to 10 if less
than 10 and to 80 if more than 80. In making final decisions on model inclusion,
we considered clinical meaningfulness and the gain in discrimination by including
a variable as measured by the c statistic and the
Aikake Information Criterion (both available through SAS statistical software).27 We also reconsidered key variables based on the literature,
such as age and respiratory rate, which had not been retained in stepwise
The result of these analyses was an 11-variable model. Because this
was an excessive number of variables for the size of our validation data set,
prior to the final model validation, we drew 5 other random samples from the
entire data set. Three of the 11 variables originally fit to the development
sample (low temperature, congestive heart failure on chest radiograph, and
bilateral infiltrate on chest radiograph) improved discrimination in only
half of the 6 samples, so they were dropped from the model.
We then used coefficients for the 8-variable model, as estimated in
the development sample, to test the model's discrimination and calibration
in the original validation sample.28 To assess
discrimination, we primarily used the c statistic,
which evaluates among all possible pairs of individuals whether those with
higher predictive risk are more likely to die. The c
statistic is also equal to the area under the receiver operating characteristic
curve. The Hosmer-Lemeshow goodness-of-fit statistic was used to measure calibration
by assessing agreement between predicted and observed risk by decile of predicted
Finally, the 8-variable model was fit to the entire data set and used
to create an approximation in the form of a simple scoring system for clinicians.
The predicted probability of mortality associated with each point total was
computed by averaging predicted probability from this logistic model for all
episodes with a given point total. Statistical analyses were performed with
S-Plus24 and SAS statistical software.27
Project nurses evaluated residents in 2592 episodes with symptoms or
signs suggesting an LRI. From these evaluations, we identified 1406 episodes
in 1044 individuals for inclusion in our outcome analysis. Figure 1 shows how we derived our sample. Most residents (n = 794)
had a single episode, 176 had 2 episodes, 48 had 3 episodes, 18 had 4 episodes,
and 8 had more than 4 episodes. In all but 37 of the 1406 episodes, chest
radiographs were available. Based on the assessments of radiographic reports,
186 (13.2%) had possible pneumonia and 748 (53.2%) had probable pneumonia.
There were 207 deaths (14.7%) from all causes within 30 days, with 143 in
the nursing home, 62 in the hospital, and 2 in an extended care unit following
hospitalization. Nineteen percent were hospitalized within 48 hours and 27%
were hospitalized within 30 days.
Table 1 shows selected characteristics
of the development and validation samples at the onset of the LRI episode.
Of note, 75% of episodes occurred in subjects who were older than 80 years. Table 2 and Table 3 show the bivariable relationship of selected variables to
30-day mortality in our entire sample. A large number of variables are associated
with 30-day mortality, including most factors seen in previous studies.
Based on clinical and statistical considerations, we selected an 8-variable
model of 30-day LRI mortality, including serum urea nitrogen, white blood
cell count, BMI, pulse rate, ADL score, low total lymphocyte count (<800/µL
[0.8 × 109/L]), male sex, and decline in mood over 90 days. Table 4 shows estimates derived using generalized
estimating equations for the entire data set. As shown in Table 5, the model fit to the developmental sample showed good discrimination
(c = 0.82) and calibration (Hosmer-Lemeshow goodness-of-fit
statistic P = .85 with nonsignificant values indicating
acceptable calibration). When the coefficients from the developmental sample
were applied to the validation sample, discrimination declined (c = 0 .76) but calibration remained acceptable (P = .54). The validation sample estimate is more likely to be representative
of the model's discriminating ability in an independent sample. Another useful
measure of discrimination is the ratio of mortality in the highest-risk and
lowest-risk quintiles as predicted by the model. In the development set this
ratio was 17.2, and in the validation set it was 13.8. In contrast to these
findings, testing of the 11-variable model showed that although it performed
well in the development data (c = 0.83 and Hosmer-Lemeshow
statistic P = 0.35), it did not perform as well in
the validation set (c = 0.74 and P = .001, which indicates poor calibration).
We used the logistic model based on the entire data set to develop a
simplified risk score, which approximates our logistic model, and can be more
easily applied by clinicians (Table 6
and BOX 2). Table 7 shows how
individual scores correspond to average predicted probabilities (from the
logistic model) and observed mortality. The left portion of the table shows
the risk score applied to the entire data set. Table 7 also shows how a similar score based on the logistic model
from the development set performs in the development and validation sets.
Consider a hypothetical male nursing home resident with an LRI who exhibits the following: serum urea nitrogen of 20 mg/dL (7.14 mmol/L); white blood cell count of 8000/µL (8.0 × 109/L) with 15% lymphocytes; pulse of 80/min; requires extensive assistance in hygiene and locomotion, limited assistance in using the toilet, and supervision in eating; weight, 66 kg (145 lb); height, 170.3 cm (5' 8"); and has not had a recent decline in mood.
To calculate absolute lymphocyte count, multiply white blood cell count
by percentage of lymphocytes: (8000/µL) × .15 = 1200/µL.
To convert height to meters, recall that there are 2.54 cm per inch. Therefore,
body mass index equals 66 divided by (68 × .0254)2 = 22.1.
The Missouri LRI Project risk score is calculated as follows:
(1 point for serum urea nitrogen) + (0 points for white blood cell count)
+ (0 points for absolute lymphocyte count <800/µL) + (1 point for
pulse) + (1 point for sex) + (1 point for activities of daily living) + (2
points for body mass index) + (0 points for mood change) = 6 total points.
Table 7 shows that individuals
with a score between 5 and 6 have a predicted 30-day mortality risk of 6.9%.
Alternatively, using the logistic model in Table 4, a 6.7% mortality risk would
be obtained as follows:
sum = (−4.53 + [0.046 × 20] + [0.052 × 8] + [0.613 × 0]
+ [0.017 × 80] + [0.555 × 1] + [0.31 × 2] − [0.089 × 22.1] + [0.97 × 0]) = −2.63
predicted mortality = esum/(1 + esum) = 0.067 or 6.7%
We developed a new risk-prediction model for LRI in nursing home residents.
In a large sample, our simplified scoring system identified 52% of residents
with a low (score of 0-4) or relatively low (score of 5-6) 30-day mortality
risk. Although many of these residents are likely candidates for nursing home
management, 30% of those hospitalized within 48 hours in our study had scores
of 0 to 6. Hospitalization rates for nursing home residents with infection
and other acute illnesses vary substantially among nursing homes,30- 33 and
some of such hospitalizations may be unnecessary.34
If confirmed in other settings, our model could be helpful in assessing the
need for hospitalization. For higher-risk residents, decisions about treatment
location will depend on individualized treatment goals and weighing the hazards
of hospitalization against the nursing home's capability to provide adequate
For patients with community-acquired pneumonia, the current standard
for estimating risk is the PSI.14 It uses age,
sex, nursing home residence, altered mental status, vital signs, serum urea
nitrogen, glucose, pH, serum sodium, oxygen saturation, presence of pleural
effusion, and selected comorbid diseases to classify individuals into 5 risk
groups. However, its structure (adding points for each year of age and for
nursing home residence) places most nursing home residents in high-risk categories.
In a retrospective study, the PSI predicted mortality reasonably well in 158
episodes of nursing home–acquired pneumonia9;
however, 85% were classified in the highest-risk categories (classes IV and
V). Although we studied the broader category of LRI and not just pneumonia,
the PSI classifies 87% of our subjects in risk classes IV and V. While the
PSI remains an important tool in the more general context for which it was
developed, our model better distinguishes lower-risk episodes of LRI in the
nursing home setting.
Our predictors bear some similarities but also notable differences to
those in the PSI. Common variables to both predictive models include pulse,
serum urea nitrogen, and male sex. Age, a key determinant of the PSI, dropped
out early in our modeling process and is not statistically significant if
added to our final model. This likely reflects the old age of our sample and
the nursing home population in general. Among such individuals, functional
measures, such as ADL status, provide more useful prognostic information than
As with our model, ADL dependency has been repeatedly associated with
LRI or pneumonia mortality in nursing home samples.5- 7,10,11,35
Poor nutritional status has been linked to a variety of poor outcomes.36 Low BMI and low total lymphocyte count,37
which are 2 markers of poor nutritional status, were strongly associated with
mortality from LRI in our model. It is possible that other nutritional variables
might be superior, but both of these are readily available.
Our final 2 risk factors were elevated white blood cell count and decline
in mood over the previous 90 days. While elevated white blood cell count has
not been a significant predictor in previous multivariable models of mortality
from pneumonia in nursing home settings, it was in 2 hospital-based studies
of pneumonia outcomes38,39; one
of these included just elderly patients.39
Major depression40 and comorbid depression41- 43 have been associated
with mortality in nursing home residents, but not specifically from LRI. In
our study, mood decline was a better predictor than summary depression scores,
so it is not clear if this reflects depression or is a marker for general
Several variables do not appear in our model. Rapid respiratory rate
predicts mortality not only in the PSI, but also in 3 previous nursing home
studies using multivariable analyses, including our pilot study.5,7,35
In the current study, pulse rate was highly correlated with respiratory rate
and was a better predictor than respiratory rate. Nonetheless, absence of
respiratory rate as a variable is a potential weakness of our risk prediction
score, and it will be an important variable to assess in future studies evaluating
our prediction rule. Oxygen saturation is also an important variable in the
PSI, but such measurements were relatively uncommon in nursing homes during
our study. For the 27% of subjects who had such data, adding oxygen saturation
did not improve our prediction rule. Oxygen saturation data may play an important
role in assessing treatment decisions in the future. Although fever, low temperature,
and bilateral infiltrates have clinical appeal, and on a bivariable level
are strongly related to mortality, they did not improve our model. In fact,
low temperature and abnormal radiographic findings were included in our penultimate
model (the 11-variable model) but were removed because they weakened the model
in other random samples. Finally, several comorbid conditions, such as congestive
heart failure, are important indicators in the PSI. None were significant
in our multivariable modeling. Their lack of importance in our models may
reflect their high prevalence and the high degree of disability among nursing
home residents (Table 1).
Our findings are subject to several limitations. A key issue is generalizability
to other settings. All study facilities were in central or eastern Missouri.
While they were similar in size and ownership to facilities nationally,15 factors affecting mortality could differ in other
states or countries. More importantly, predictive models often perform less
well in independently derived samples. Internal validation samples help avoid
overfitting models to the peculiarities of a particular data set, but that
is not sufficient to determine the ultimate utility of a prediction rule.
Our model and its associated scoring system should be validated in other studies
of nursing home residents to confirm their usefulness.
Second, important data may have been missing or misclassified. Although
we identified subjects prospectively, some examination information had to
be obtained from hospital records in 9.2% of evaluations. Hospital record
data may not have been as detailed as project nurse assessments. Further,
though all project nurses had strong assessment skills and additional training
for this project, they might have missed some important findings. However,
among variables ultimately included in our model, biases are unlikely. Most
represented objective findings with high reliability, including pulse, sex,
and the 3 laboratory results. Weight and height to compute BMI may be unreliable
in nursing home records, but it is unlikely that they would be systematically
biased across the study. Patient ADL status and information on mood decline
were obtained from interviews with nursing home staff familiar with the resident.
Finally, we combined pneumonia and other LRIs in our analysis because
clinically distinguishing between the 2 is often difficult, particularly in
the nursing home setting, where physicians, advanced-practice nurses, or physician
assistants are frequently unavailable to assess acutely ill residents. Portable
radiographs obtained in nursing homes are of variable quality and require
cautious interpretation. Although we made special efforts to ensure consistency
in classifying radiology reports as possible, probable, or negative for pneumonia,
we reviewed reports rather than radiographs in most cases. We may have misclassified
some subjects as to whether their radiograph suggested pneumonia. We chose
to review reports since reports and not radiographs are usually available
to physicians caring for nursing home residents. Furthermore, because of our
broader definition of LRI, a chest radiograph positive for pneumonia was not
essential for study inclusion. However, pneumonia on chest radiograph was
not a significant predictor of mortality in our multivariable model. These
choices were intended to make our findings optimally useful to physicians
making treatment decisions for ill nursing home residents with LRIs.
We identified a new predictive model for 30-day mortality risk in nursing
home residents with LRIs. Our results are notable for identifying relatively
low-risk residents. Our prediction rule could aid clinicians and researchers
in optimizing care for nursing home residents with LRIs. As with all prediction
rules, it should be validated in other settings.