ARTICLE

Brief Hospitalization and Pulse Oximetry for

Predicting Amoxicillin Treatment Failure in Children
with Severe Pneumonia
Linda Y. Fu, MD, MSca, Robin Ruthazer, MPHb, Ira Wilson, MD, MScb, Archana Patel, MD, DNB, MSCEc, LeAnne M. Fox, MD, MPHd,
Tran Anh Tuan, MD, MSce, Prakash Jeena, MBChB, FCPf, Noel Chisaka, MBChB, MScg, Mumtaz Hassan, MBBS, FRCPh, Juan Lozano, MD, MSci,
I. Maulen-Radovan, MDj, Donald M. Thea, MD, MScd, Shamim Qazi, MB, BS, DSH, MSc, MDk, Patricia Hibberd, MD, PhDb
aDepartment of General and Community Pediatrics, Childrens National Medical Center, Washington, DC; bInstitute for Clinical Research and Health Policy Studies, TuftsNew England Medical Center, Boston, Massachusetts; cClinical Epidemiology Unit, Indira Gandhi Medical Center, Nagpur, India; dCenter for International Health, Boston
University, Boston, Massachusetts; eRespiratory Department, Childrens Hospital 1, Ho Chi Min City, Vietnam; fPaediatrics Department, Nelson R. Mandela School of
Medicine, University of Kwazulu-Natal, Durban; gDepartment of Clinical Medicine, Tropical Disease Research Centre, Ndola, Zambia; hDepartment of Pediatrics, Childrens
Hospital, Islamabad, Pakistan; iDepartment of Pediatrics, Javeriana University School of Medicine, Bogota, Columbia; jPediatrics Department, Hospital Angeles Lomas,
Mexico City, Mexico; kWorld Health Organization, Geneva, Switzerland

The authors have indicated they have no nancial relationships relevant to this article to disclose.

ABSTRACT
OBJECTIVE. In settings with limited assessment tools, we sought to determine whether
early clinical signs and symptoms and blood oxygen saturation would predict
amoxicillin treatment failure in children with severe pneumonia (as defined by the
World Health Organization).
METHODS. Data were from a previously reported, multinational trial of orally admin-

istered amoxicillin versus injectable penicillin for the treatment of World Health
Organization defined severe pneumonia in children 3 to 59 months of age. We
assessed all 857 participants assigned randomly to the experimental amoxicillin
arm. Six multivariate logistic regression models were created and evaluated for
their ability to predict failure after 48 hours of therapy. Regression models included vital signs, symptoms, and laboratory data collected at baseline and after 12
or 24 hours of observation. Oxygen saturation data were included in 3 models.
RESULTS. Clinical treatment failure occurred for 18% of children. Younger age,

increased initial respiratory rate, and baseline hypoxia predicted treatment failure
in all models. Data available after 24 hours improved the ability to predict failure
compared with data available at baseline or 12 hours. The inclusion of oximetry
data improved the predictive ability at baseline, 12 hours, and 24 hours. The ability
to predict failure after 12 hours of observation with oximetry data was similar to
the predictive ability after 24 hours without pulse oximetry data.
CONCLUSIONS. Assessment of clinical parameters at presentation and after 24 hours
improved the ability to predict clinical failure of oral amoxicillin therapy, compared with assessment at presentation alone or at presentation and after only 12
hours, for children with World Health Organization defined severe pneumonia.

e1822

FU et al

www.pediatrics.org/cgi/doi/10.1542/
peds.2005-2673
doi:10.1542/peds.2005-2673
We attest to the fact that we, the authors,
are responsible for this research review. We
have all participated in the concept and
design of the manuscript, interpretation of
the articles reviewed, and drafting and
revising of the manuscript. We have all
approved the documents submitted.
Key Words
amoxicillin, pneumonia, hospitalization,
oximetry, child
Abbreviations
ARIacute respiratory infection
LCIlower chest wall indrawing
RSVrespiratory syncytial virus
WHOWorld Health Organization
CI condence interval
Accepted for publication Jul 11, 2006
Address correspondence to Linda Y. Fu, MD,
MSc, Department of General and Community
Pediatrics, George Washington University,
Childrens National Medical Center, 111
Michigan Ave, NW, Washington, DC 20010.
E-mail: lfu@cnmc.org
PEDIATRICS (ISSN Numbers: Print, 0031-4005;
Online, 1098-4275). Copyright 2006 by the
American Academy of Pediatrics

CUTE RESPIRATORY INFECTION (ARI) is one of the

leading causes of death throughout the world for
children 5 years of age and is responsible for an estimated 1.9 million childhood deaths each year.1 In contrast to the situation in industrialized nations, childhood
pneumonia in developing countries is caused commonly
by bacteria and not viruses.2,3 Bacterial infection is associated with higher ARI mortality rates than viral infection.2,4 Accordingly, clinical trials have suggested that
antimicrobial therapy may reduce ARI-associated mortality rates by one half in developing nations.2 Because of
the difficulties associated with diagnosing bacterial
pneumonia in the developing world,2,3,5 the World
Health Organization (WHO) has based case management
solely on clinical signs for the initiation of empiric antibiotic therapy.4,6
For children who have cough or difficulty breathing,
the WHO ARI case management guidelines require only
an assessment of the respiratory rate and the presence of
visible and audible signs of respiratory distress. The presence of rapid breathing (60 breaths per minute for
children 2 months of age and 50 breaths per minute
for children 2 months of age) confers a diagnosis of
nonsevere pneumonia, which is treated with orally administered cotrimazole or amoxicillin at home. Lower
chest wall indrawing (LCI) as an additional sign signifies
severe pneumonia, which is treated with intramuscularly administered benzylpenicillin or ampicillin and
hospitalization for at least 48 hours. A child whose disease improves from severe pneumonia to nonsevere
pneumonia or well status after 48 hours of therapy is
sent home with orally administered antibiotics, whereas
a child whose condition remains unchanged or worsens
remains hospitalized and receives expanded antibiotic
coverage.3 The WHO management definitions of pneumonia do not rely on the typical diagnostic methods
used in developed countries, such as bacterial culture or
radiography, and inevitably include more cases of nonpneumonia respiratory disorders, such as reactive airway disease and viral bronchiolitis. Nevertheless, because of the high incidence and severity of disease
associated with bacterial pneumonia in developing nations, use of empiric antibiotic therapy based on these
guidelines has been estimated to reduce mortality rates
in the developing world by 27%.7
In a recent, hospital-based trial, orally administered
amoxicillin was found to be equivalent to injectable
penicillin for the treatment of severe pneumonia in developing countries.8 Although this finding supports the
home administration of orally administered amoxicillin,
the safety of home-based therapy has yet to be established. An important concern is that therapies such as
supplemental oxygen and nebulized bronchodilator
therapy cannot be administered routinely in the home.
In addition, continuous observation of a patients response to therapy by trained health care workers is not

possible. To guide practitioners in their use of orally

administered amoxicillin, our study used data from the
previously published trial to determine the incremental
gain in ability to predict amoxicillin treatment failure
achieved with information gathered after 12 or 24 hours
of hospitalization, compared with information from a
single baseline assessment, for children with severe
pneumonia.8 We also evaluated whether changes in
blood oxygen saturation further improved the ability to
predict treatment failure.811
METHODS
Sample and Data
This posthoc cohort analysis was nested within the previously completed, randomized trial comparing orally
administered amoxicillin with injectable penicillin for
children 3 to 59 months of age with severe pneumonia
and included only children assigned randomly to the
amoxicillin arm.8 The trial was conducted in the pediatric departments of tertiary care facilities in 9 locations in
8 countries, namely, Colombia, Ghana, India, Mexico,
Pakistan, South Africa (2 sites), Vietnam, and Zambia.
Children who presented to the emergency department
with a history of cough or difficulty breathing and LCI
were considered for study entry. Children with LCI associated with reactive airway disease (determined on the
basis of the response to 2 courses of nebulized bronchodilator therapy) were excluded. In addition, children
with danger signs (inability to drink, abnormal sleepiness, central cyanosis, or convulsions) indicating very
severe disease, a recent history of a very severe infectious/noninfectious disease, a chronic or congenital illness, known asthma, clinically evident HIV infection,
persistent vomiting, a known penicillin allergy, or 48
hours of antibiotic therapy before presentation were excluded.
Baseline clinical assessment (history and physical examination) and laboratory (oxygen saturation and nasopharyngeal isolate respiratory syncytial virus [RSV]
testing) data were obtained within 1 hour after enrollment, before the first dose of antibiotics. The children
included in this analysis received their first dose of orally
administered amoxicillin syrup (45 mg/kg per day, divided into 3 doses) within the first hour of enrollment.
Oxygen saturation was determined through pulse oximetry in room air, when the child was calm and not crying
(Nellcor N-200E pulse oximeter, with N-25 sensor; Nellcor, Hayward, CA). Study personnel were trained and
certified in assessing respiratory rate, presence of LCI,
and danger signs by using WHO standard ARI case management modules and a WHO training videotape.
All subjects were hospitalized for 48 hours. Vital,
clinical, and danger signs and oxygen saturation data
were recorded every 6 hours. Indications for providing
oxygen (oxygen saturation of 90% or respiratory rate
PEDIATRICS Volume 118, Number 6, December 2006

e1823

of 70 breaths per minute), antipyretics (fever of

39C), and bronchodilator treatments (audible
wheeze) were standardized across study sites. Supportive therapies were given as needed after vital signs were
assessed and recorded for each time interval. Designation of treatment failure occurred only after the site
principal investigator reviewed all study data to confirm
failure. All variables collected were entered into a SASformat database.8 Written informed consent was obtained from parents or legal guardians of children before
enrollment. Ethical approval was obtained from every
local institution and both sponsoring organizations. The
study was monitored by an independent data safety
monitoring board.

12 months of age: very fast: 70 breaths per minute;

fast: 50 breaths per minute; not fast: 50 breaths per
minute; for children 12 months of age: very fast: 60
breaths per minute; fast: 40 breaths per minute; not
fast: 40 breaths per minute).3 Hypoxemia at baseline
was assessed on room air and was considered present if
oxygen saturation was 90% at sea level or 88% at
elevation (in Columbia and Mexico). Temperature, respiratory rate, and oxygen saturation were analyzed at
12 hours and at 24 hours as changes from baseline
values. Variables assessed at 12 hours and 24 hours were
coded as missing for children who were designated as
experiencing treatment failure before the assessment
time.

Coding of Treatment Outcomes

Treatment success was considered to be improvement to
nonsevere pneumonia or well status (normal respiratory
rate) after 48 hours of therapy in both the original trial
and this analysis. Treatment failure in this analysis
included children with persistent LCI after 48 hours of
therapy and those with danger signs or severe hypoxia
(oxygen saturation in room air of 80% at sea level
and 75% at elevation). Case report forms for children
with nonpneumonia-related outcomes that were classified as failure in the original trial (children who received another antibiotic, had a newly diagnosed comorbid condition, withdrew from the study voluntarily,
or left against medical advice at any time after study
enrollment)8 were reviewed and reclassified as treatment success or treatment failure or excluded if they
did not meet 1 of the 2 definitions. Reclassification of
cases was performed independently by 2 investigators,
and discrepancies were resolved through discussion and
consensus. Inter-rater agreement was assessed with a
statistic.

Statistical Analyses
The relationship of each predictor variable to oral amoxicillin treatment failure was examined in univariate analyses by using the 2 test (discrete data) or unpaired t test
(continuous data). A crude odds ratio with 95% confidence interval (CI) was calculated for the effect of each
predictor on the outcome by using logistic regression
modeling. All analyses were conducted with SAS 9.0
(SAS Institute, Cary, NC). Test levels for significance
were P .05.
To address our study question, 3 multivariate logistic
regression models were created. The first model used
data obtained at baseline only, the second used data
from baseline and from 12 hours after therapy initiation,
and the third used data from baseline and from 24 hours
after therapy initiation. A forward stepwise selection
process was used each time to choose variables to predict
failure. Pulse oximetry data were excluded from these
models. Three additional models were then created by
allowing pulse oximetry data to enter the selection process. All variables used in the development of regression
models were checked for colinearity; a Pearson correlation coefficient of 0.8 was considered highly colinear.
Because the range of Pearson correlation coefficients
was 0 to 0.73 for all pairings, no variable was excluded.
Because of a relatively high percentage of missing values
for the variables of immunization status (5%) and presence of RSV (11%), dummy variables were created in all
6 models for missing values for these 2 variables. Significance levels for entry and elimination in the stepwise
procedure were set at .1.
The final regression models included study site as a
random effect. Random-effects modeling was performed
by using the SAS macro program Glimmix (available at:
http://ftp.sas.com). The predictive abilities of the 6 models generated were compared by using receiver operating
characteristic curve areas and likelihood ratios. The Hosmer-Lemeshow goodness-of-fit test was used to determine the fit of the models to the data.

Coding of Predictor Variables

A priori and on the basis of findings from previous
studies,4,9,1217 we selected the following as potential predictor variables: age (6 months, 6 11 months, or 12
months); use of antibiotics within 48 hours before enrollment; immunization status (up to date or not);
breastfeeding status (breastfeeding at time of enrollment, not breastfeeding, or not applicable because child
was 24 months of age); dehydration at baseline; nutritional status (based on weight-for-age z score); presence of RSV; respiratory rate at baseline, at 12 hours,
and at 24 hours; oxygen saturation at baseline, at 12
hours, and at 24 hours; temperature at baseline, at 12
hours, and at 24 hours; and presence of nonauscultatory
audible wheeze at 12 hours and at 24 hours. Wheezing
at baseline was not included in the analysis because this
was a study exclusion criterion. Respiratory rate at baseline was recorded as a 3-level class variable (for children
e1824

FU et al

RESULTS
Study Group
A total of 857 children (50.4% of 1702 patients enrolled
in the study) who were assigned randomly to the amoxicillin arm of the clinical trial represented the present
study sample. We identified 23 children in the trial with
treatment failure for reasons other than failure to exhibit improvement or worsening pneumonia (8 children
received another antibiotic, 4 had newly diagnosed comorbid conditions, 2 had newly diagnosed conditions
and received another antibiotic, and 9 withdrew from
the study voluntarily or left against medical advice).
Two investigators (Drs Fu and Fox) independently reclassified these cases according to our definitions of
treatment success and failure and agreed on 22 of 23
cases. Inter-rater agreement was high ( 0.93; 95%
CI: 0.80 1.00). Of the 23 cases in question, 3 were
reclassified as treatment success, 8 were reclassified as
treatment failure, and 12 were excluded from analysis
because they were not classifiable with the information
available (Fig 1). After reclassification, the rate of treatment failure was 18% (152 of 843 cases), which was
very similar to the percentage obtained in the amoxicillin arm of the clinical trial (19%). Among all of the
patients classified as experiencing treatment failure, the
condition of 13% had worsened, compared with baseline, whereas that of the rest had not changed.
According to univariate analysis, age was a strong
predictor of treatment failure (P .001) (Table 1). The

odds of failure for the youngest children (6 months of

age) was 3.6 times the odds of failure for the oldest
children (12 months of age). Immunizations not being
up to date and antibiotic use for up to 48 hours before
trial entry were both associated with twofold greater
odds of treatment failure (P .004 and P .02, respectively). In addition, baseline fast or very fast respiratory
rate and baseline hypoxia were associated with greater
odds of treatment failure (P .03 and P .004, respectively). The apparent association of breastfeeding with
treatment failure disappeared after adjustment for age
(P .54).
Assessment of the Benets of 12 or 24 Hours of Patient
Observation for Predicting Amoxicillin Treatment Failure
In univariate analysis, at 12 hours after enrollment in
the amoxicillin arm of the trial, a 1C increase in temperature over baseline was associated with 27% increased odds of treatment failure (P .01) (Table 1).
After 24 hours of treatment, a 1C increase in temperature was associated with 52% increased odds of failure,
and an increase in respiratory rate of 5 breaths per
minute was associated with 18% increased odds of failure (P .001 for both variables) (Table 1).
We created 3 multilevel, multivariate, logistic regression models to compare the ability to predict treatment
failure with (1) baseline data only, (2) data from baseline and from 12 hours after therapy initiation, or (3)
data from baseline and from 24 hours after therapy

FIGURE 1
Reclassication of treatment outcomes. a No children in
the amoxicillin arm died during the clinical trial.

PEDIATRICS Volume 118, Number 6, December 2006

e1825

TABLE 1 Odds of Amoxicillin Treatment Failure According to Patient History and Signs and Symptoms
at Baseline and 12 and 24 Hours After Study Enrollment
Variable

Patient history and baseline variables

Age, %
35 mo
611 mo
1259 mo
Antibiotics taken in previous 48 h, %
Immunizations not up to date, %
Breastfeeding, %
Not applicable because age 24 mo
No and age 24 mo
Yes and age 24 mo
Evidence of clinical dehydration, %
Weight-for-age z score, mean SD
Positive RSV status, %
Respiratory rate, %b
Very fast
Fast
Not fast
Temperature, mean SD, C
Hypoxia, %d
Variables assessed after 12 h
Change in respiratory rate, mean SD,
breaths per min
Change in temperature, mean SD, C
Audible wheezing, %
Change in oxygen saturation, mean SD, %
Variables assessed after 24 h
Change in respiratory rate, mean SD,
breaths per min
Change in temperature, mean SD, C
Audible wheezing, %
Change in oxygen saturation, mean SD, %
a Odds

Proportion or Mean

Crude Odds Ratio of

Treatment Failure
(95% CI)

Treatment Failure
(n 152)

Treatment Success
(n 691)

57.2
23.0
19.7
13.8
16.6

31.2
28.6
40.2
7.8
8.7

3.60 (2.305.63)
1.59 (0.952.66)
1.00
1.90 (1.113.26)
2.09 (1.253.49)

4.0
18.4
77.6
1.3
0.4 1.4
28.4

18.8
17.5
63.8
0.7
0.6 1.4
24.4

0.20 (0.090.44)
0.87 (0.551.37)
1.00
1.82 (0.359.44)
1.07 (0.951.21)a
1.23 (0.811.87)

32.9
60.5
6.6
37.5 0.9
15.8

23.4
67.8
8.9
37.6 0.9
8.1

1.93 (0.924.05)
1.20 (0.592.43)
1.00
0.87 (0.711.07)c
2.12 (1.273.54)

6 13

6 11

1.01 (0.941.09)e

0.4 1.0
44.1
0.7 4.9

0.6 0.9
40.7
0.1 3.0

1.27 (1.051.54)c
1.15 (0.811.64)
1.14 (1.081.19)f

6 13

11 12

1.18 (1.091.28)e

0.5 1.1
44.4
0.1 4.0

0.9 1.0
37.5
1.1 3.2

1.52 (1.251.85)c
1.33 (0.931.90)
1.20 (1.111.32)f

ratio and 95% CI reported for an increase in weight-for-age z score of 1 unit.

b If the child was 12 months of age, then very fast respiratory rate was 70 beats per minute, fast was 50 beats per minute, and not fast was

50 beats per minute. If the child was 12 months of age, then very fast was 60 beats per minute, fast was 40 beats per minute, and not fast
was 40 beats per minute.
c Odds ratio and 95% CI reported for an increase in temperature of 1C.
d Hypoxia was dened as oxygen saturation in room air of 90% at sea level or 88% at elevation (in Columbia and Mexico).
e Odds ratio and 95% CI reported for an increase in respiratory rate of 5 breaths per minute.
f Odds ratio and 95% CI reported for a decrease in blood oxygen saturation of 1%.

initiation. These models included only variables that

could be assessed with methods promoted currently by
the WHO for care at the first referral level in developing
countries (ie, excluding pulse oximetry data). In the
model that included only baseline data, younger age
and increased respiratory rate were significant predictors
of treatment failure (Table 2). In the model that included
data from baseline and from 12 hours after therapy
initiation, younger age, high respiratory rate at baseline,
and increasing respiratory rate over time were significant predictors of treatment failure. In the model that
included data from baseline and from 24 hours after
therapy initiation, younger age, increased respiratory
rate at baseline, higher temperature at baseline, increasing respiratory rate over time, and increasing temperae1826

FU et al

ture over time were significant predictors of treatment

failure. When we compared the 3 models likelihood
ratios, we found that treatment failure was predicted
most accurately by using data from baseline and from
24 hours after therapy initiation (P .001 for comparisons with both other models) (Table 3). There was no
significant difference in predictive ability between the
regression model that included baseline data alone and
the model that included data from baseline and from 12
hours after therapy initiation (P .35) (Table 3).
Assessment of the Benets of Pulse Oximetry Data for
Predicting Amoxicillin Treatment Failure
In univariate analysis, hypoxia at baseline was associated
with more than twofold increased odds of treatment

TABLE 2 Signicant Predictors of Amoxicillin Treatment Failure in Multivariate Analyses

Multivariate Logistic Regression Models

Adjusted odds ratio (95% CI)a

Age
35 mo
611 mo
1259 mo
Respiratory rate at baselineb
Very fast
Fast
Not fast
Hypoxia at baselinec
Increasing respiratory rate after 12 hd
Increasing respiratory rate after 24 hd
Decreasing oxygen saturation after 12 he
Decreasing oxygen saturation after 24 he
Higher temperature at baselinef
Increasing temperature after 24 hf
Wheezing after 24 h
Area under the curve (95% CI)
a All

Baseline Data
(n 842)

Baseline and
12-h Data
(n 842)

Baseline and
24-h Data
(n 838)

Baseline Data
With Pulse
Oximetry
(n 838)

Baseline and
12-h Data With
Pulse Oximetry
(n 833)

Baseline and
24-h Data With
Pulse Oximetry
(n 832)

3.65 (2.255.93)
1.74 (1.022.98)
1.00

4.01 (2.456.57)
1.91 (1.113.30)
1.00

4.48 (2.627.67)
2.07 (1.153.70)
1.00

3.60 (2.225.85)
1.74 (1.022.98)
1.00

3.89 (2.236.81)
1.87 (1.023.45)
1.00

3.90 (2.326.57)
1.79 (1.013.19)
1.00

3.90 (1.768.65)
1.68 (0.803.51)
1.00

6.89 (2.8816.46)
2.29 (1.074.91)
1.00

12.20 (4.7831.19)
2.86 (1.276.43)
1.00

3.60 (1.627.98)
1.56 (0.753.24)
1.00
2.37 (1.204.68)

6.34 (2.4516.44)
2.08 (0.914.76)
1.00
3.99 (1.798.90)
1.15 (1.042.85)

11.16 (4.5127.64)
2.57 (1.175.60)
1.00
3.83 (1.828.09)

1.16 (1.061.26)
1.28 (1.031.41)

1.32 (1.201.45)
1.17 (1.101.24)
1.20 (1.091.32)

0.67 (0.630.72)

1.80 (1.262.55)
1.97 (1.412.74)
2.19 (1.223.93)
0.73 (0.680.77)

0.68 (0.630.73)

0.69 (0.650.74)

0.72 (0.670.77)

2.05 (1.173.61)
0.76 (0.710.80)

6 logistic regression models were adjusted for study site.

b If the child was 12 months of age, then very fast respiratory rate was 70 beats per minute, fast was 50 beats per minute, and not fast was 50 beats per minute. If the child was 12 months

of age, then very fast was 60 beats per minute, fast was 40 beats per minute, and not fast was 40 beats per minute.
was dened as oxygen saturation in room air of 90% at sea level or 88% at elevation (in Columbia and Mexico).
d Odds ratio and 95% CI reported for an increase in respiratory rate of 5 breaths per minute.
e Odds ratio and 95% CI reported for a decrease in blood oxygen saturation of 1%.
f Odds ratio and 95% CI reported for an increase in temperature of 1C.
c Hypoxia

TABLE 3 Abilities of Logistic Regression Models to Predict Amoxicillin Treatment Failure (n 828)
Logistic Regression Model

Likelihood
Ratio

Degrees of
Freedom

Predictive
Abilitya

Pb

24 h of observation with pulse oximetry

24 h of observation
12 h of observation with pulse oximetry
Baseline assessment only with pulse oximetry
12 h of observation
Baseline assessment only

105.3
86.6
85.2
62.9
51.8
50.9

8
8
7
5
5
4

1
2 (tie)
2 (tie)
4
5 (tie)
5 (tie)

.001
.23
.001
.001
.35

a Predictive
bP

ability was best for the regression model ranked 1 and worst for the models ranked 5.
value for comparison with the next highest ranking regression model.

failure (P .004) (Table 1). After 12 and 24 hours of

treatment, decreasing blood oxygen saturation was associated with 14% and 20% increased odds of failure,
respectively (P .001 for both variables) (Table 1). We
created 3 new predictive models that allowed variables
related to blood oxygen saturation to enter the selection
process. The 3 models used (1) baseline data only, (2)
data from baseline and from 12 hours after therapy
initiation, or (3) data from baseline and from 24 hours
after therapy initiation. In the logistic regression model
that used baseline data alone, younger age, increased
respiratory rate, and hypoxia were significant predictors
of treatment failure (Table 2). This model predicted
treatment failure better than did the model that used
baseline data and excluded hypoxia as a possible predic-

tor variable (P .001) (Table 3). In the regression model

that used data from baseline and from 12 hours after
therapy initiation, younger age, higher respiratory rate
at baseline, hypoxemia at baseline, increasing respiratory rate over time, and decreasing oxygen saturation
over time were significant predictors of treatment failure
(Table 2). This 12-hour model predicted treatment failure better than did the 12-hour model that excluded
pulse oximetry data (P .001) (Table 3). In the regression model that used data from baseline and from 24
hours after therapy initiation, younger age, increased
respiratory rate at baseline, hypoxia at baseline, development of audible wheezing, increasing respiratory rate
over time, and decreasing oxygen saturation over time
were significant predictors of treatment failure (Table 2).
PEDIATRICS Volume 118, Number 6, December 2006

e1827

When we compared this model with the one that included 24 hours of data without pulse oximetry information, again we found that treatment failure was predicted more accurately with pulse oximetry data than
without such data (P .001) (Table 3).
In all 3 regression models that included pulse oximetry variables, baseline hypoxia increased the odds of
treatment failure twofold to fourfold, and each 1% decrease in oxygen saturation over 12 or 24 hours was
associated with 20% increased odds of treatment failure (Table 2). Variables related to pulse oximetry were
such strong predictors of treatment failure that there was
no difference in predictive ability between the model
that used data collected after 12 hours of therapy and
included pulse oximetry information and the model that
used data collected after 24 hours of therapy and excluded pulse oximetry information (P .24) (Table 3).
All 6 logistic regression models fit the data well, according to the Hosmer-Lemeshow goodness-of-fit test (P
.52.99).
DISCUSSION
Our study has 2 main findings. First, information obtained after at least 12 hours but ideally 24 hours of
patient observation provides more valuable information
than baseline assessments alone for predicting which
children with severe pneumonia are likely to experience
failure of oral amoxicillin therapy. Second, assessments
of blood oxygen saturation at baseline and saturation
trends over time improve the ability to predict which
children will experience failure of oral therapy. Specifically, if pulse oximetry is available, then the ability to
predict treatment failure with data from baseline and
from 12 hours after initiation of therapy is almost identical to the predictive ability obtained with data from
baseline and from 24 hours without the benefit of pulse
oximetry. Having 24 hours to observe a child with access
to pulse oximetry provides the best predictive ability for
treatment failure.
These findings extend the observations of the recently
published trial that demonstrated equivalence between
injectable penicillin and orally administered amoxicillin
for the treatment of severe pneumonia in children.8 The
original trial established the equivalence of oral and
parenteral therapy in a facility setting, without affirming
the safety of home oral therapy. Our study suggests that
an identifiable subset of children presenting with severe
pneumonia may need to be hospitalized for optimally
safe care if treated with oral therapy. Hospitalization
provides patients with monitoring by trained professionals, as well as the opportunity to receive ancillary therapies as appropriate. Patient observation has been cited
as one of the most common methods through which
health care providers acquire information on which to
base decisions.18 One fifth of all children in our study
received supplemental oxygen at least once during the
e1828

FU et al

first 24 hours, and two fifths received bronchodilator

therapy. Appropriate administration of oxygen and
bronchodilators by health care workers, according to
well-defined criteria, during 24 hours of hospitalization
likely contributes to the high treatment success rate
achieved with amoxicillin.
Our findings that demonstrate the benefits of blood
oxygen saturation measurements in the prediction of
amoxicillin treatment outcomes are consistent with the
findings of other researchers. Studies from Indonesia,
Kenya, Zambia, and the Gambia have found the risk of
death with acute lower respiratory tract infection to be
1.4 to 4.6 times higher for children with hypoxia than
for those without.9,1921
In our study, low blood oxygen saturation conferred a
risk of treatment failure even with adjustment for respiratory rate. This suggests that the 2 factors may not be
measures of exactly the same physiologic process. A
study of children in the Gambia found that high respiratory rate was a poor predictor of hypoxemia.22 Oxygen
saturation trends may prove to be as helpful as clinical
criteria such as respiratory rate trends for guiding pneumonia care. In one study in Papua, New Guinea, children with severe pneumonia who were given supplemental oxygen on the basis of saturation criteria had a
marginally lower risk of death (P .07) than those
treated according to clinical signs.10 This might be attributable to the fact that pulse oximetry readings are more
consistent and therefore more reliable than clinical assessments, which vary from clinician to clinician.23 Ideally, oximetry data could be used in conjunction with
clinical guidelines to assess patient status and progress.
In one study, blood oxygen saturation of 96.6% and
WHO criteria for pneumonia (nonsevere or severe) together identified 87% of cases of radiologically evident
pneumonia.11
Despite the growing literature supporting their use,
pulse oximeters are not routinely available at first-referral health facilities in developing countries. The 2 greatest obstacles to their widespread use are perceived high
cost and impracticality. These obstacles are not insurmountable. Pulse oximeters are less expensive to use
than radiographic technology and can operate with car
batteries if necessary. The machines are easy to calibrate,
to use, to maintain, and to repair when broken.11,24 Because oximeters improve the ability to differentiate children with pneumonia who are more likely to fare poorly
from those who are less likely, they may assist practitioners in deciding which children can be sent home with
oral therapy. In this way, they may be a cost-effective
addition for families and overall health care. Of course,
achieving the maximal benefit of pulse oximeters involves using them to guide administration of supplemental oxygen. As new technologies such as oxygen
concentrators,25 which provide oxygen inexpensively
and reliably, become more available in developing na-

tions, there will be even more inducements for widespread distribution of pulse oximeters.
Our study has several limitations. To make the results
of our study readily applicable to practitioners internationally, we used the WHO ARI algorithm as the basis for
our definition of treatment failure. Because of the minimal resources needed to follow the guidelines, the WHO
ARI algorithm is the accepted method for diagnosing
pneumonia in much of the world. Use of empiric antibiotic therapy based on these guidelines has been estimated to reduce mortality rates in the developing world
by 27%.7 However, the WHO definition of pneumonia
has a significant false-positive rate and misclassified 35%
of well children as having pneumonia in one study.11
Patients with ARI attributable to pathologic processes
not expected to be responsive to amoxicillin were likely
included as cases of antibiotic failure in our study. At the
Durban and Zambia study sites, the adult seroprevalence
of HIV was as high as 30%. Children from those sites
who experienced failure of therapy were at greater risk
of having pneumonia caused by Pneumocystis jiroveci, because amoxicillin is ineffective against that organism.
Overall, 67% of the children in our study were afebrile
at the time of study enrollment; some of those children
with failure of antibiotic therapy were likely suffering
from reactive airway disease rather than pneumonia,
despite our baseline exclusion of children with known
asthma or positive responses to bronchodilator therapy.
Our study identified children infected with the 2 most
common pathogens causing community-acquired pneumonia. Twenty-seven percent of the children had Streptococcus pneumoniae isolated from their nasopharynx, and
20% had Haemophilus influenzae. Negative culture results
were not a risk factor for treatment failure. It is likely
that a higher percentage of children in the study had
bacterial lung infections, because the percentage of children with ARI symptoms and positive lung aspirate cultures ranged from 21% to 74% in other studies.2 Because of the diagnostic ambiguity inherent in using the
WHO definition of severe pneumonia for determination of treatment failure, the conclusions of our study
and the original trial may not be directly applicable to
settings in which more-precise methods of diagnosing
bacterial pneumonia, such as bacterial culture and radiography, are routine.
Additional study is necessary to translate our findings
into simple practical guidelines for individual-level patient care in developing countries. We found incrementally increasing value of 12-hour and 24-hour patient
data over baseline data alone, as well as benefits of
knowing blood oxygen saturation at all time points, for
predicting amoxicillin treatment failure. Future treatment guidelines will need to balance the maximization
of predictive ability with practicality. Although the techniques for assessing clinical status in this trial were the
same as the methods promoted by the WHO for care at

first-referral health facilities, this trial was conducted in

tertiary care hospitals, to ensure patient safety and study
quality. The feasibility of providing observational stays
and pulse oximetry for patients at first-referral health
facilities in developing nations has yet to be determined.
Oral therapy may now be considered appropriate
treatment for severe pneumonia in developing nations.8
Because oral therapy does not require technical skill for
administration, the hospital length of stay for severe
pneumonia may be reduced. However, caution must be
exercised to ensure that children at risk of oral therapy
failure are identified properly. Identification of children
at high risk can be performed with 12 hours of patient
observation if the capacity to measure blood oxygen
saturation is available. If it is not, then information gathered after 24 hours of observation without pulse oximetry may be equally beneficial in predicting treatment
failure.
ACKNOWLEDGMENTS
This work was supported in part by the Agency for
Healthcare Research and Quality (grant T32 HS00060);
the Department of Child and Adolescent Health and
Development, World Health Organization (Geneva,
Switzerland); and the Applied Research in Child Health
Project, Boston University (Boston, MA), under the US
Agency for International Development cooperative
agreement HRN-A-00-96-900100-00.
We are grateful to Gregory Koblentz, Robert Goldberg, and Olivier Fontaine for their careful review of the
manuscript and helpful suggestions.
REFERENCES
1. Williams BG, Gouws E, Boschi-Pinto C, et al. Estimates of
world-wide distribution of child deaths from acute respiratory
infections. Lancet Infect Dis. 2002;2:2532
2. Shann F. Etiology of severe pneumonia in children in developing countries. Pediatr Infect Dis. 1986;5:247252
3. World Health Organization. WHO Programme for the Control of
Acute Respiratory Infections: Technical Bases for the WHO Recommendations on the Management of Pneumonia in Children at First-Level
Facilities. Geneva, Switzerland: World Health Organization;
1991. Publication WHO/ARI/91.20
4. Tupasi TE, Velmonte MA, Sanvictores ME, et al. Determinants
of morbidity and mortality due to acute respiratory infections:
implications for intervention. Pediatr Infect Dis. 1988;157:
615 623
5. Korpi M, Leinonen M, Markku K, et al. Bacterial coinfection in
children hospitalized with respiratory syncytial virus infections. Pediatr Infect Dis. 1989;8:687 692
6. World Health Organization. WHO Programme for the Control of
Acute Respiratory Infections: Acute Respiratory Infections in Children:
Case Management in Small Hospitals in Developing Countries. Geneva, Switzerland: World Health Organization; 1990. Publication WHO/ARI/90.5
7. Sazawal S, Black RE, Pneumonia Case Management Trials
Group. Effect of pneumonia case management on mortality in
neonates, infants, and preschool children: a meta-analysis of
community-based trials. Lancet. 2003;3:547556
8. Addo-Yobo E, Chisaka N, Hassan M, et al. Oral amoxicillin

PEDIATRICS Volume 118, Number 6, December 2006

e1829

9.

10.

11.

12.

13.

14.

15.

versus penicillin for severe pneumonia in children aged 3 to 59

months: a randomized multicentre equivalency study. Lancet.
2004;364:11411148
Djelantik IG, Gessner BD, Sutanto A, et al. Case fatality proportions and predictive factors for mortality among children
hospitalized with severe pneumonia in a rural developing
country setting. J Trop Pediatr. 2003;49:327332
Duke T, Mgone J, Frank D. Hypoxaemia in children with
severe pneumonia in Papua New Guinea. Int J Tuberc Lung Dis.
2001;5:511519
Madico G, Gilman R, Jabra A, et al. The role of pulse oximetry:
its use as an indicator of severe respiratory disease in Peruvian
children living at sea level. Arch Pediatr Adolesc Med. 1995;149:
1259 1263
Sehgal V, Sethi GR, Sachdev HP, Satyanarayana L. Predictors of
mortality in subjects hospitalized with acute lower respiratory
tract infections. Indian Pediatr. 1997;34:213219
Straus WL, Qazi SA, Kundi Z, Nomani NK, Schwartz B. Antimicrobial resistance and clinical effectiveness of co-trimoxazole
versus amoxicillin for pneumonia among children in Pakistan:
randomized controlled trial. Lancet. 1998;352:270 274
Shann F, Barker J, Poore P. Clinical signs that predict death in
children with severe pneumonia. Pediatr Infect Dis J. 1989;8:
852 855
CATCHUP Study Group. Clinical efficacy of co-trimoxazole
versus amoxicillin twice daily for treatment of pneumonia: a
randomized controlled clinical trial in Pakistan. Arch Dis Child.
2002;86:113118

e1830

FU et al

16. Delport SD, Brisley T. Aetiology and outcome of severe community-acquired pneumonia in children admitted to a paediatric intensive care unit. S Afr Med J. 2002;92:907911
17. MASCOT Study Group. Clinical efficacy of 3 days versus 5 days
of oral amoxicillin for treatment of childhood pneumonia: a
multicentre double-blind trial. Lancet. 2002;360:835 841
18. Hedberg B, Larsson U. Observations, confirmations and
strategies: useful tools in decision-making process for nurses in
practice? J Clin Nurs. 2003;12:215222
19. Onyango FE, Steinhoff MC, Wafula EM, et al. Hypoxaemia in
young Kenyan children with acute lower respiratory infection.
BMJ. 1993;306:612 615
20. Smyth A, Carty H, Hart CA. Clinical predictors of hypoxaemia
in children with pneumonia. Ann Trop Paediatr. 1988;18:31 40
21. Usen S, Weber M, Mulholland K, et al. Clinical predictors of
hypoxaemia in Gambian children with acute lower respiratory
tract infection: prospective cohort study. BMJ. 1999;318:86 91
22. ODempsey TJD, Todd JE. Chest infections in African children
[letter]. BMJ. 1993;306:1342
23. Wang EE, Milner RA, Navas L, Maj H. Observer agreement for
respiratory signs and oximetry in infants hospitalized with
lower respiratory infections. Am Rev Respir Dis. 1992;145:
106 109
24. Reuland DS, Steinhoff MC, Gilman RH, et al. Prevalence and
prediction of hypoxemia in children with respiratory infections
in the Peruvian Andes. J Pediatr. 1991;119:900 906
25. Dobson MB. Oxygen concentrators and cylinders. Int J Tuberc
Lung Dis. 2001;5:527532