Brief report. Intestinal parasites in Malaysian children with cancer
Journal of Tropical Pediatrics
Intestinal Parasites in Malaysian Children with Cancer
by Bina S. Menon <> 0 1 2
Mohd. Shukri Abdullah 0 1 2
Faridah Mahamud 0 1 2
Balbir Singh 0 1 2
0 Department of Paediatrics, Universiti Sains Malaysia , Kubang Kerian 16150, Kelantan, Malaysia. Tel. 00 60 9 760 2195; Fax 00 60 9 765 3370. E-mail
1 Departments of
2 Immunology and Infectious Diseases, Harvard School of Public Health , 677 Huntington Avenue, SPH1, Room 815, Boston, MA 02115 , USA
In this prospective study, we examined stool specimens from children with cancer receiving chemotherapy who were admitted for fever to the Universiti Sains Malaysia Hospital in Kota Baru, Kelantan. Stool specimens were examined for ova and cysts of parasites. Over a period of 15 months, there were 129 febrile episodes in 50 children with cancer and, in all, 237 stool specimens were examined. Sixty-six per cent of febrile episodes were associated with neutropenia and 9 per cent were associated with diarrhoea. Stool parasites were found in 42 per cent of children. The most common were helminths, followed by protozoa. Trichuris trichiura was the most common parasite (24 per cent), followed by Ascaris lumbricoides (22 per cent). Hookworm was found in 2 per cent. Giardia lamblia was found in 6 per cent of children, Blastocystis hominis in 4 per cent, and Cryptosporidium parvum in 2 per cent.
Enteric parasitic diseases are prevalent in Malaysia,
particularly in disadvantaged communities.1 Kelantan is
a state in the north-east of peninsular Malaysia, with a
largely rural population. A study in 1992 showed a high
prevalence of Cryptosporidium parvum (11.4 per cent)
in children with diarrhoea from this state.2 This organism
has been reported as a cause of life-threatening diarrhoea
in children with cancer.3 The aim of this study was to
determine the prevalence of Cryptosporidium parvum as
well as other stool parasites in children with cancer.
Patients and Methods
This was a prospective study over 15 months from
August 1996 to October 1997. Three stool specimens
were collected on consecutive days from children with
cancer receiving chemotherapy who were admitted for
fever to the Hospital Universiti Sains Malaysia in Kota
Baru. Fever was de®ned as a temperature of 388C on two
occasions 4 h apart or > 388C on one occasion. A
questionnaire was completed for each patient
documenting clinical information such as diarrhoea, animal
contact, and neutropenia. Diarrhoea was de®ned as an
One child had a positive stool specimen for C. parvum.
This was a 2-year-old girl with Down syndrome and
bilateral retinoblastoma. She was not neutropenic at the
time and did not have loose stools. However, she did
have bronchopneumonia. Three months earlier she had
diarrhoea lasting 1 week but stool samples were negative
for C. parvum. Four further stool samples 1 and 3 months
following the positive sample were also negative. There
was no history of animal contact.
Three children had positive stool samples for Giardia
lambliaÐtwo had symptoms of diarrhoea. One child had
profuse diarrhoea (20 times/day) and abdominal pain.
She was neutropenic at the time and had numerous
trophozoites and cysts in her stool. She was treated with
metronidazole as well as broad-spectrum antibiotics.
Despite this, she died, most probably due to a
multiresistant bacterial septicaemia.
Two children were positive for Blastocystis hominis, one
of whom had diarrhoea. This child also had B. hominis in
his stool 2 weeks prior to the diarrhoea when he was
febrile but had no loose stools.
Sixteen children (32 per cent) had helminthiasis and
seven children (14 per cent) had more than one helminth.
Twelve children were positive for Trichuris trichiura
and 11 for Ascaris lumbricoides. Only one child had
hookworm ovaÐthis child had a mixed infection with
Giardia lamblia as well as Ascaris lumbricoides.
Parasites are reported to be rare pathogens (1 per cent of
infectious episodes) in neutropenic patients in developed
countries.4 The prevalence in developing countries is not
known. Our study showed that 42 per cent of children
with cancer were positive for stool parasites. Our
numbers, however, were too small to show any
signi®cant association with neutropenia. The majority
of children infected with helminths were asymptomatic;
only one child had hookworm infection which might
have exacerbated the anaemia.
C. parvum was found in only 2 per cent of children
with cancer despite a history of animal exposure in 50
per cent. C. parvum has been transmitted from infected
domestic pets5 as well as cattle.6 The index case had
bronchopneumonia but no diarrhoea. C. parvum is
known to cause respiratory disease.7 However, in this
case, we cannot be certain that the protozoan was the
cause as bronchial washings were not done. Only one
of the 10 stool samples from this patient was positive for
C. parvum. Multiple stool samples are necessary due to
intermittent oocyst excretion.
In Mexico, C. parvum was found only in the diarrhoeal
stools of adult cancer patients.8 There were few
diarrhoeal episodes in our study, which may explain
the low prevalence of the organism. However, a large
study in India in 560 cancer patients with diarrhoea
showed a similar prevalence to ours of 1.3 per cent.9
Two children had signi®cant symptoms due to
giardiasis. Severe giardiasis has been reported previously
in a child on chemotherapy.10 The death in our case was
attributed to bacterial infection rather than giardiasis.
There is controversy as to whether Blastocystis hominis
is a pathogen in humans.11 Both children with this
organism had negative stool samples subsequently
without any speci®c treatment.
In conclusion, we found a high prevalence of enteric
parasites in children with cancer in Kelantan, Malaysia.
However, this was mainly due to helminthiases rather
than protozoal infections and the majority of patients
Cure of b-Thalassaemia Major by Umbilical Cord Blood
Transplantation ± A Case Report of Malaysia's First Cord
by Chan Lee-Lee MRCP (UK) and Lin Hai-Peng FRCP (Edin)
Department of Paediatrics, University of Malaya, Kuala Lumpur, Malaysia
A 25-month-old boy with b-thalassaemia major was presented with an opportunity for umbilical cord
blood transplantation when his unborn sibling was diagnosed in utero to be a b-thalassaemia carrier
and also human leucocyte antigen compatible. A barely adequate amount of cord blood was collected
at the birth of his sibling and infused into the patient after appropriate chemo-conditioning.
Engraftment occurred without major complications. The subject is now alive and well 9 months
posttransplant, thus marking our ®rst success in umbilical cord blood transplantation.
b-Thalassaemia major is one of the commonest
inherited haematological disorders in the Malaysian
population.1 Current optimal treatment requires regular
transfusion with leuco®ltered blood and adequate iron
chelation. There is ample evidence that this disease can
be cured by human leucocyte antigen (HLA) matched
sibling bone marrow transplantation (BMT) both from
experience overseas2;3 and locally.4 Indeed, in a country
where blood transfusion services or iron chelation
therapy are not optimal, bone marrow transplantation
should become the treatment of choice. The patient's
genetically abnormal marrow is destroyed by
chemoconditioning and replaced by healthy haematopoietic
stem cells from an appropriate donor who is almost
always an HLA-matched sibling. The use of mismatched
sibling or matched unrelated donors is not generally
recommended5 as such transplants have been beset by
graft rejections and early deaths.6
Another source of haematopoietic stem cells is
umbilical cord blood (UCB). This usually discarded
material has been shown to contain the requisite stem
cells for engraftment in patients with malignant and
non-malignant diseases.7;8 The greater proliferative
potential of UCB cells9 and their relative immunological
naivety10;11 have resulted in their increased application
in the ®elds of stem cell transplantation. The ®rst
matched sibling umbilical cord blood transplantation
(UCBT) occurred in 1988 for a patient with Fanconi's
anaemia.12 The ®rst UCBT for b-thalassaemia major
Correspondence: Associate Professor Chan Lee-Lee,
Department of Paediatrics, University Hospital, 50603 Lembah Pantai,
Kuala Lumpur, Malaysia. Tel. 603 750 2065; Fax 603 755
6114. E-mail <>.
was performed in Thailand in June 1993.13 Thus far only
10 patients with this disease have received transplanted
UCB.8;14ÿ16 Eight out of 10 were completely successful.
Success with matched sibling UCBT for b-thalassaemia
major is dependent on several factors. One of these is
the ability to diagnose the fetus in utero as being free
of the disease. Normal or thalassaemia trait donors
are acceptable. Another requirement would be HLA
compatibility between the fetus and the recipient. The
expertise to collect UCB and the technology to
cryopreserve stem cells play an important role. Last but not
least the adequacy and suitability of the collected UCB
can determine the success of UCBT. We report our ®rst
experience in the emerging ®eld of UCBT in a patient
with b-thalassaemia major.
TEH presented at the age of 3 months with pallor and was
diagnosed as having b-thalassaemia major. As his only
sibling, an elder sister, was found to be HLA
incompatible, the patient could only be treated conservatively
with regular blood transfusions.
A new sibling was conceived when the patient was
7 months old. In utero examination of a sample of
chorionic villous biopsy by molecular techniques for
the b gene mutation identi®ed the fetus to be a carrier. In
October 1996 a live female baby was born by normal
vaginal delivery and 60 ml of UCB were collected. A
small portion of the UCB was sent for HLA typing while
the remainder was cryopreserved in 10 per cent
dimethylsulfoxide (DMSO) and kept frozen in a liquid
nitrogen storage tank. The total numbers of nucleated
cells and CD34 cells in the collected cord blood were
65 ´ 107 and 0:54 ´ 106 respectively. Infection screen
on the UCB and newborn sibling con®rmed the absence
of hepatitis B, hepatitis C, and human immunode®ciency
viruses. Results of HLA typing showed the UCB to be
compatible with the patient at six out of six HLA loci.
The recipient was prepared for cord blood
transplantation by a combination of busulphan at 22 mg/kg and
cyclophosphamide at 200 mg/kg recipient body weight.
On 31 July 1997, the cryopreserved cord blood was
rapidly thawed in a water bath at 378C and infused into
the recipient. Apart from some patient discomfort
manifested by crying and mild restlessness, there were
no untoward side-effects.
The post-transplant recovery was complicated by mild
veno-occlussive disease of the liver, which responded
to conservative management, and methicillin-resistant
Staphyloccoccus epidermidis septicaemia, necessitating
a change of his central venous catheter. Haematopoiesis
was slow and the patient's total white blood cell count
exceeded 1000=ml on day 59 with platelets exceeding
50 000=ml on day 74. Bone marrow examination on day
88 con®rmed trilineage engraftment while chromosomal
analysis demonstrated full chimerism with a karyotype
of 46XX. When reviewed in April 1998 the patient
remained well and free of blood transfusions.
Patients with b-thalassaemia major who are treated
conservatively with optimal blood transfusions and iron
chelation are now living up to their late 30s.17ÿ19 The
introduction of iron chelation with desferrioxamine has
contributed to this improved survival,20;21 albeit
accompanied by restrictions in lifestyle related to daily
subcutaneous infusions, drug toxicities, and late
endocrine abnormalities. The possibility of cure from this
autosomal recessive disorder by stem cell
transplantation brings hope and cheer to patients, families, and
physicians alike. Unfortunately not many patients with
thalassaemia major will meet the criteria for successful
transplantation. These include availability of a matched
sibling donor, low number of blood transfusions received
(related to alloimmunization and graft rejection), and
absence of signi®cant liver damage from hepatitis or
On average only 30 per cent of patients would be able
to ®nd a matched sibling donor for transplantation.22 The
elder sister of our patient was HLA incompatible and he
was lucky when the newborn sibling was found to be
HLA compatible with him.
Couples who already have a child with thalassaemia
major are faced with at least two immediate dilemmas.
One involves the question of having more children. This
case demonstrates the utility of DNA technology in their
decision making. In utero diagnosis by chorionic villous
sampling using molecular studies on b gene mutations
are informative for 90 per cent of couples.23 Having
made the decision to have another child and being able to
exclude the diagnosis of b-thalassaemia major in utero,
couples are then confronted with the anxiety of whether
the fetus would be a suitable donor for their affected
child. In fact some couples try for another child on the 25
per cent chance of conceiving an HLA-matched sibling.
A second dilemma involves the decision to transplant
or not. Indeed, some couples agonize over the decision to
proceed with transplantation even when a suitable donor
has been identi®ed because of the current
transplantrelated mortality of approximately 5±10 per cent for the
recipient. Determination of compatibility by HLA
testing can be performed on chorionic villous samples but
this is not often done and hence expertise and accuracy
are limited. In utero determination of HLA compatibility
has not been performed in Malaysia. More often, as with
our patient, a small sample of the cord blood collected
at delivery is sent for HLA typing.
UCBT is undoubtedly gaining momentum. The
relative ease of UCB collection, lack of risks to the
donor, and immediate availability make it an attractive
option compared with bone marrow or peripheral blood
stem cell transplantation. Although all but one of the 10
cases of thalassaemia treated by UCBT used matched
sibling donors,8;14ÿ16 it is expected that more
mismatched UCBT will be performed in future. Early results
from mismatched UCBT are encouraging with low
levels of graft rejection and lower than expected rates of
graft-versus-host disease.24 This may be due to the lower
T-cell reactivity in UCB, hence the term
`immunologically naive'.11 This is one of the major advantages of
UCB over bone marrow or peripheral blood stem cells
and whether the lower incidence of graft-versus-host
disease will stand the test of time remains to be seen.
The availability of expertise and facilities for in utero
diagnosis of b haemoglobinopathy and stem cell
collection and cryopreservation in the University
Hospital, Kuala Lumpur is the culmination of years of work
in the ®eld of stem cell transplantation. Traditionally,
HLA-matched bone marrow stem cells have been used
for thalassaemia transplantation with reasonable success.
Our experience using matched sibling UCBT in this
patient is encouraging but insuf®cient for any de®nite
conclusions to be drawn. Even as issues relating to
incidence of acute and chronic graft-versus-host
disease, UCB banking,25ÿ27 and inadvertent
transmission of genetic disease are being addressed, the
application of UCBT, especially in the unrelated donor setting,
is bound to widen.
Adenosine Deaminase in Childhood Pulmonary
Tuberculosis: Diagnostic Value in Serum
by Necdet Kuyucu MD, CemsËit Karakurt MD, Eris BilalogÏ lu MD, Candemir Karacan MD and Tahsin TezicË MD
Dr Sami Ulus Children's Hospital, Ankara, Turkey
The diagnostic value of serum adenosine deaminase (ADA) activity was evaluated in childhood
pulmonary tuberculosis. Serum ADA levels were measured in 20 children diagnosed with pulmonary
tuberculosis (group 1) and 150 children (group 2) including 128 with tuberculosis infection (Mantoux
test positive) and 22 healthy children. In group 1, the mean serum ADA activity was 74:06 6 18:5 U/l,
which was signi®cantly ( p < 0:001) higher than that of group 2 (40:36 6 12:0 U/l). A serum ADA level
of $53:76 U/l had a sensitivity of 100 per cent, speci®city of 90.7 per cent, positive predictive value of
58.8 per cent, and a negative predictive value of 100 per cent in children with tuberculosis disease. To
conclude, measurement of serum ADA activity was a useful diagnostic tool in childhood pulmonary
Tuberculosis is still one of the leading causes of
morbidity and mortality in the world despite continued
multinational efforts to control the disease.
Unfortunately, diagnosis of tuberculosis infection and disease is
somewhat dif®cult in children.1 In recent years there has
been a desire for the development of new microbiologic,
genetic, immunologic, and biochemical methods for the
rapid and accurate diagnosis of tuberculosis. One such
biochemical method is measurement of the adenosine
deaminase (ADA) activity, which has been proposed
to be a useful marker for tuberculosis disease in the
pleura, pericardium, peritoneum, and central nervous
The aim of this study was to evaluate the diagnostic
value of serum ADA activity measurement in pulmonary
Materials and Methods
The study group comprised 170 2±14-year-old children
who had undergone tuberculosis disease and infection
screening at the outpatient clinics. In all children, a
detailed history about symptoms suggestive of
tuberculosis and a past history of tuberculosis were obtained.
Physical examinations and chest X-rays were performed
and BCG scar numbers were determined in all
participants. Tuberculin skin testing with 5 U of puri®ed
protein derivative (PPD) was administered by the
standard technique. Acid-fast bacilli examination in
early morning gastric aspirates or sputum by Ziehl±
Neelsen staining was done on three consecutive days if
there was a clinical suspicion of pulmonary tuberculosis.
If necessary, specimens were examined by radiometric
culture (BACTEC) and polymerase chain reaction (PCR)
for M. tuberculosis. In all children, a detailed family
history of tuberculosis was obtained and family
screening by tuberculin testing, chest X-ray, and, when needed,
smear examination of acid-fast bacilli in sputum was
According to the ®nal diagnosis, cases were
subdivided into two groups. Group 1 included 20 children
diagnosed as having pulmonary tuberculosis by physical,
radiological, and microbiological ®ndings compatible
with the disease. Group 2 consisted of 150 children who
did not have tuberculosis disease. Of these children, 128
had a PPD induration >15 min and positive family
history, but were lacking physical and laboratory
Correspondence: Necdet Kuyucu MD, ZiyagoÈkalp Cad. 62/6
06600 oÈncebeci, Ankara, Turkey. Tel. 90 312 317 0707/281;
Fax 90 312 317 0353. E-mail <>.
®ndings of tuberculosis. The remaining 22 were healthy
children having a PPD #15 mm.
The adenosine deaminase activity was determined
in the serum of all 170 children by the calorimetric
method of Giusti,5 which is based on the measurement
of ammonia produced when adenosine deaminase acts
on an excess of adenosine.
The Mann±Whitney U test was used in statistical
evaluation of the data. The diagnostic value of the ADA
was assessed in terms of sensitivity, speci®city, and
positive and negative predictive values.
The age, tuberculin reaction size, and serum ADA
enzyme activity values of the subjects are presented in
Table 1. Mean induration size of group 1 and group 2
were similar ( p < 0:05). Serum ADA activity was
signi®cantly ( p < 0:001) higher in group 1 than in
Taking the 80th percentile value (53.76 U/l) serum
level of ADA as a cut-off point, there was a speci®city
of 90.7 per cent and sensitivity of 100 per cent (Table 2).
However, with higher values of ADA activity (85th,
90th, and 95th percentiles) the speci®city increased
but the sensitivity decreased. In contrast, with lower
values of the activity (70th and 75th) the sensitivity
didn't change but the speci®city decreased. A cut-off
value of $53:76 U=l for serum ADA therefore seems to
suggest a diagnosis of pulmonary tuberculosis with a
positive predictive value of 58.8 per cent and negative
predictive value of 100 per cent.
Adenosine deaminase is essential for the differentiation
of lymphoid cells, particularly T cells, and plays a role
in the maturation of monocytes to macrophages.6 ADA
is considered to be an indicator of cell-mediated
immunity.7 In recent years, measurement of ADA in
pleural, pericardial, meningeal, and peritoneal effusions
has gained importance in the diagnosis of
tuberculosis.2ÿ4 The diagnostic value of serum ADA is pulmonary
tuberculosis has been investigated in only a few
studies4;8ÿ10 which revealed its usefulness as a
The immunologic reactions taking place during the
course of tuberculosis infection and disease are
complex. Especially, the factors responsible for the
elevation of serum ADA activity during tuberculosis disease
are not yet clear. CD4 and CD8 a b T cells play a pivotal
role in the development of delayed type hypersensitivity
and in the control of both tuberculosis infection and
All values mean 6 SD.
disease.8 Elevation of ADA activity, especially in pleural
effusions, has been proposed to result from an increase in
the T-cell population, notably the immature and reactive
cells.9;10 However, in other studies no correlation has
been found between ADA activity and T-cell CD4 and
CD8 numbers or the CD4/CD8 ratio.11;12
Our results suggest that the determination of serum
ADA activity has a high sensitivity and speci®city for
the diagnosis of tuberculosis disease. A serum value
of $53:76 U=l has been found to be optimal in
differentiating tuberculosis disease from the infection and
healthy controls. On the other hand, the present
study could not reveal any diagnostic value of the
measurement of serum ADA activity in tuberculosis
Bhargava et al.4 found a mean serum ADA activity
of 78:12 6 17 U=l in adult patients with pulmonary
tuberculosis and they concluded that values above
54 U/l had a speci®city of 97.6 per cent and sensitivity
of 81.5 per cent in the diagnosis of pulmonary
tuberculosis. Our results are in agreement.
In conclusion, we would like to point out that the
measurement of ADA activity in the serum as well as
peritoneal or pleural ¯uids may be a bene®cial
diagnostic tool in childhood tuberculosis disease.
Classi®cation Trees and Logistic Regression Applied to
Prognostic Studies: A Comparison using Meningococcal
Disease as an Example
by Guilherme L. Werneck*;** MD, MSc, Diana M. de Carvalho** MD, DPH, David E. Barroso** MD, PhD,
Earl F. Cook* DSc, and Alexander M. Walker* MD, DPH
*Department of Epidemiology and Department of Immunology and Infectious Diseases, Harvard School of Public
Health, Boston, USA **Department of Preventive Medicine, NESC, Federal University of Rio de Janeiro, Brazil
The authors used logistic regression and classi®cation trees to develop prediction models for fatal
outcomes in meningococcal disease in a cohort of 829 children hospitalized for meningococcal disease
during 1989±1990 in Rio de Janeiro. The area under the receiver operator characteristic (ROC)
curve was 92 per cent for logistic regression and 88 per cent for classi®cation trees. Logistic
regression may be preferred when the main objective is to obtain explicit measures for statistical
inference and measures of the force of the association between each variable and the outcome.
However, estimation of the probability of dying for each patient involves manipulation of the logistic
regression formula, which would not easily be done in an emergency room. Classi®cation trees
provided comparable discrimination between fatal and non-fatal outcomes, and yielded a graphical
display of the results that is easier to understand and is straightforward to apply in clinical settings.
Meningococcal disease (MD) is hyperendemic in the city
of Rio de Janeiro, Brazil, with incidence rates of around
5 per 100 000 for the past decade.1;2 During these years
the case-fatality rates have remained between 15 and 20
One possible way to deal with the problem of high
fatality rates is to identify prognostic factors that can
easily be assessed and used to aid clinical decision
making. Much effort has been invested in the
development of prognostic scores to predict mortality from
MD.3±11 However, much of this work has been done in
developed countries and the results may not be directly
applicable in other settings. As a preliminary effort to
build a predictive model for MD in a developing country,
this study compares the ability of logistic regression and
classi®cation trees to discriminate between fatal and
Dr Werneck was partially supported by the Ministry of
Education/CAPES (Brazil). This work was supported by the
Harvard Pharmacoepidemiology Teaching and Training Program.
non-fatal cases in a cohort of children from Rio de
Subjects and Methods
During 1989±1990 a total of 829 MD cases (< 16 years)
were admitted to the Instituto Estadual de Infectologia
SaÄo SebastiaÄo in the city of Rio de Janeiro, Brazil. Cases
of MD were de®ned by the presence of one of the
) isolation of N. meningitidis from blood or
cerebrospinal ¯uid (CSF);
) identi®cation of Gram-negative diplococci or
meningococcal antigens in CSF; or
) typical clinical picture with fever and haemorrhagic
Outcome status and putative prognostic factors were
obtained from medical and epidemiological surveillance
We undertook the analysis in two stages. First, we
used multiple logistic regression and classi®cation trees
to identify the most important independent prognostic
factors for death in MD. Second, the predicted probability
of death for each individual in the cohort was estimated
from each model, and then compared to the actual
outcomes. We used the area under the receiver operator
characteristic (ROC) curve, sensitivity, speci®city, and
positive and negative predictive values as the criteria to
assess the performance of these models. Multiple logistic
regression was carried out in STATA.12 Classi®cation
trees were performed using S-Plus.13
In logistic regression, backward elimination was used
to select the signi®cant prognostic factors to be in the
®nal model. A 10 per cent signi®cant level was chosen.
We tested for two-way interactions, but none were
Classi®cation trees (or CART, for classi®cation and
regression trees) provide an alternative to logistic models
for classi®cation problems.13;14 CART builds a binary
classi®cation system (tree) through recursive
partitioning, so a data set is successfully split into increasingly
homogeneous subgroups.13 At each stage (node) the
CART algorithm selects the explanatory variable and
splitting value that gives the best discrimination between
two outcome classes.
A full CART algorithm adds nodes until they are
homogeneous or contain few observations (#5 is the
standard cut-off in S-plus).14 CART creates a full tree
that has a minimal misclassi®cation rate, but may have a
poor predictive power for a new sample, since it may
be too closely tied to the original data (the `learning
sample').15 The problem of creating a useful tree is to
®nd a suitable guideline to cut back (to `prune') the tree.
The general principle of pruning is that the tree of best
size would have the lowest misclassi®cation rate for
individuals not included in the learning sample.15 If a
second data set is available (the `validation sample'), one
could apply trees of various sizes to it and then choose
the one with the lowest misclassi®cation rate. If no
validation sample is available it is possible to make one
by dividing the learning sample. CART performs this
approach using the method of cross-validation.14
Crossvalidation works by dividing the learning sample into
groups of equal size, building the tree on part of the data,
and then assessing the tree misclassi®cation rate on the
remaining part of the data. We used cross-validation,
splitting the data into four groups.
Table 1 presents the variables signi®cantly associated
with death in the logistic regression model. Age above 1
year, neck stiffness, and a longer duration of disease
were associated with a lower risk of death. Seizures,
diarrhoea, a clinical diagnosis of shock, focal
neurological sign, and residence outside Rio de Janeiro city were
positively associated with death.
Figure 1 presents the best classi®cation tree obtained
by the cross-validation procedure. There are six terminal
nodes (shaded boxes), to be compared with a total of 53
in the complete generated tree. For each node in the tree,
the numbers of fatal and non-fatal cases and the variable
used to split the parent node are displayed. The
percentages displayed under each terminal node
represent the risk of death among those who eventually
reached this node.
Table 2 compares the predictive power of the two
aThe ROC curve provides a summary of the discriminant ability of the model over all possible predicted values associated with
variations in the cut-off point in the estimated probability of dying.
bSensitivity, speci®city, positive and negative predictive values are calculated on the basis of a `prediction' of death for any individual
for whom the estimated probability of dying is greater than 10 per cent.
techniques in terms of the area under the ROC curve,
speci®city, sensitivity, and positive and negative
predictive values. Logistic regression has a marginally
better overall performance than the classi®cation tree.
The main objective of this study was to compare the
performance of logistic regression and classi®cation
trees in identifying the best prognostic system for death
in MD. The prognostic models generated by the two
techniques are statistically equivalent. The two methods
seem to complement each other. Logistic regression
provides explicit measures for statistical inference and
measures of the force of the association between each
variable and the outcome. However, estimating of the
probability of dying for each patient involves
manipulation of the logistic regression formula, which would not
easily be done in an emergency room. The tree model
provided comparable discrimination between fatal and
non-fatal cases using only ®ve variables (logistic
regression used eight). The graphical display of the
results from a tree is easier to understand and is
straightforward to apply in clinical settings.
Prognostic systems based only on clinical data are
thought to be not as powerful as those based on
laboratory ®ndings. Nevertheless, they are simple and
inexpensive, and they may be more useful in developing
countries. The prognostic models developed in this study
are still preliminary and need to be improved and
validated using more recent samples. In order to take
advantage of the complementary information provided
by the two techniques we suggest considering both
approaches when developing prognostic systems.
1. Gama SGN , Marzochi KBF , Silveira-Filho GB . CaracterizacË aÄo epidemioloÂgica da doencËa meningocoÂcica na aÂrea metropolitana do Rio de Janeiro , Brazil, 1976 a 1994 . Rev SauÂde puÂbl 1997 ; 31 : 254 ± 62 .
2. Noronha CP , Baran M , Nicholai CCA et al. Epidemiologia de doencËa meningocoÂcica na cidade do Rio de Janeiro: modi®cacËaÄo apoÂs vacinacË aÄo contra os sorogrupos B e C . Cadernos de SauÂde PuÂblica 1997 ; 13 : 295 ± 303 .
3. Kirsch EA , Barton RP , Kitchen L , Giroir BP . Pathophysiology, treatment and outcome of meningococcemia: a review and recent experience . Pediatr Infect Dis J 1996 ; 15 : 967 ± 79 .
4. Leclerc F , Delepoulle F , Diependaele JF et al. Severity scores in meningococcal septicemia and severe infectious purpura with shock . Intensive Care Med 1995 ; 21 : 264 ± 5 .
5. Flaegstad T , Kaaresen PI , Stokland T , Guttenberg T. Factors associated with fatal outcomes in childhood meningococcal disease . Acta Paediatr 1995 ; 84 : 1137 ± 42 .
6. Nurnberger W , Platonov A , Stannigel H et al. De® nition of a new score for severity of generalized Neisseria meningitidis infection . Eur J Pediatr 1995 ; 154 : 896 ± 900 .
7. Algren JT , Lal S , Cutliff A , Richman BJ . Predictors of outcome in acute meningococcal infection in children . Crit Care Med 1993 ; 21 : 447 ± 52 .
8. Tuysuz B , Ozlu I , Aji DY , Erginel A . Prognostic factors in meningococcal disease and a new scoring system . Acta Paediatr 1993 ; 82 : 1053 ± 6 .
9. Fakhir S , Ahmad SH , Ahmad P . Prognostic factors in¯uencing mortality in meningococcal meningitis . Ann Trop Paediatr 1992 ; 12 : 149 ± 54 .
10. Tesoro LJ , Selbst SM . Factors affecting outcome in meningococcal infections . Am J Dis Child 1991 ; 145 : 218 ± 20 .
11. Gedde-Dahl TW , Bjark P , Hoiby EA , Host JH , Bruun JN . Severity of meningococcal disease: assessment by factors and scores and implications for patient management . Rev Infect Dis 1990 ; 12 : 973 ± 92 .
12. StataCorp. Stata Statistical Software: Release 5 .0. Stata Corporation , College Station 1997 ; 271 ± 307 , 343 ± 67 . (Software information available at http://www.statacom/.)
13. Statistical Sciences . S-PLUS guide to statistical and mathematical analysis , version 3 .3. StatSci, Seattle 1995 : 10 .1± 10 . 35 . (Software information available at http:// www.mathsoft.com/.)
14. Clark LA , Pregibon D . Tree-based models . In: Chambers JM , Hastie TJ (eds), Statistical Models. Chapman & Hall , New York, 1993 ; 337 ± 419 .
15. Efron B , Tibshirani R . Statistical data analysis in the computer age . Science 1991 ; 253 : 390 ± 5 .