Portal de Periódicos da CAPES

Methodology for a multicenter study of serious infections in young infants in developing countries

1999; Lippincott Williams & Wilkins; Volume: 18; Issue: Supplement Linguagem: Inglês

10.1097/00006454-199910001-00003

1532-0987

NULL AUTHOR_ID,

Child Nutrition and Water Access

This study was designed to evaluate both the clinical features that predict serious infections and the etiologic agents responsible for those infections among infants younger than 3 months of age in developing countries. The rationale for the study is outlined in an accompanying paper.1 Previous studies addressing this question have suffered from a number of methodologic problems including the limited number of infants with severe outcomes and the difficulty of identifying cases at the time of care seeking. From the standpoint of developing a clinical prediction rule, these issues are important because of the need to have an adequate number of outcome events so that multiple predictor variables can be included simultaneously in a model. From the standpoint of identifying etiologic agents, low isolation rates mean that a large number of sick infants must be studied to understand the relative importance of the different pathogens. Thus a multicenter study was required to address the study goals. This paper describes the design of the study and how such methodologic and logistic issues were addressed. In addition we have attempted to summarize lessons learned, because these may be relevant to future similar studies. METHODS Study sites. Previous studies of the etiology of neonatal sepsis in developing countries contained many hospital-acquired infections, yet most deaths from neonatal infections occur in the community. We chose to focus on community-acquired infections. There are a number of possible points in the referral system at which such a study could be considered. A proposal for a community-based component in which infants would be enrolled in their homes was rejected because the objective was to assess the value of clinical findings at the time families seek care and because enrolling infants who became sick in the home would markedly increase the number of children needed. Sites were chosen where reported infant mortality exceeded 40/1000. Other criteria were the availability of laboratory facilities, previous experience in this type of study and the presence of a clinical facility in which infants could be enrolled in an outpatient clinic where they sought care. Four research institutes participated in the project: the Medical Research Council Hospital, Fajara, The Gambia; the Ethio-Swedish Children's Hospital, Addis Ababa, Ethiopia; the Research Institute for Tropical Medicine, Manila, Philippines; and the Papua New Guinea Institute of Medical Research, Goroka, Papua New Guinea (PNG). All the institutes are associated with clinical sites that serve as both primary care facilities for large urban or periurban populations and secondary or tertiary referral centers. A fifth site in Latin America was abandoned in the planning phase of the study. The study began in The Gambia in September, 1990, and continued until July, 1993, when the final infant was recruited in Ethiopia. In The Gambia the study was performed in two sites in each of 2 successive years (September, 1990, to December, 1992). The infant mortality rate in The Gambia is ∼100/1000 live births, of which ∼40 deaths/1000 live births are in neonates.2 In the first year of the study the outpatient facility at the Medical Research Council Hospital in Fajara was used to recruit infants. This facility provides primary care for a densely populated urban and periurban coastal region of The Gambia. In the second year the Royal Victoria Hospital in the Gambian capital, Banjul, was used for enrollment. This is the only pediatric referral hospital in The Gambia. It receives cases from health centers in the western half of the country and from the other referral hospital in the eastern part of the country. Thus infants recruited in the first year were self-referred, usually presenting for the first time at a health facility, whereas those seen in the second year of the study were mostly referred from health centers or smaller hospitals for the evaluation of potentially serious illness. In Ethiopia the study was carried out over 2 years at the Ethio-Swedish Children's Hospital in Addis Ababa (August, 1991, to July, 1993). This hospital acts as a primary care facility as well as a pediatric referral hospital for Addis Ababa and the surrounding districts. The infant mortality rate in Ethiopia is estimated to be 92/1000 live births3 with a range of 66/1000 live births in urban areas to >190 in rural areas. Acute respiratory infections account for ∼20% of those deaths.4, 5 More than 50% of the infant mortality in Ethiopia is believed to be the result of deaths in infants <2 months of age. In The Philippines three hospitals were used to recruit infants over the 2-year study period (March, 1991, to March, 1993). In the first study year recruitment took place at the Research Institute for Tropical Medicine and the Philippine General Hospital. These are government hospitals that serve the poor communities of Manila and the surrounding semiurban areas. In the second study year Quezon City General Hospital was included. Like the other 2 hospitals it is a government facility located in the northern part of metropolitan Manila. All 3 institutions see referred patients from the surrounding health facilities as well as primary care patients who present directly to the hospital. The Philippines Ministry of Health estimates infant mortality to be 21/1000 live births of which 14/1000 occur in infants younger than 1 month of age and 5/1000 occur in infants younger than 3 months of age.6 These figures are thought to be underestimates. A report from the United Nations Development Program estimated the infant mortality to be 44/1000 live births.7 In Papua New Guinea (PNG) the study was carried out for 2 years at the children's outpatient department of Goroka Base Hospital (March, 1991, to April, 1993). Although the hospital serves as a referral hospital for the highlands region, most children attending the outpatient department come from the urban and periurban areas surrounding the town. Infant mortality in the study area was estimated to range from 32/1000 in the urban area to 84/1000 live births in the more remote rural areas. Neonatal mortality varies from 5/1000 in the town to 33/1000 in the remote rural areas.8 Patients. To evaluate the importance of age in the pattern of serious infections in young infants in developing countries, infants up to 90 days of age were recruited. At each of the study sites, all infants 90 days of age or less who were seen during study hours with an illness that began at home were triaged by a study nurse. Infants were eligible for inclusion if their rectal temperature was 37.5°C or more, or 35.5°C or less, or if any of the following symptoms were volunteered by the mother, agreed by her on questioning or observed by the health worker: cough; fast, noisy or difficult breathing; poor feeding; abnormally sleepy or difficult to wake; convulsions; or fever. They were also eligible if the mother volunteered that the infant was very sick, was irritable or had suffered a cyanotic or apneic episode, or if the health worker felt that the child looked ill. Excluded from the study were infants with: clinic attendance for routine care (e.g. immunizations); a chief complaint of trauma or burn; a birth weight <1500 g and less than 48 h old; illnesses that began in the hospital; an episode of sepsis, pneumonia or meningitis in the previous 3 weeks; transfer from another facility where the child had been hospitalized; congenital malformations; or previous study enrollment. Documentation of signs and symptoms. If informed consent was given, infants who met the criteria for inclusion in the study underwent a standardized history and physical examination to assess the presence or absence and the degree of severity of 51 presenting signs and symptoms commonly believed to be associated with the presence and severity of bacterial disease.9 Candidate predictors considered included demographic variables, historical variables, vital signs and physical examination findings (Table 1). In the Gambia, Ethiopia and the Philippines a pediatrician or a clinician with considerable pediatric experience performed the examination. In Papua New Guinea the examination was undertaken by study nurses with pediatric training or by the study pediatrician.TABLE 1: Clinical predictors evaluated in model development The importance of having full agreement between sites on the definitions of the symptoms and signs being documented in the study was recognized at the outset. Definitions of clinical signs were agreed on and circulated. Variability between observers at each site was minimized by training exercises before and during the study. Site visits were undertaken by the coordinator and expert consultants to help standardize the measurement of clinical findings. On-site training sessions were conducted in which all investigators observed sick infants and assessed them independently. Site teams used these sessions to discuss how to observe the clinical findings and how to complete data collection forms. Tape recordings of chest findings were used to help standardize auscultatory findings. Some principal investigators also visited other sites to compare clinical examination techniques and conduct training exercises. During the study, site teams conducted periodic exercises in which nonstudy infants were observed for clinical findings so that observations of the same infant by different examiners could be compared. After the examination was completed, nurses measured the oxygen saturation with a Nellcor N-200 pulse oximeter (Nellcor Inc., Hayward, CA). The measurements were made while the child was as near a resting state as possible and when there was a stable pulse tracing. The highest stable reading was recorded. This procedure was performed twice. Each reading was ranked as acceptable or poor according to the pulse signal and the infant's state during the reading. The mean of the readings was used as the infant's oxygen saturation if the two readings were rated equally. If one reading was rated as poor, the other reading was used. Indications for investigations. After the clinical evaluation infants who met the prespecified criteria underwent a laboratory evaluation that included blood culture, white blood cell count and chest radiograph Table 2 summarizes the criteria for laboratory evaluation. The protocol indicated that infants with a temperature of 38°C or more, or 35.5°C or less, or clinical signs suggesting meningitis should have a lumbar puncture performed. However, clinicians used their discretion on this point to avoid unnecessary lumbar punctures. Urine cultures were not obtained consistently because it was not feasible to collect suprapubic specimens or to catheterize infants in all the settings. Attending physicians could also investigate any child not meeting these criteria if they believed it was indicated clinically. Because some infants without presenting findings suggestive of bacterial illness might still have bacterial illness, the study protocol called for a random sample of 20% of the patients who did not meet criteria for laboratory testing to undergo chest radiograph, blood culture and blood count.TABLE 2: Criteria for investigation of infants who had completed a full history and physical examination* Pneumonia was diagnosed on the basis of chest radiographs. Chest radiographs were interpreted by a panel of pediatric radiologists in the US, UK and Switzerland. The radiologists received radiographs from the study sites and interpreted them without clinical information. Films were examined in a standardized manner and interpretations were recorded on a precoded form.10 After reviewing the films the radiologists indicated whether pneumonia was absent, probable or definite. Definite pneumonia required evidence of alveolar consolidation, whereas probable pneumonia was defined as the presence of parenchymal changes such as interstitial consolidation or atelectasis that fell short of the requirements for definite pneumonia. For the purposes of the analysis, study infants were defined as having pneumonia if all radiologists who read the film and ruled it interpretable, classified the radiograph as definitely or probably abnormal. Infants with uninterpretable films or with missing radiographic information were classified as having a normal radiograph. Treatment. Decisions regarding treatment were made on a clinical basis. Most infants who were admitted were treated with a penicillin plus gentamicin parenterally. Treatment was adjusted according to clinical state and bacteriologic and radiologic results. Duration of admission and arrangements for follow-up varied according to local conditions. An attempt was made to follow up all infants 1 to 2 weeks after discharge, or after the initial attendance for infants who were not hospitalized. Laboratory procedures.Collection and transport of specimens. After thorough cleansing with alcohol or iodine, the skin was allowed to dry and blood was collected, usually with a disposable 23-gauge butterfly needle from a vein in the cubital fossa or the dorsum of the hand. An equal volume of blood was injected into each of two blood culture bottles (Septicheck, Hoffmann la Roche, Basel, Switzerland). In most cases 1.0 ml of blood was injected per bottle. The remaining blood was injected into a tube containing EDTA for hematology and malaria blood film (in endemic areas) and a plain tube for the collection of serum. Nasopharyngeal aspirates (NPAs) were collected with a mucous aspirator and a mechanical suction device. An appropriately sized catheter was passed into the posterior nasopharynx and withdrawn with suction applied. The tubing was then washed through with 1.0 ml of phosphate-buffered saline. The extent of the virologic investigations varied between sites. Urine for culture was collected by one of three methods: suprapubic aspiration; use of an adhesive perineal bag; or a clean catch technique in which an assistant catches urine passed during the examination. Lumbar punctures were performed by sterile technique in the lateral position with 21- or 23-gauge hypodermic needles. Bacteriology. Blood for culture (0.5 to 1.5 ml) was inoculated into one bottle each of tryptic soy broth (TSB) and brain-heart infusion containing sodium polyanethol sulfonate (Roche Diagnostica, Basel, Switzerland). Where only a small volume of blood was obtained, TSB was given priority. The bottles were mixed and transported directly to the laboratory for incubation at 35-37°C. TSB bottles were vented with hemoline venting needles (BioMérieux, Lyon, France) before incubation and BCB slides (Roche Diagnostica, Basel, Switzerland) were screwed onto the bottles 6 h later. The slides were flooded daily with the broth of the blood culture bottles until growth became apparent. The brain-heart infusion broths were subcultured after 1, 2 and 7 days, and when turbidity was noted, onto 5% horse blood agar, MacConkey agar and Vitox (Unipath Ltd., Basingstoke, UK)-enriched chocolate agar plates. Plates were incubated for 24 to 48 h at 35-37°C aerobically (horse blood agar and MacConkey agar), anaerobically (horse blood agar) and in 5 to 10% CO2 (Vitox-enriched chocolate agar). Specimens of cerebrospinal fluid (CSF) were processed without delay.11 The macroscopic appearance was noted as clear, turbid or xanthochromic. A loopful of turbid CSF or the sediment from a centrifuged specimen was inoculated onto medium and incubated as for blood cultures. Uncentrifuged CSF was examined microscopically and the white blood cells counted by standard methods. Smears were stained by Gram and Giemsa methods to detect bacteria and to differentiate white blood cells. In addition in sites other than PNG, CSF specimens were tested for capsular polysaccharide antigens of H. influenzae type b, S. pneumoniae and Neisseria meningitidis by latex agglutination with Slidex Meningite kit 5 (BioMérieux). Bacterial isolates from blood or CSF were classified as to the likelihood that they were genuine pathogens. Criteria used included evidence that an organism is a known pathogen in neonates, the speed with which they were detected and the number of bottles in which the organism was observed. The urine specimens were processed as follows. A loopful of well-mixed urine was inoculated onto a plate of cystine lactose electrolyte-deficient agar with a sterile calibrated wire loop (0.001 ml). Plates were incubated aerobically at 35°C for 24 h and examined to determine bacterial counts. Only urine samples containing ≥105 colony-forming units/ml were regarded as having significant growth. All isolates were identified by standard biochemical and serologic techniques.11 Where cultures for Mycoplasma spp. and Ureaplasma spp. were performed, NPAs were taken to the laboratory in Mycoplasma and Ureaplasma transport media (Oxoid, Basingstoke, UK), inoculated into growth media and incubated at 35°C for 1 to 7 weeks and 24 to 48 h for identification of Mycoplasma spp. and Ureaplasma spp., respectively. In addition some NPAs in transport media were frozen at −70°C and shipped to the University of Alabama in Birmingham for further isolation and characterization of Mycoplasma spp. and Ureaplasma spp. In the Gambia NPAs were also inoculated onto chocolate agar containing 300 mg/l bacitracin (Sigma, St. Louis, MO) and blood agar containing 5.0 mg/l gentamicin sulfate (Rotexmedica, Trittau, Germany) for the selective isolation of H. influenzae and S. pneumoniae, respectively. Plates were incubated at 35-37°C in 5 to 10% CO2 for 24 to 48 h. For the isolation of Bordetella spp. calcium alginate nasopharygeal swabs were placed in half-strength charcoal agar supplemented with 20 mg/500 ml cephalexin (Unipath Ltd.), amphotericin B solution (5 ml/l) and 10% horse blood. Isolation cultures were set up in full-strength charcoal agar (Unipath Ltd.), supplemented as above and incubated in air at high humidity and 35-37°C for up to 7 days. Virology. Viral procedures varied among the sites according to the capabilities and experience of the local staff. Variations from these procedures are detailed in the site-specific papers.12-15 All performed indirect fluorescent antibody detection. Viral culture was performed only in the Philippines and the Gambia. In general the procedure was as follows. On arrival at the laboratory portions of NPAs were inoculated into cryotubes containing virus transport medium and 2-sucrose-phosphate Chlamydia transport medium and stored at −70°C. Broths for the recovery of Mycoplasma pneumoniae and Ureaplasma urealyticum were also inoculated with NPA specimens. Swabs dipped in the NPA were used to prepare smears for direct immunofluorescence with the use of commercially available monoclonal antibodies to Chlamydia trachomatis (Syva Corp., Palo Alto, CA). Confirmation of these findings by PCR is under way at two sites and will be reported in later publications. Cells in the remainder of the NPA were washed free of mucus, spotted onto microscope slides, dried, fixed in acetone and stored at −20°C. They were then stained with antisera with an immunofluorescence marker. Monoclonal antibodies were used for influenza A, adenovirus and respiratory syncytial virus and polyclonal antiserum was used for influenza B, parainfluenza types 1 and 3 provided by WHO16 and measles (Karolinska Institute, Stockholm, Sweden). Each specimen in virus transport medium was inoculated onto monolayers of Vero, NCI-H292, HEp-2 and MRC-5 cells (American Type Culture Collection, Manassas, VA) and incubated stationary at 37°C and rolling at 33°C. The monolayers were inspected every 2 to 3 days for a minimum of 14 days for evidence of viral replication. Isolates of influenza, parainfluenza, respiratory syncytial virus, measles and herpes simplex were identified by immunofluorescence. Isolates of adeno- and enteroviruses were typed by neutralization tests. Reference laboratories. Isolates of S. pneumoniae, H. influenzae and S. pyogenes were confirmed and serotyped at the reference laboratories. (Table 3). The reference laboratories also provided advice about processing of specimens at the sites and transportation. They were available to confirm identifications and to serologically type isolates of S. pneumoniae, H. influenzae and S. pyogenes. In PNG and The Gambia S. pneumoniae isolates were typed by either latex agglutination or the Quellung reaction with pool, group and type antisera obtained from Statens Seruminstitut, Copenhagen, Denmark. Factor typing and typing of isolates from the other sites were done by the reference laboratory. Encapsulated strains of H. influenzae were serotyped at each of the above laboratories by slide agglutination with type-specific antisera a to f (Murex Diagnostics, Dartford, UK). Beta-hemolytic streptococci isolated in PNG and The Gambia were Lancefield-grouped by latex agglutination (Streptslide II from Porton Cambridge, Newmarket, UK, and Streptex from Murex Biotech Ltd., Dartford, UK, respectively).TABLE 3: Microbiologic reference laboratories participating in the study The sites that did not use viral culture techniques (Ethiopia and PNG) sent slides to the reference laboratory for viral immunofluorescence. Because of concerns about the isolation rate for Mycoplasma spp. and Ureaplasma spp., special media were prepared at the reference laboratory and transported frozen to some of the sites, where they were stored at −70°C. Nasopharyngeal swabs were inoculated into the thawed media. Specimens were then refrozen and transported to the reference laboratory within 3 months. Sample size calculations. Because of the complexity of the proposed multivariate analysis of clinical predictors, the sample size for the study was estimated by estimating the sample size required for an agestratified analysis of individual clinical signs along the following lines. To evaluate a sign that is present in 50% of infants, of whom 10% have a serious outcome, and that has a sensitivity of 80% (95% confidence interval, 70%, 90%), 1200 infants would be needed in each of 3 proposed age strata. Therefore the study set out to enroll a total of 4000 infants in the 4 sites. This provides a conservative estimate of the sample size required for an analysis based on sophisticated multivariate modeling. Data management. As part of the necessary intersite standardization process, it was agreed at the outset that each site would use the same set of forms, with translations into local languages where appropriate. Data processing and quality control procedures were performed on-site, using standardized methodology. Customized software for data entry, corresponding to the data forms, was prepared using dBase III+. Partway through the study some sites changed to an equivalent data entry program using Epi-Info Version 5.0. Technical assistance regarding data entry and cleaning was made available to each site before data collection started and later through on-site visits. After data entry was completed it was submitted in batch form to the data manager where it was subjected to range checks and other basic data-cleaning procedures. Final cleaning of the data set was undertaken at WHO in Geneva. The study produced a large amount of data for each child that was investigated (∼1 kilobyte). The data collection instruments included in the study are listed in Table 4 (copies of the data collection instruments are available on request). Items administered as questions were translated and back translated into local languages. For example three sites identified local terms for pneumonia.TABLE 4: Data collection instruments used in the study Ethical issues. The study was approved by the Secretariat for Research In Human Subjects of the World Health Organization and by the ethical review boards at each site. At each site verbal consent was obtained from mothers entering the study. The institutional review boards determined that written consent was not required because the examinations and investigations performed on the infants were consistent with good clinical management. Analysis. The basic analysis was performed in Geneva using SAS for Windows statistical software. Analysis of the clinical predictors of serious illness was undertaken at Duke University, Durham, NC, and the University of Virginia using S-PLUS (MathSoft Inc., Seattle, WA). The details of the analysis are presented elsewhere.17 Decisions regarding the allocation of infants to diagnostic outcome categories required the use of the following variables, which were taken as key indicators of bacterial illness and its severity: positive blood culture; positive CSF; culture; and oxygen saturation. Positive bacteriologic outcomes were defined before the study, and definitions of contaminants were agreed at meetings during the study. Particular attention was given to Staphylococcus epidermidis which was finally regarded as a contaminant after an analysis of positive cases showed them to be indistinguishable from culture-negative cases. The definitions of contaminants were supported by the observation that at the end of the study, the mortality in the culture-negative group of infants was similar to the mortality of infants whose blood cultures were positive for organisms deemed to be contaminants. The cutoff points used for definitions of mild and severe hypoxemia were based on clinical experience and the views of experts. Most clinicians agreed that infants with SaO2 concentrations <90% are significantly hypoxemic and in need of supplemental oxygen. This is consistent with the finding that this corresponds to a PaO2 of 60 to 70 mm Hg and the point below which the hemoglobin oxygen dissociation curve falls sharply, indicating substantially less oxygen availability at tissue level with lower SaO2 levels. The dynamics of oxygen availability are different at high altitude; therefore adjustment for altitude was required because two of the sites were not at sea level (Goroka, Papua New Guinea, 1600 meters; Addis Ababa, Ethiopia, 2800 meters). Oxygen saturation measurements obtained at these sites were adjusted to correspond to readings at sea level. This was accomplished by determining the oxygen saturation at each non-sea level site at which the risk of an adverse outcome was equivalent to the risk at the sea level sites. The adjusted oxygen saturation measurements were used to place infants into three risk categories (equivalent to SaO2 values of 94%). RESULTS Patient characteristics. A total of 8418 infants younger than 3 months of age were triaged in the 4 study sites, of whom 4552 satisfied the criteria for enrollment and underwent a full history, physical examination and pulse oximetry. The characteristics of infants enrolled at the different sites are summarized in Table 5. Although the site in PNG enrolled nearly one-half of the study infants, infants seen at this site were less severely ill than at the other sites. The nutritional status of these infants was better, fewer were hospitalized and fewer died. These differences appeared to relate to the site at which infants were enrolled (a primary care clinic). Case-fatality and hospitalization rates at the three other study centers were similar.TABLE 5: Characteristics of infants enrolled at each of the four study sites Of 2398 infants who met the criteria for laboratory investigation, 2277 (95%) had blood cultures performed and 1868 (82%) had chest radiographs. Failure to perform blood cultures usually occurred because of an inability to obtain a blood specimen. Chest radiographs were not performed if an infant was clinically unstable or the radiographic facility was unavailable at the time the infant was seen. Lumbar punctures were performed on 507 infants. Of the 2154 enrolled infants who did not meet the criteria for laboratory investigation, 175 (8%) were randomly selected for laboratory evaluation, so that the total number of infants who had blood cultures performed was 2452. The proportion evaluated varied considerably between sites because of differences in the proportion of families who refused blood drawing. Outcome. Of the 4552 infants enrolled 1297 (28%) were hospitalized and 247 (5%) died. Of the blood cultures taken 162 (7%) were positive, and 49 of those infants (30%) died. Of the 41 culture-positive cerebrospinal fluid samples, 18 (44%) died. For the analysis of clinical signs 2 levels of positive outcome were defined in addition to death. The less severe of these included children with radiologic pneumonia (chest radiograph showing alveolar consolidation) or mild hypoxemia (SaO2 90 to 95%). Four hundred fifty (10%) of the children fitted into this category. The more severe level included children with bacteremia (positive blood culture), meningitis (positive CSF culture) and severe respiratory disease associated with hypoxemia (SaO2 <90%). There were 386 children in this category. Challenges in implementation of the study. Implementing a multicenter study in developing countries created a number of special logistic, methodologic and implementation challenges. Logistical challenges included the limited availability and quality of laboratory methods (e.g. bacteriology, virology, radiology), the need to transport data to the coordinating center at the WHO, specimens to reference laboratories and radiographs to and between radiologists. In addition, in developing countries challenges are accentuated by problems with infrastructure and communication and by cultural differences between groups. The study also posed several important methodologic challenges including the need to evaluate clinical predictors independently of outcome. A range of possible clinical predictors was needed to select a simple set of indicators that could be used by health workers in developing countries. Previous studies of the value of the clinical examination in identifying serious bacterial illness have obtained diagnostic information more often on infants with positive clinical findings than those with negative findings.18-20 This approach to the study of clinical examination may lead to overestimates of the accuracy of findings as false negatives may be missed and consequently sensitivity overestimated. If these patients are assumed to be "true negatives," the approach may also lead to overestimates of the specificity. This problem has been called verification bias.21 Implementation challenges included containment of data collection to the essential elements. This was difficult to achieve because of the widely divergent views among those involved in planning the study as to which symptoms or signs are likely to be useful. These problems were complicated by incomplete data management software before the beginning of the study. This led to inconsistencies in the data early in the study that had to be corrected after data were collected. Because much of the data cleaning was conducted centrally in Geneva without the forms, settling inconsistencies in the data was laborious and time-consuming. Finally funding constraints and a lack of experience in large multicenter project coordination led to staffing limitations relative to the magnitude of the project. As a consequence the project ran considerably behind schedule. DISCUSSION Multicenter studies present special challenges in design, conduct and analysis. The usual reason to conduct a multicenter study is to overcome sample size problems such that the study objectives can be achieved within a reasonable time. To do this the central components of the results need to be pooled. The validity of this exercise carries the assumption that the true results being sought are the same in each of the sites. Where the study involves the drug treatment of individuals with a specified condition, this is usually a reasonable assumption. In the present study there are two types of information being sought, clinical indicators of severe illness in infants and etiologic agents responsible for severe illness in infants. Clinical indicators of severe illness in infants younger than 3 months of age are likely to be the same in different settings, provided the means of measuring the indicators are identical. For evaluation of symptoms there is likely to be less similarity because of wide ranging cultural differences in the interpretation of symptoms by parents. Clinical signs are more objective, yet teaching and interpretation of clinical signs varies considerably. These issues were identified as a major challenge for this study and the study coordinators and principle investigators spent a great deal of time and effort trying to standardize the interpretation of clinical signs and symptoms. This task would have been much simpler if a smaller number of signs were measured and the educational level of the observers was more comparable. To the extent that these efforts were ineffective, variation in the application of the clinical findings would reflect how the findings would be applied under actual conditions. The assessment of clinical predictors is also affected by the study population. Analyses of this type are significantly affected by the level in the health care system at which the study is undertaken. For example at a first level health facility most children are well, so the few seriously ill children are being compared with a well group, which should improve the discriminating power of a physical sign. However, because the well group is numerically much larger, it will generate some false positives, and this compromises the predictive power of the sign. On the other hand at the level of a tertiary facility most children are seriously ill, including many children who do not have the outcome variable against which the sign is being evaluated. Thus the comparison group is smaller but more ill, and it becomes debatable whether the children with the sign in question, but not the outcome parameter being used, are really false positives or sick children who do not happen to have the that outcome. Despite these differences among sites, we did not observe differences in the predictive accuracy of the models among sites.17 The other primary outcome of this study was the identification of etiologic agents. In this case there is likely to be considerable variability in the spectrum of agents causing disease in these four very different settings. Although these differences would be identified only by a much larger study, it is understood that the final, pooled etiologic results of this multicenter study would constitute an estimate of the "average" etiologic spectrum, rather than the picture that would have been obtained had a study of similar overall size been conducted at any one of the four sites. Standardization of the bacteriologic and virologic methods was essential to ensure the validity of the conclusions at each of the sites, but this does not mean that the results would converge. Because the study was designed to focus on community-acquired infection, nosocomial infections were not included, so it is unlikely that the site substantially affected the bacterial agents identified unless there are true differences in the spectrum of agents responsible for community-acquired infections. These points must be kept in mind when interpreting the results of this multicenter study. The predictive ability of clinical signs is probably truly the same in different sites, although the analysis will be affected by variable interpretation of signs, and differences in study population. The etiologic analysis probably represents the average of the true situations at the different sites. These issues should be carefully thought through at the beginning of future multicenter studies of this type. At the time the study was conducted it was one of the first large scale studies of the clinical examination in developing countries. Thus the project served as a kind of beginning of this type of effort. Much has been learned about the efficient conduct of multicenter studies in recent years. However, because the knowledge and experience about how to conduct them have not been widely shared, new projects often begin the learning process from the beginning.22 Recent publications have outlined many of the systems required to support effective clinical trial execution, as well as the characteristics of the organizational structures necessary to coordinate them.23 Despite these challenges the project was important in creating the experience needed to develop the use of empirical studies to the develop policy recommendations about the use of clinical findings in managing acutely ill children. The value of this type of project is demonstrated by the results presented in this supplement.