Study design and setting

The study was based on case-control analyses within a prospective study aimed at establishing the precise cause of the fever, which was conducted from April to December 2008 in two outpatient clinics in Tanzania: one in Dar-es-Salaam, the economic capital (urban site) and one in Ifakara, Morogoro region, situated in the center of the country (rural site).

Study population, inclusion and exclusion criteria

The methodology of the study is detailed elsewhere [5]. In brief, 1005 consecutive children aged 2 months to 10 years (94% were less than five years) presenting with an axillary temperature of ≥38°C were screened. Two additional inclusion criteria were used: 1) visit was the first for the present illness; 2) fever duration ≤1 week. Exclusion criteria were: 1) chief complaint was injury or trauma; 2) receipt of antimalarials or antibiotics in the preceding week; 3) severe malnutrition according to WHO definition [25]; 4) requirement of immediate life-saving procedures. To record clinical findings, all items relevant to a febrile presentation were listed in a standardized case report form. In all cases, blood samples (5ml whenever possible), pooled nasal and throat swabs were taken for microbiologic testing on-site (rapid tests and blood cultures) and further serologic and molecular work-up. A systematic set of investigations was performed in all patients while additional investigations (including chest X-ray) were carried out according to several non-mutually exclusive decision charts developed for chief complaint(s). The most appropriate microbiologic method of detection was selected for the diagnosis of each pathogen in an acute episode.

The details of the study were discussed with the children’s guardians who provided written informed consent. The protocol was approved by the National Institute for Medical Research Review Board in Tanzania and the local ethical committee in Basel, Switzerland.

Predictors’ definitions

The parameters were divided in four groups: 1) symptoms based on history taking; 2) signs based on clinical examination; 3) epidemiological and demographic parameters (age, rural/urban site, season, chronicity, recent travel); and 4) laboratory test results. All parameters were included in the statistical analysis for each outcome, except if they were part of the criteria used to define this outcome (Table 1) or if present in less than 10 children. Laboratory parameters measured in a selected subgroup of patients were not included in the analyses (eosinophils, creatinine, C-reactive protein (CRP), procalcitonin (PCT) and other host biomarkers) and, for pneumonia, are the subject of another publication [26]. When two predictors were similar from the clinical or pathophysiological point of view [for example erythropenia and anemia (based on hemoglobin)], and thus strongly correlated, only the easiest parameter to assess or measure (in this case anemia) was retained for the multivariate analysis.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Pre-defined case definitions.

Pre-defined case definitions based on clinical and laboratory criteria, for each febrile illness included in the present analysis.

https://doi.org/10.1371/journal.pone.0173314.t001

Diseases’ definitions

The criteria and diagnostic procedures to define each disease causing fever have been described in a previous publication [5]. Briefly, the final diagnoses were established based on pre-defined criteria derived from WHO and Infectious Diseases Society of America guidelines, as well as systematic reviews. In order to maximize consistency and to minimize subjectivity, these were generated by means of a computer-based algorithm. Diseases that were documented by at least one microbiological or radiological test and affecting at least 30 patients were selected for the analyses. They included diseases documented with laboratory tests performed in all children (malaria, typhoid fever, acute HHV6), or diseases documented on one laboratory or radiological test combined with a constellation of clinical parameters (UTI, radiological pneumonia). Diseases were not mutually exclusive and some patients had more than one diagnosis. The pre-defined case definitions for each disease included in the analysis, as well as the parameters excluded from the analysis because they were part of this definition, are presented in Table 1.

An analysis was also performed for viral disease that included all clinically or microbiologically confirmed viral diagnoses, i.e. acute respiratory infection with one or more positive nasopharyngeal PCR results for a respiratory virus, gastroenteritis with a positive rapid diagnostic test result for adenovirus or rotavirus, systemic viral infections (acute HHV6, parvovirus B19, EBV, CMV and hepatitis A) documented by PCR or serology, and nasopharyngeal viral infections (defined as fever with no documented cause except a positive nasopharyngeal PCR result for a respiratory virus). ‘Bacterial diseases’ included all clinically or microbiologically confirmed bacterial diagnoses, i.e.: pneumonia with an end-point consolidation on chest X-ray (see Table 1) or severe acute respiratory infections, with a positive nasopharyngeal PCR result for Streptococcus pneumoniae; typhoid fever; UTI; sepsis due to bacteremia; gastroenteritis with a positive stool culture result for Salmonella or Shigella; rickettsioses; Q fever; leptospiroses and significant skin and mucosal purulent infections (see the supplementary appendix of the Tanzanian Fever study publication for more details [5].

Statistical analysis

Data were analyzed using STATA version 12 (College Station Texas, USA) and R software (version 2.15.3). To calculate the crude likelihood ratios (LR) for the bivariate analysis, the qualitative parameters were coded as dichotomous and the continuous parameters were dichotomized using accepted thresholds for normality. Multivariate analyses, using the Spiegelhalter-Knill-Jones (SKJ) approach [1] to obtain adjusted likelihood ratios (aLR), were performed. Only clinical parameters with a crude LR+ >2 or LR- <0.5 were introduced in the stepwise backward logistic regression models. Parameters were then excluded from the model if the p-value was ≥0.05 according to the Wald test. Confidence intervals were calculated using a bootstrap method.

To reduce the statistical hazard, parameters were also tested using a different type of statistical analysis: the classification and regression tree (CART). CART analyses have the added benefit to evaluate the combined accuracy of predictors to correctly predict the outcomes. The aim of performing this second analysis was not to provide ready-to-use clinical guidelines but to identify the best combinations of predictors. The same clinical parameters as for the SKJ approach (crude LR+ >2 or LR- <0.5) were introduced in the formula to develop the CART. Continuous predictors were both introduced as such in the CART models to let the software decide the best cutoff and were introduced as dichotomous variables with usual cutoffs. A minimum of 10 cases for parent nodes and 5 cases for child nodes, and a maximum tree depth of 4 nodes were used as conditions to generate the decision trees. The cost of misclassifying the presence versus the absence of an outcome was systematically varied to generate models favoring higher sensitivity or specificity. The pruning was varied to avoid overfitting. CART models were compared based on sensitivity, specificity and LRs. Analyses were performed using the ‘mass’, ‘boot’, ‘rpart’, ‘caret’ and ‘rpart.plot:prp’ libraries in R software.

Malaria

Most predictors of malaria obtained in our analyses (Table 3 and Fig 1A) are consistent with the findings of other studies [2,11,12,29–33]. This gives credit to the predictors we identified for diseases that have been less studied than malaria. High temperature, recent travel and jaundice were retrieved by two different types of analyses (logistic regression and CART). High temperature had the highest LR and was used as the tree’s first node in the CART model; in the latter, 38.7°C was calculated as the best cutoff for temperature to predict malaria. In a systematic review on clinical prediction rules for malaria, fever >37.7°C or >38°C were found as predictive of the disease [11]. Other studies showed an association of high temperature >40°C with severe malaria [13]. No study using CART analyses was found to be able to confirm the threshold for temperature that we found. Recent travel can be explained by the fact that malaria transmission is higher outside than in the towns of Dar es Salaam and Ifakara, and that children use less protective measures, such as impregnated bed nets, during their stay in the rural area than at home. Splenomegaly, which was the second strongest predictor according to the logistic model, does not appear in the CART model and contrariwise, hepatomegaly, which was selected by the chosen CART model, has not been retained by the logistic model. For the latter, a significant interaction with jaundice prevented it from being selected in the logistic model, which explains why these two predictors were mutually exclusive in the different generated CART models.

Typhoid fever

In the present study, 73% of typhoid fevers occurred in children older than 2 years and the variable age, with a cutoff of 2 years, was in fact used for the first node in the CART model (Fig 1B). Other studies have also shown this association with age [14,16,17]. Age is an easy parameter to assess for clinicians and could be used in guidelines to decide on performing a laboratory test or giving presumptive treatment for typhoid fever. Jaundice, abdominal tenderness and lymphadenopathy were predictive of typhoid in both types of analyses. Jaundice was present in 22% of the typhoid cases, which is a higher prevalence compared to other studies (10 to 16%) [34,35]. However, this sign is not described in most studies as a clinical predictor of typhoid fever in children [14–17,36–39]. This could be explained by the difficulty for many clinicians to assess this sign combined with its relatively low frequency in children in general. Typhoid fever should thus be included in the differential diagnosis of jaundice, besides malaria and hepatitis. Abdominal tenderness was the strongest predictor resulting from the logistic regression and, combined with age in the CART model, these two predictors identified only 17 false positive cases, providing a high specificity of 98%. Abdominal signs in general are well described in the literature as predictors of typhoid fever, and liver pain, which corresponds to an abdominal tenderness in the upper right quadrant, was also a good predictor in our study. Although very specific, this latter sign was however still quite rare among typhoid fever cases (8%), which would preclude its use in clinical guidelines. Rainy season was significantly associated with typhoid fever: 80% of the cases occurred during that season. We did not find any similar result in other studies. The reason might be that during heavy rains, the sewers tend to overflow and contaminate the pipes of clean water with fecal bacteria, including Salmonella enterica serova Typhi. Interestingly, 14% of the children with typhoid presented lymphadenopathies. This sign is not clearly described in the literature and further studies would be necessary to really evaluate its clinical predictive value. Prolonged fever, commonly described as a predictor of typhoid fever and one of the reasons for Integrated Management of Childood Ilnesses (IMCI) to recommend referral of children whenever fever lasts more than one week, was significantly associated in the bivariate analysis with a crude LR+ of 2 for fever duration of ≥4 days. In the present study, which included only fevers lasting less than one week, a significant number of typhoid cases were still identified, among which more than a quarter were severe cases based on objective criteria. This means that diagnosing typhoid regardless of the fever duration, either based on clinical predictors such as those found in the present study, or by a reliable test when available, is essential.

Urinary tract infection (UTI)

In the present study, 95% of children with UTI were under the age of 3 years (the best age cutoff in the bivariate analysis, the CART models (Fig 1C) providing a cutoff of 2 years instead) and 100% under 5 years, which confirms the strong relationship between age and UTI as described in other studies [18,40]. Temperature ≥40°C and fever duration of ≥3 days were also associated with UTI as described in two studies [18,41]. A known symptom of UTI, usually found in school-aged children and adults, pollakiuria, was strongly associated with UTI as well. When present, specific symptoms of UTI are thus of great help to suspect the diagnosis in older children. Considering their good sensitivity, accessibility and frequency, one or more of these parameters could be integrated in guidelines, for example to improve the pre-test probability before performing a urine dipstick. Although the dipstick test has the advantage over urine culture of not being subject to bacterial contamination, it still lacks specificity. To avoid overdiagnosis, this test should thus not be used in all febrile children regardless of the clinical presentation, but rather in a sub-group of children selected on the basis of these predictors. White mouth (oral candidiasis) was associated with UTI as well; both UTI and oral candidiasis most often affect young children and therefore, this predictor probably simply represents a confounding factor.

Radiological pneumonia

Recent studies conducted in US emergency departments have shown that WHO criteria used to diagnose radiological pneumonia lack sensitivity and specificity [42] and that tachypnea did not discriminate between children with positive and negative chest X-ray [43]. In a systematic review and meta-analysis, fast breathing and lower chest wall indrawing showed poor diagnostic performance [44]. Thus, for settings where radiology is not available, looking for alternative predictors could be useful to improve diagnostic rules. The performance of the widely studied sign tachypnea [19,20,43,45–48], also named fast breathing (WHO terminology), could not be measured in the present study as it was the entry point for deciding to perform a chest X-ray and thus to diagnose radiological pneumonia. However, very fast breathing and other respiratory signs or symptoms, such as dyspnea, chest indrawing, nose flapping and abnormal chest auscultation can still be studied and the bivariate analysis showed that respiratory signs or symptoms presented the highest +LRs. In bivariate and multivariate analyses, the best predictors were respiratory signs, abdominal tenderness, high temperature and leukocytosis. The strongest predictor, abdominal tenderness, which probably reflects the presence of a serious infection in general, should be easy to assess by clinicians and could thus be potentially useful in an algorithm aimed at identifying pneumonia. The predictors abnormal chest auscultation or leukocytosis are clinician dependent for the first and not always available for the second, which makes them unsuited for guidelines. On the contrary, the predictors very fast breathing and temperature ≥39°C are easy to assess. In the CART model (Fig 1D), respiratory signs and leukocytosis were selected but the final sensitivity (0.20) was very low and despite a high specificity (0.97), the performance of the CART is not sufficient. As authors concluded in a recent systematic review on clinical features for diagnosis of pneumonia in outpatients [44], our findings highlight that new point-of-care tests to measure host biomarkers or specific pathogens need to be further investigated to be able to provide better diagnostic tools to clinicians.

Acute HHV6 infection

With 79 cases, acute HHV6 infection is the second most frequent diagnosis after malaria (105 cases). Only 1% of the children with HHV6 presented a rash, which is described as the most typical feature of this infection (roseola). A US study conducted in 1994 [49] found a prevalence of only 6% of roseola’s typical exanthema at the beginning of the episode and a total of 17% when including exanthema appearing when fever subsided. An explanation could be that the faint rash of acute HHV6 infection is difficult to detect on black skin and generally appears after the resolution of the fever. Both the logistic regression (Table 3) and the CART models (Fig 1E) found that age under 2 years was useful to predict HHV6 infection. Leucopenia was also predictive of HHV6 in accordance with the literature [50]. Interestingly, the CART model shows that it is the absence rather than the presence of each symptom or sign that rules in HHV6. We therefore looked at the association between the variable absence of clinical symptom or sign and HHV6 and found a crude LR+ of 1.8 (1.1–2.8). This finding supports the recommendation provided by the WHO and other guidelines not to treat febrile children with antibiotics in the absence of specific symptoms or signs and to rather follow-up with the patient, as most of these cases correspond to self-limited viral infections.

Bacterial disease

The predictors found for (any type of) bacterial disease, temperature ≥40°C and seizures, were described in the systematic review of Thompson et al. [51]. We also found chest indrawing and nose flapping, predictors which were not described in other studies. These signs reflect the presence of a respiratory distress that can be due to severe sepsis in general, and not only to a respiratory infection. Severe anemia, that is known to lead to a high mortality rate, is well known in malaria but not specifically as a predictor of bacterial infection. Hemoglobinometer could thus potentially be used as a triage tool at primary care level to detect children at risk of suffering from a serious infection, and thus allow to decide to refer them at a higher level of care with an antimalarial and an antibiotic as pre-referral treatment. The decision tree of the CART model (Fig 2A) used chest indrawing, chronic condition, temperature >40°C and fever duration of >3 days to predict bacterial disease, resulting in an overall good specificity of 0.83, but unfortunately a rather poor sensitivity of 0.45. The aLR+ of the predictors in the logistic regression model were indeed quite high, which allows to rule in the diagnosis, but their aLR- being included between 0.85 and 0.97, their absence does not allow to exclude a bacterial infection. Abdominal tenderness, which was a good predictor of the three infections at highest risk of bad outcome (radiological pneumonia, typhoid and malaria), did not come out for bacterial disease. This is because this category did not include malaria and included also milder bacterial infections, such as UTI or skin infections. Abdominal tenderness thus seems to be a predictor of a serious infection and could potentially be used at primary care level to either refer the child and/or treat him/her with an antibiotic.

Viral disease

Regarding viral disease, the most interesting predictors were negative predictors. These corresponded to the absence of severity criteria, such as seizure, temperature > 40°C or jaundice, which illustrates that the clinical presentation of viral diseases is more often mild than that of bacterial diseases. However, chest indrawing, a criterion used to classify pneumonia in WHO guidelines, was in fact associated with viral disease in the multivariate regression. This is due to the fact that most acute respiratory infections, especially in young children, are due to viruses rather than bacteria, even in children with a more severe presentation. In the northern countries, bronchiolitis is nowadays one of the main causes of hospital admission in children. With the epidemiological transition due to many factors, including vaccines against pneumococcus and Haemophilus, the southern countries are also progressively shifting from bacterial to more viral diseases. To avoid overtreatment of febrile children with antibiotics, a public health priority to mitigate the spread of microbial resistance, it is essential to also provide clinicians with prediction rules able to identify children with viral infections.