A longitudinal observational study was conducted at two hospitals in Thailand: (1) the Queen Sirikit National Institute of Child Health (QSNICH) in Bangkok during 1994–97, 1999–2002, and 2004–07, and (2) the Kamphaeng Phet Provincial Hospital (KPPH) in the Kamphaeng Phet providence in a rural northern section of Thailand during 1994–97. The study methods have been described in detail elsewhere [15]. In brief, children between the ages of six months and 15 years presenting with temperature ≥38°C for no more than 72 hours and no localizing symptoms were identified in the outpatient department or on the hospital ward were eligible for enrollment with parental consent. Exclusion criteria included: signs of shock at presentation, chronic illness, or an initial alternate non-dengue diagnosis. All subjects were admitted to the hospital and monitored until 24 hours after defervescence. Fluid management (oral rehydration or intravenous fluid) was guided by the ward physician based upon clinical necessity. The study protocol was approved by the Institutional Review Boards of all participating institutions.

A blood sample was obtained on the day of enrollment and daily thereafter until discharge or for a maximum of five consecutive blood collections. Serological assays (IgM/IgG ELISA and hemagglutination inhibition assay), virus isolation, and/or RT-PCR were used to confirm all dengue cases. Patients were observed and daily clinical and laboratory measurements were recorded using standardized data collection forms.

After defervescence (2 consecutive temperatures below 38°C), serial finger-stick hematocrits were measured to capture hemoconcentration. A right lateral decubitus chest x-ray was taken the day following defervescence and a pleural effusion index (PEI) was measured as 100×(maximum width of right pleural effusion)/(maximum width of right hemithorax). After completion of the case record, a single expert physician (author S.N.), who was not directly involved in patient care, assigned a final diagnosis of DF, DHF, or OFI based upon chart review following WHO guidelines [3].

Given that not all DHF cases are severe and not all DF cases are mild, we applied several different categories of dengue disease severity using data from each patient's entire hospital course : (1) dengue shock syndrome (DSS, as defined by WHO criteria); (2) DSS or PEI>15; (3) DSS or required intravenous fluid; (4) DSS or platelet count < = 50,000 anytime during illness; (5) DSS or received fluid intervention (oral or intravenous) in any 24-hour period that exceeded maintenance volume +5% volume deficit [16], [17].

The input variables used for establishing each tree were platelet count, hematocrit, WBC count, percent monocytes, percent lymphocytes, percent neutrophils, AST, ALT, tourniquet test (+/−), age, and gender, all of which were obtained on the day of presentation.

Descriptive characteristics of the study sample were compared using t-tests and Pearson's χ². CART analysis was performed using SPSS Answer Tree 3.0 software (see Text S1) [18]. CART analysis is a non-parametric analytic tool that has many advantages over logistic regression models, such as its ability to detect interaction between variables used to generate a tree [19], [20]. Age, gender, and clinical laboratory data on the day of presentation were used to establish diagnostic decision trees to distinguish between patients with severe dengue illness and those with non-severe illness or OFI. Stopping rules were: (1) no terminal node could contain <5% of the original sample size, (2) no more than 5 levels per tree, and (3) a minimum improvement in impurity of .0001.

Additional analyses were performed to examine differences in diagnostic trees according to the day of presentation among the low risk, non-severe group. The final trees selected were those that had minimum misclassification of severe dengue illness in low risk nodes (high sensitivity) and maximum correct classification of non-severe dengue and OFI in low risk nodes (high specificity). In each terminal node, patients were classified as low risk or elevated risk of severe dengue illness where optimal sensitivity could be achieved. For all analyses, sensitivity was weighted more heavily than specificity by using misclassification cost ratio of 1∶10 severe dengue vs. non-severe. Each tree was validated using the k-fold cross validation method [21], [22]. We used k = 5 in our analysis.

In total, 1384 patients were enrolled in the study. Of these, 1311 had a final diagnosis of DHF, DF, or OFI. Of the remaining 73 patients, 32 had an undetermined diagnosis due to lack of convalescent blood sample for serology, and 41 had a presumed non-viral infection. An additional 81 patients (28 DHF, 20 DF, and 33 OFI) were missing one or more variables of interest on the day of presentation and were excluded from the analysis. Table 1 describes the 1230 patients included in the analysis. Among these, 208 had a final physician diagnosis of DHF (53 grade 1, 118 grade 2, 36 grade 3, 1 grade 4), 374 had DF, and 648 had OFI. Secondary infections accounted for 81.9% of all dengue infections (74.6% of DF cases and 95.2% of DHF cases). The majority of dengue infections were DENV1 (40.7%). Table 2 indicates the number of patients with severe dengue based on different definitions as well as how trees produced with each of the definitions (Outcome variable for tree) performed when severity was defined differently (Outcome variable for evaluation of tree).

Trees were generated for each of the five categories of severe dengue illness. As summarized in Table 2 and shown in Figure 1 (Tree 1), the tree that provided the best distinction on the day of presentation categorized severe dengue as DSS. The initial splitting variable in the tree is WBC count; other variables in the tree include percent monocytes, platelet count, and hematocrit. The tree resulted in five terminal nodes, of which three are considered low risk and two are considered elevated risk. The three low risk nodes are 1) WBC>8500, 2) WBC< = 8500 and percent monocytes >9.0, and 3) WBC< = 8500, percent monocytes< = 9.0, platelet count >160200, and hematocrit>40%. The two nodes considered elevated risk of severe dengue were 1) WBC< = 8500, percent monocytes < = 9.0, and platelet count < = 160200 (64.9% of patients with severe dengue) and 2) WBC< = 8500, percent monocytes< = 9.0, platelet count >160200, and Hct< = 40 (32.4% of patients with severe dengue).

A total of 576 (48.3%) patients with non-severe dengue are classified correctly in the low risk group at the cost of misclassifying one patient who later manifested DHF grade 3. The initial splitting variable correctly classified 384 (32%) of the patients with non-severe dengue. The patients that were correctly classified as low risk included 63.7% of all OFI, 32.1% of all DF, 41.5% of all DHF grade 1, and 17.8% of all DHF grade 2. Patients with non-severe dengue illness were more likely than patients with OFI to be classified as elevated risk of severe dengue (70.1% of non-severe dengue versus 36.3% of OFI).

Among the 617 (51.7%) patients with non-severe illness that were classified as elevated risk, the median day of presentation was 72 hours after illness onset and the average length of hospital stay was 6.8 days; patients with non-severe dengue that were correctly classified had a median day of presentation of 48 hours after illness onset and an average length of hospital stay of 7.3 days. To assess differences according to the day of presentation, the tree was applied using data from patients with non-severe illness at 72 hours among patients who were still febrile. In this group of low risk patients, the percent correctly classified as low risk decreased slightly from 48% to 44% (data not shown).

Figure 2 shows a diagnostic decision tree in which severe disease was defined as DHF grade 3 or 4 or PEI>15 (Tree 2). This disease categorization added nine patients with DHF grade 1 and 37 patients with DHF grade 2. No patients diagnosed with OFI or DF had a PEI>15. For this tree, the initial splitting variable was WBC count; other variables in the tree include AST, percent neutrophils, platelet count, and age. There are eight terminal nodes, of which five are considered low risk and three are considered elevated risk. The five low risk nodes are 1) WBC>13700, 2) WBC< = 13700, AST 36–50, and platelet count >282000, 3) WBC< = 13700, AST 36–50, platelet count < = 282000, and age< = 6.75, 4) WBC< = 13700, AST< = 35, and percent neutrophils< = 68%, and 5) WBC< = 13700, AST< = 35, percent neutrophils>68%, and platelet count>291000. The three elevated risk nodes are 1) WBC< = 13700, AST>50 (72.3% of patients with severe dengue), 2) WBC< = 13700, AST 36–50, platelet count < = 282000, and age>6.75 (16.9% of patients with severe dengue), and 3) WBC< = 13700, AST< = 35, percent neutrophils>68%, and platelet count< = 291000 (9.6% of patients with severe dengue).

This tree correctly classified 505 (44%) patients with non-severe dengue at the cost of misclassifying one patient with severe dengue. The misclassified patient was diagnosed with DHF grade 2 and had PEI of 25.8. All patients with DHF grade 3 or grade 4 were correctly classified in this tree as elevated risk of severe dengue. Among the 505 patients correctly classified as low risk of severe dengue, 380 were OFI (58.6% of OFI), 105 were DF (28.1% of DF), and 20 were DHF grade 1 or 2 (16.0% of non-severe DHF). Patients with non-severe dengue illness were more likely than patients with OFI to be classified as elevated risk. When the tree was applied using data from patients with non-severe illness at 72 hours, the percent of non-severe cases correctly classified as low risk increased from 44% to 50%.

We assessed the generalizability of our trees using other categories of dengue disease severity (Table 2). For example, when applying the tree that was generated using DSS as the only criterion for dengue disease severity (Tree 1) to different categories of severity, the percentage of patients with a severe dengue illness that were misclassified as low risk ranged from 12.5% to 17.6% and the percentage of patients with non-severe illness that were correctly classified ranged from 48.6% to 52.1%. All additional trees (Trees 3–5) had moderate specificity but limited sensitivity (Table 2), with a misclassification of severe dengue as low risk ranging from 34.9% to 42.6% and a correct classification of non-severe illness ranging from 72.1% to 81.5%. Each tree shared the same initial splitting variable of WBC count (data not shown).