Early diagnosis of severe dengue illness not only has the potential to reduce morbidity and mortality, but could also reduce the economic impact of dengue illness by decreasing the duration of hospitalization and the number of patients who will develop shock. We identified two diagnostic algorithms using early clinical laboratory indicators and patient characteristics that could distinguish patients with severe dengue from those with non-severe dengue or other febrile illnesses within the first 72 hours of illness.
When applying these trees to other (broader) categories of disease severity, a high sensitivity was still achieved. Previous studies have shown that modified definitions of dengue disease severity have better agreement with a treating physician's assessment when compared to strict adherence to WHO criteria 
. For any classification of dengue disease severity utilized, a high proportion of patients with non-severe dengue or other febrile illness were correctly classified as low risk of severe dengue (). These data suggest that patients classified as ‘elevated risk’ of severe dengue based on these algorithms should be treated and managed more aggressively; in comparison, our data suggest that patients classified as ‘low risk’ of severe dengue could be safely managed on an outpatient basis.
The single patient with severe dengue that was misclassified in Tree 1 presented within the first 24 hours of illness, had an initial WBC count of 13700, and was diagnosed with DHF grade 3. Five other patients with severe dengue in Tree 1 also presented within the first 24 hours and yet were correctly classified as elevated risk. When we further investigated the effect of day of presentation by using day 3 data from all non-severe cases, we found that day of presentation had little effect on the sensitivity of Trees 1 and 2 (within the first 72 hours); Tree 1 still correctly classified 44% of the non-severe cases as low risk of severe dengue infection and, in Tree 2, the percent correctly classified as low risk increased from 44% to 50%.
Many of the variables used in our decision algorithms have been shown to distinguish between patients with dengue and patients with OFI in other settings 
. Trees 1 and 2 have an initial splitting variable of WBC count, which reinforces the reported utility of this variable in distinguishing severe dengue illness within the first days of illness 
. Both trees included nodes using platelet count as the splitting variable. Thrombocytopenia is a hallmark of severe dengue disease, although it frequently occurs in DF as well 
. Platelet counts are able to distinguish between patients with dengue and OFI 
. However, when producing a tree using a minimum platelet count of <
50,000 as part of the categorization of severity (Severity category 4), the tree misclassified 42.5% of patients with severe dengue (). These data suggest that thrombocytopenia is not a specific marker for severe disease in the early febrile phase of dengue illness.
One criticism of CART analysis is that the cutoff values may not be clinically meaningful. However, when we re-defined the cutoff values for Trees 1 and 2 the results maintained a high sensitivity. For example, in Tree 1 when we rounded platelet count to 160,000, the results remained the same. In Tree 2, when we rounded the cutoffs of platelet count to 290,000 and 280,000, percent neutrophils to 70%, and age to 7, the tree correctly classified 45.9% of the non-severe cases while still achieving 94.0% sensitivity for severe cases.
Interestingly, many of the cutoff lab values in our decision trees fall within the ‘normal’ range; this suggests that established ‘normal’ ranges for routine laboratory tests have low sensitivity to detect clinically relevant changes. For example, some variables, such as hematocrit are normal early in the course of disease but appear to be able to predict those children who will later develop severe disease, often associated with hemoconcentration, as seen in the final node of Tree 1. We interpret this to indicate an interaction with the outcome of severe dengue and the other clinical values used in this Tree.
Tanner and colleagues published an analysis establishing dengue decision trees; however, their analysis was based on only three WHO-defined DHF cases and it was unclear if these three cases met other objective criteria for severity 
. In contrast, our study has 37 cases of more severe WHO-defined DSS and 171 cases of DHF grade 1 or 2. We also applied other criteria that could classify patients with dengue as having severe illness. Their study included a platelet count of <50,000 as part of the definition of severe dengue, and the resulting tree was limited in its sensitivity (82.6%) 
. Although the tree had a high specificity, sensitivity is a more important clinical consideration in the detection of severe disease. A more recent decision tree study by Lee and colleagues found a history of clinical bleeding, serum urea and serum protein to distinguish between patients with DF and patients with DHF; however, both studies have limited clinical utility as a predictive algorithm for patients with severe dengue because virologic confirmation of dengue infection is not known at presentation 
. Our study identifies those with severe dengue illness among all suspected dengue cases.
Our study is subject to some limitations. First, our study included only pediatric patients at two hospitals in Thailand. However, because the majority of dengue cases in Thailand and other regions of Southeast Asia are children, our findings are clinically relevant 
. Additionally, some patients may have received early fluid intervention as part of standard care which may have modified disease progression and development of severe plasma leakage. Furthermore, because our study enrolled patients only during the initial 72 hours of illness, these algorithms may not adequately reflect clinical practice outside of a research setting where many patients present for medical attention after the first 72 hours of illness. Therefore, we cannot make any conclusions regarding the sensitivity and specificity of these classification trees at later time points in illness. We recognize that clinical algorithms cannot replaced by good clinical management. Further validation using datasets from additional prospective cohort studies conducted in other dengue endemic regions is needed to establish the clinical utility of our algorithms in other populations.
We provide two decision tree algorithms using 12 years of systematically collected clinical data from a well-defined cohort of pediatric patients in a dengue-endemic region. Our algorithms have minimal misclassification of WHO-defined DSS cases among all patients with suspected dengue infection who present within the first 72 hours of illness. These algorithms also have minimal misclassification of other severe dengue illnesses using different categorizations of severity. A robust, validated decision algorithm can be easily implemented in resource limited settings to identify patients who are at risk for developing a more severe dengue illness and limit the number of unneeded hospitalizations.