Predicting the severity of dengue fever in children on admission based on clinical features and laboratory indicators: application of classification tree analysis

Dengue fever causes a high burden of disease and mortality across tropical and subtropical regions in Southeast Asia, Africa, the Western Pacific, and the Americas [1]. Dengue virus comprises five serotypes, DENV-1, DENV-2, DENV-3, DENV-4 and DENV-5, which are transmitted by Aedes aegypti mosquitoes [2‐4]. An estimated 2.5 billion people worldwide are at risk of dengue. More than 50 million dengue infections are estimated to occur annually, of which approximately 500,000 result in hospital admissions for severe dengue in the form of dengue haemorrhagic fever (DHF) or dengue shock syndrome (DSS), principally among children [5].

Dengue infection is frequently confounded with other febrile illnesses (OFI), presenting with non-specific clinical symptoms and clinical features analogous to OFI. During the early stages of dengue, the presence of non-specific febrile illness makes precise diagnosis strikingly difficult, resulting in inefficient treatment and possible increases in morbidity and mortality [2, 6]. Severe dengue fever, if not appropriately managed, may lead to rapid death, particularly in children [7, 8]. In addition, the lack of necessary laboratory facilities, particularly in remote, rural areas, may cause difficultly in discriminating dengue infection from OFI [9]. Dengue is one of the most common vector-borne diseases in Southeast Asia, and one of the most important mosquito-borne viral diseases with an epidemic potential in the world [10].

Dengue was first included in Cambodia’s national surveillance programme in 1980. Since 2000, between 10,000 and 40,000 dengue cases have been reported annually by the Dengue National Control Program [11], from a total population of approximately 13.5 million people [12]. The true incidence of the burden of disease in Cambodia remains under-reported due to difficulties in diagnosing dengue infection, especially in hospitals [13]. In this study, data from a cohort of children that were admitted with febrile illness to Angkor Hospital for Children, Siem Reap, Cambodia, during a one-year period were retrospectively analysed using a data mining approach. This approach used a classification and regression tree, or CART, which was first introduced by Breiman et al. [14]. This is a common tool used in data mining, which creates a model or algorithm that predicts the value of a target variable based on several input variables. In our study, CART was constructed to predict the severity of dengue infection based on early clinical and laboratory indicators. The model was then evaluated against the final diagnoses.

There were 3225 patient admissions during the study year, of which 1361 (42.2%) met the inclusion criteria. Of these, 136 (10.0%) were not enrolled, leaving 1225 febrile episodes in 1180 children, with 1144 children having a single episode, 31 children having two episodes, one child having three episodes, and four children having four episodes. The patients were mainly diagnosed as having lower respiratory tract infection (38.3%), undifferentiated fever (25.5%), or diarrhoeal disease (19.5%) [15]. Out of 1180 enrolled children, there were 69 deaths, the causes of which were: clinical pneumonia with no organism/virus identified (12 cases, 27.5%), dengue virus infection (eleven cases, including one with co-existent melioidosis, two with co-existent scrub typhus, and four with co-existent clinical pneumonia, 15.9%), and melioidosis (four cases, 5.8%). 941 non-dengue episodes and 86 episodes with no samples available were excluded from this analysis. Further details can be found in the original report [15].

Out of 198 confirmed dengue episodes, 43 episodes required ICU admission, with 29 of those classified as severe dengue based on their clinical signs, supported by two independent clinical opinions. Nine additional severe dengue episodes were included from non-ICU admissions, making a total of 38 episodes of severe dengue. There were eleven in-hospital deaths amongst all ICU-admitted patients with dengue virus infection, however dengue was the primary diagnosis in just five of these. Therefore, only these five cases were included in the severe dengue group. The flowchart of the study is shown in Fig. 1.

Clinical features, including blood in the stool, liver enlargement, ICU admission, number of days in ICU, low or high haematocrit, low or high WBC count, high creatinine, high urea, low platelet count, rapid pulse, rapid respiratory rate, low Glasgow Coma Score (GCS), pleural effusion (only one case), abdominal pain, urinary protein, urinary RBC, and high ALT, were considered on a case-by-case basis when clinicians classified dengue as severe or non-severe. The clinical features and laboratory indicators of the 38 severe dengue cases are shown in Table 1. The three most common features among patients with severe disease were ICU admission (76.3%), rapid respiratory rate (81.5%), and rapid pulse (65.7%). Severe dengue was more prevalent in children aged less than five years old. Vomiting and abdominal pain were significantly more common in the severe dengue group, as were rapid pulse and respiratory rate, increased capillary refill time, and low GCS. A significantly higher proportion of patients with severe dengue presented with a lower haematocrit, higher WBC and lymphocyte count, higher ALT level, together with the presence of urinary RBC (Table 2).

The final decision tree algorithm included five clinical and laboratory parameters: haematocrit, GCS, urinary protein, creatinine, and platelet count. The sensitivity and specificity of the model were 60.5% and 65%, respectively (Fig. 2). The accuracy of the model was 64.1%, where the clinical diagnosis was used as the reference value. The area under the receiver operating characteristic (ROC) curve for logistic regression was 0.616. The final decision tree was then restructured using logistic regression analysis to estimate the impact of each CART-selected variable as represented by the OR and 95% CI.

Table 3 gives the estimated OR for each parameter selected by CART. Low haematocrit, low GCS, low platelet count, presence of urine protein, and high creatinine increased the probability of a diagnosis of severe dengue, with significant OR ranging from 1.47 to 13.73. The parameters that were statistically associated with severe dengue were 1) low haematocrit (OR = 7.114, 95% CI = 3.00–16.87, p < 0.001) and 2) low GCS (OR = 13.73, 95% CI = 3.46–54.50, p < 0.001). Although low platelet count (OR = 2.33, 95% CI = 0.95–5.76), presence of urine protein (OR = 1.83, 95% CI = 0.78–4.32) and increased serum creatinine (OR = 1.47, 95% CI = 0.51–4.25) were associated with an increased risk of severity, they were not shown to be statistically significant by regression analysis (Table 3).

Using a data mining approach, we have developed an algorithm using both simple clinical manifestations and laboratory indicators to predict the severity of dengue during the early phase of the illness. The final algorithm for predicting severe dengue (Fig. 2) comprised six components in order of their significance. The most significant factor in predicting severe dengue was low haematocrit, followed by a GCS of 11 or below as the second split if haematocrit was greater than 28, the presence of urine protein and creatinine above 84 μmol/l as the third split if GCS was above 11, and finally a platelet count of 146,000 per mm³ or less as the final split, if the presence of urine protein and creatinine was below 84 μmol/l.

Comparing the algorithm we derived with those reported in previous studies, we found both similarities and differences. Potts et al. constructed decision algorithms for predicting dengue shock syndrome (DSS) or dengue with significant pleural effusion [26]. The algorithm achieved a high sensitivity of 97%. Both low haematocrit and platelet counts were also identified as predictive factors in their work, although the cut-off values used in our algorithm were more extreme, i.e. for haematocrit ≤ 28 vs. ≤ 40, and for platelet count ≤ 146,000 vs. ≤ 160,200. The mechanism by which thrombocytopaenia is caused by dengue virus is complex [27]. Previous studies suggest that the virus probably contributes to bone marrow suppression and platelet destruction [28, 29]. To meet the WHO guidelines for classifying patients with DHF, thrombocytopaenia (platelet count ≤ 100,000) is required. Srikiatkhachorn et al. demonstrated that thrombocytopaenia was related to dengue severity and that not all severe cases would have been classified as DHF according to the WHO criteria [30]. Although thrombocytopaenia suggests that dengue infection is severe, a low platelet count is also common among OFI such as malaria and scrub typhus [31]. The 1997 WHO definition of DHF stated that a low platelet count (≤ 100,000), together with an increased haematocrit of ≥ 20% above the baseline value, is indicative of plasma leakage. In contrast, our results and those of Potts et al. suggested a drop in haematocrit as a sign of severity, especially among patients with internal bleeding in areas such as the gastrointestinal tract [26]. Our results also suggested a more extreme haematocrit value compared with the previous study (28% vs 40%) [27]. Although Potts et al. identified WBC count and monocyte percentage as important, our analysis did not identify monocyte results as significant even when included in the decision tree algorithm. In addition, Potts et al. evaluated predictors for DSS and dengue with significant pleural effusion, whereas in our study severe dengue was differentiated based on clinical features and laboratory indicators.

Another recent study, by Tamibmaniam et al., used simple logistic regression and identified three parameters, including vomiting, pleural effusion, and low systolic blood pressure, to predict severe dengue based on the 2009 WHO criteria [32]. This study did not specifically focus on children and included only female patients. The sensitivity and specificity achieved in its decision algorithm were 81% and 54%, respectively. Of the three parameters they identified, vomiting was the only parameter available in our study, and although it initially appeared to be significant in the severe group, it was not selected for the final tree.

Despite using a similar approach to predict somewhat similar outcomes to the aforementioned studies, we identified additional parameters that related to the severity of dengue, including GCS, urine protein, and serum creatinine. There are a number of possible explanations for these differences, as outlined below.

GCS is used to measure the level of consciousness (mental status changes) [33]. In our results, the node with GCS ≤ 11 (considered to be moderate) in the model was significant. Rao et al. showed that patients with dengue encephalitis had a GCS of 7–8 and recommended intubation and mechanical ventilator support during their hospitalisation [34].

Previous studies in which urine protein was associated with DHF or DSS used the urine protein-to-creatinine ratio [35, 36], but we used only a urine dipstick for this measure. The presence of urine protein in severe dengue could be due to plasma leakage.

An increased serum creatinine level indicates kidney dysfunction. In patients with DHF, a mild increase in serum creatinine is common, in contrast to the higher levels seen in severe dengue cases. Our model showed that a serum creatinine level > 84 mmol/l (4.6 mg/dl) was associated with severe dengue, a value similar to that found in Thai paediatric patients with DHF, whose mean serum creatinine was 4.9 mg/dl. That analysis also showed that 24 of 25 patients with acute kidney injury (AKI) had DSS as a final diagnosis. Of the 25 patients with DHF-associated AKI, 16 (64%) died as a result of profound shock, together with other conditions such as liver failure, respiratory failure, and severe bleeding [37]. Studies in adults have reported an AKI incidence of 14.2% among dengue patients, and those with AKI saw significant morbidity and mortality, longer hospital stays, and poor renal outcomes [38]. Early diagnosis of dengue infection, known clinical and laboratory characteristics and risk factors together with early detection of AKI using appropriate criteria [39], and monitoring for warning signs of severe dengue, are essential if AKI and other complications are to be avoided [40].

Although the two sets of WHO criteria from 1997 and 2009 are still debatable in terms of their ability to appropriately differentiate dengue from OFI and to classify disease severity [20‐22, 41, 42], the problem is compounded by a lack of key data in resource-poor settings, making it difficult to apply the criteria. For instance, we lacked information on clinical bleeding sites, and were only able to detect gastrointestinal bleeding based on stool examinations. In addition, there was a lack of information on blood pressure or narrow pulse pressure to indicate whether a patient was in shock [43], no data on restlessness suggestive of circulatory failure, and no chest X-ray results to evaluate pleural effusion or abdominal ultrasound to detect ascites, both of which are important for identifying plasma leakage. The 1997 and 2009 WHO dengue guidelines also include a tourniquet test as a diagnostic tool for dengue in the early febrile phrase. However, the tourniquet test has been shown to have low sensitivity for dengue diagnosis, such that a negative result does not exclude dengue infection [44‐46]. The tourniquet test had not been performed for the majority of patients in our data set and was thus not included in the analysis.

Regarding the two approaches used in this study, CART versus the more conventional approach of logistic regression, some points are worth mentioning here. Firstly, our main focus was on building the decision tree model from CART analysis. CART is non-parametric, and can manipulate numerical data which may be highly skewed, multi-modal, ordinal or non-ordinal in structure. CART is not significantly impacted by outliers in the input variables. The output of CART in the form of a decision tree is easy to follow and gives some visual information on the hierarchical importance of the variables from the top to the bottom of the tree, although calculating the importance matrices of the predictors in CART is not straightforward. In this study, therefore, we quantified the importance of each decision tree predictor via the odds ratio as calculated by logistic regression. Secondly, the ways in which the decision boundaries are generated in the two approaches are different. While the logistic regression generates a single boundary, a decision tree essentially partitions the data space into half-spaces using axis-aligned linear decision boundaries, giving a non-linear decision boundary. Either approach may be more applicable depending on the setting. Finally, the model’s accuracy was measured in different ways for each of the two approaches. For the decision tree model, the out-of-sample accuracy was estimated via cross-validation, i.e. the 10-fold cross-validation function in Weka allowed us to conveniently perform the cross-validation and directly report the model’s accuracy. For the logistic regression, however, the model’s accuracy was estimated from the classification table, which showed the number of observed against predicted outcomes, using a default cut-off value of 0.5 for the predicted probability. For all of the above reasons, it would have been difficult to compare the relative merits of the two methods used in our study.

There were several limitations with regard to the dataset used for this study. Firstly, the data came from just one hospital, a further indicator of the poor resources in the Southeast Asia region where dengue is endemic. Secondly, due to the lack of IgG antibody results it was not possible to interpret whether cases were primary or secondary dengue infections. This information could potentially be a useful early indicator for the severity of a dengue infection. Thirdly, the algorithm was derived from data collected within 48 h of admission among children aged less than 16 years old. If the model were to be used for older patients or in different regions, some adjustments to it may be necessary.

Although the cohort of 198 patients with confirmed dengue was relatively small, with an even smaller subset of just 38 severe dengue cases, the simple model we derived is still likely to be useful because it includes a small number of predictive variables that would probably be available in similar settings. In addition, a previous study by Carter et al. showed that the DENV rapid diagnostic test (RDT) had low sensitivity for the diagnosis of dengue infection [47]. However, the development of diagnostic tests for dengue has advanced rapidly. The NS1 test in particular has become widely available in many resource-limited settings. It is simple to use and has acceptable accuracy. If rapid diagnosis of dengue using NS1 can be achieved, our algorithm would prove very useful. This also highlights the importance of children attending for testing as soon as dengue is suspected, because NS1 detection is optimal during the first seven days of infection. The algorithm will become more relevant and useful as the rapid diagnosis of dengue becomes more common. By using our simple algorithm to help identify and predict severe dengue, we believe that there would be more room to focus on other, more serious bacterial diseases, which are all-too-common in these types of resource-poor settings.

Our decision tree algorithm using simple clinical and laboratory indicators has a moderate classification accuracy for predicting the development of severe dengue fever among paediatric patients with confirmed DENV infection. The model demonstrates the importance of haematocrit and platelet levels for monitoring the severity of dengue, as indicated by the WHO criteria and previous studies. Our algorithm offers simple indicators for severity, including haematocrit, GCS, urine protein, creatinine, and platelet count, all of which are measured on admission. This model is potentially useful for guiding inpatient monitoring and outpatient management of fever cases. The model does require further validation against other datasets from cohort studies conducted in a variety of settings, with the goal of establishing a universal algorithm for guiding clinical management of severe dengue in resource-limited settings.