4.1.2 Distinguishing Dengue from other Febrile Illnesses in an early Stage of Disease…
4.1.2.1 White Blood Cell Count, absolute Numbers of Lymphocytes and Temperature, but not Symptoms can Diagnose Dengue Infection in an early Stage of Disease…
A recent study reported a predictive model based on logistic regression to diagnose dengue fever within the first two days of admission (Chadwick et al., 2006). This study mainly used clinical data combined with symptoms to predict dengue infection and showed a sensitivity of 84%. By contrast, our model (DENPRE_TOTAL_453; Figure 3.1; Page 94) using clinical data within 72 hours of onset of disease also shows a sensitivity of 82%, and integration of cytokine data (DENPRE_INCYTA_291; Figure 3.7; Page 105) resulted in an increased sensitivity of 90%. It is intriguing that our classification tree based on clinical data and Chadwick’s model show striking similarity to each other. The first split chosen by our tree is represented by a white blood cell count equal or lower than 4800 cells/mm3 and this threshold is supported by the logistic regression model which defined a white blood cell count of lower than 5000 cells/mm3 as a genuine risk factor for dengue infection. The logistic regression model of Chadwick’s study further used rash, creatine, bilirubin and prothrombin time as discriminators whereas our presented classification tree included the absolute numbers of lymphocytes and the body temperature as further splitting criteria. We were not able to include the clinical data used by the logistic regression model because the respective variables were not measured in the EDEN study. Furthermore, model calculations only based on symptoms did not result in a reliable classifier suggesting
that detection of dengue specific symptoms is only possible in a later stage of disease.
The advantage of our decision tree over the recently published logistic regression model not only lies in the early time point of sample collection but also in the fact that we were predicting a PCR positive result which shows a sensitivity higher than 99%.
The IgM assay that Chadwick and collaborators based their logistic regression on, shows a slightly lower sensitivity. PCR is more specific than the IgM assay because IgM can be also found in other flaviviruses although these are not frequent in Singapore.
In addition, we were able to analyze different subgroups of dengue patients which clearly revealed that having a lower white blood cell count represents a main but not the only indicator for a positive dengue PCR result. The decision tree especially emphasized the decreased numbers of lymphocytes as a main characteristic of acute dengue infection. This finding is supported by another study also showing low numbers of lymphocytes in an acute stage of dengue disease (Azeredo et al., 2001).
Our clinical model shows a very high accuracy of over 90% and the stability is underlined by a narrow confidence interval of the calculated AUC of the ROC curve (Table 3.3; Figure 3.2; Page 96). Moreover, the odds ratios calculated either on the whole dataset or at each decision node show high level of significance and the similarity of the two models based either on the total dataset (453 cases) (DENPRE_TOTAL_453; Figure 3.1; Page 94) or on 291 cases (DENPRE_EXCYT_291; Figure 3.3; Page 98) proves the validity of our calculations.
It is important to note that the chosen classifier on the ROC curve is skewed towards identifying true negative cases because by definition, the program chooses the model
at a specific threshold with the highest profit for the major class. This resulted in the lower sensitivity of 82% compared to a higher specificity of 91% (Table 3.3; Figure 3.2; Page 96). The model was evaluated by stratified k-fold cross validation (k=10) which is considered to be a powerful method to prevent data overfitting (Kohavi, 1995). Nevertheless, it is necessary that our decision tree model gets evaluated in a new epidemic setting. Assessment of the classification tree on negative patients enrolled in 2006 showed similar classifier performance. However, evaluating the classifier in terms of its ability to identify dengue positive cases is of much higher relevance, especially in a setting where other flaviviruses and symptomatically similar diseases like Chikungunya are frequent.
This presented model includes only basic clinical parameters and is useful in providing assistance in dengue diagnosis in an early stage of disease. To some extent, it might be a useful substitution in epidemic settings where lack of infrastructure and under representation of trained staff exacerbate the implementation of either the commonly used virus isolation, the serologic PCR or the IgM assay. Measuring the white blood cell count and the number of lymphocytes are commonly used basic hematological tests that are not laborious, do not require highly trained staff and are probably more economical than the commonly used dengue diagnostic tools. Hence, the model may allow early dengue diagnosis at low cost. Combination of this diagnostic model with our proposed severity prediction might contribute to better case management and eventually, to a better prevention of DHF/DSS cases.