4.1.3 Severity Prediction in an early Stage of Disease
4.1.3.2 A high viral Load combined with a secondary Infection is a genuine Risk
We investigated the possibility of predicting development of a low platelet count on day five, six and seven by including clinical as well as immunological data into one reliable model. In a first attempt, we only exploited clinical data collected on the 1st visit which resulted in a classification tree (SEVERE_TOTAL_125; Figure 3.20; Page 135) with a sensitivity of 81% and a specificity of 83%. The fist decision node made use of the platelet count as a splitting criteria with a threshold set to <=108 platelets/mm3. It is evident that patients already having a low platelet count on the first visit are at greater risk for the development of an even lower platelet count during
infection. Consequently, we considered the group of patients presenting a higher platelet count at the first visit as more interesting with regard to severity prediction.
Remarkably, we observed that dengue patients having a higher viral load along with a secondary infection are at major risk to become thrombocytopenic during infection.
The calculated decision tree explicitly shows that having a high platelet count in combination with a low viremia in an early stage of disease is almost no risk for the progression into severe disease manifestations and shows a relative risk reduction (RRR) of 91%19. The RRR value may be an important consideration regarding drug discovery and vaccine development. Our model clearly shows that development of a drug and/or vaccine that may reduce viral load in a very early stage of disease is highly desirable and might prevent the progression to the more severe dengue forms.
Furthermore, it highlights secondary infections as a risk factor which is a major challenge in vaccine implementation.
The presented model also supports the finding of another study which clearly showed that high viremia titers 3 days after onset of fever correlated with severe disease 2 days later at the time of defervescence (Vaughn et al., 2000). We observed that there was no correlation regarding viral load and secondary infection on the whole dataset as well as at the specific decision node (results not shown). Hence, we can consider viral load and secondary infections as two independent risk factors that, in combination, may have a significant influence on disease outcome. This is contradictory to the theory of antibody dependent enhancement (ADE) which, strictly applied, defines a direct relationship of viral load with secondary infections (Halstead, 2003). Briefly, it
19 RRR of people with a high platelet count along with a low viral load is calculated by 1/(Relative Risk of people having a high platelet count along with a high viral load)-1; 1/11.34-1=-0.912 (For values
is thought that binding of cross-reactive antibodies from a primary infection enhances uptake of virus via Fc-receptors into target cells and subsequently leads to higher replication causing higher viremia. Our observations would not be against the theory of cross-reactive T-cells which are thought to be triggered upon a secondary infection and which subsequently influence disease outcome (reviewed in (Green and Rothman, 2006)). Looking at the sequence of decision nodes and at the frequency distributions in our proposed model, we hypothesize that a higher viral load might trigger a stronger immune response and a combined secondary infection enhances the risk of an increased inflammatory response which finally causes a more severe disease outcome.
We can assume that the calculated decision tree and its chosen features based on the overall dataset (125 cases) are reliable because constructing a tree on a dataset with smaller sample size (89 cases) results in the exact same tree (SEVERE_EXCYT_89;
Figure 3.24; Page 143) with similar performance. The importance of platelet count, viral load and secondary infections in an early stage of infection is further underlined by the models constructed only on the data of hospitalized cases. The two constructed trees (SEVHOSP_TOTAL_71; Figure 3.22; Page 139 and SEVHOSP_EXCYT_52;
Figure 3.32; Page 155) (71 cases or 52 cases) both include as a first splitting criteria the platelet count with the same threshold and use as a second decision node secondary infections followed by viral load. The switch in the splitting criteria between secondary infections and viral load might be caused (1) by the smaller datasets used, (2) by the fact that the datasheets represent more similar cases with regard to viral load thus removing statistical significance and creating higher data entropy, (3) by the fact that the frequency distribution of secondary infections still shows the same strength because its dependency on the platelet groups rather than on hospitalization and (4) by the change in class majority meaning that, in this situation, the more severe group
represents the major class. The latter suggests that the further splitting criteria were caused by data overfitting of the tree and do not indicate genuine differences which is finally underlined by comparable weak statistical values on the whole dataset.
By way of example, taking the two correctly classified DHF cases and following them through the classification steps in our decision tree, validates the calculated model.
One patient was presented with a platelet count lower than the chosen threshold and was therefore not further classified. The other patient showed a higher platelet count, a higher viral load (meaning the Ct-value was lower than the chosen threshold) and in addition, was positive for IgG antibodies indicating a secondary infection. The simplicity of the model along with the chosen features that are easy to measure as well as the high accuracy of 83% represent advantages with regard to the early assessment of disease severity in dengue patients. Hence, implementation of the model based on clinical data in combination with the severity model may have an important share in tackling severe dengue epidemics.