3.1 Decision Tree Analyses of Clinical Data
3.1.2 Prediction of Disease Severity in Dengue Patients
3.1.2.8 Severity Prediction based on Cytokine Data
It is known that cytokines are important factors in determining severity of disease and thus, a tree only including the cytokine data was constructed. This would provide us with a better picture of data interference and second, with a more detailed view of which cytokines could be important for disease pathogenesis. The following calculations revealed that classifier performance was strongly dependent on which cytokines were included into the tree. Therefore, we only picked the cytokines IP_10_1, IL_10_1 and IL_8_115 which showed the highest accuracy for severity prediction and this resulted in a good classifier performance of 0.75 for the prediction of ‘low’ (95%CI: 0.64, 0.86) as well as of ‘high’ (95%CI: 0.65, 0.85) cases (minimum cases was set to 23 with a pruning confidence of 25%). K-fold cross validation (k=10) determined an overall error rate of 29.028% with a sensitivity of 66% and a specificity of 76% (Table 3.53; Figure 3.30). The model (SEVERE_CYTOA_89_IL8) (Figure 3.28) picked IP_10_1>1697.9 (OR: 11.20; 95%CI: 7.97, 14.44) as the first and IL_8_1>3.34 (OR: 5.93; 95%CI: 2.52, 9.33) as the second splitting criterium (Table 3.51; Table 3.52). IL_10_1 was not integrated into the tree, but exclusion of this cytokine from the analysis, resulted in a decreased performance. This suggested that IL_10_1 had a supporting role on either IP_10_1 or IL_8_1 but as a predictor would be inferior to IL_8_1. Even decreasing of the tree parameter “minimum cases” did not result in the inclusion of IL_10_1.
Therefore, we checked for the influence of IL_8_1 in the tree and excluded this cytokine for further calculations. This resulted in a tree (SEVERE_CYTOA_89_IL10)
15 IP_10=interferon inducible protein 10; IL_10=interleukin-10; IL_8=interleukin-8; OR=odds ratio;
(Figure 3.29) that had IL_10_1>6.66 (OR: 5.63; 95%CI: 2.27, 8.98) as the second splitting criterium (Table 3.52; Table 3.52) and showed a higher sensitivity (70%) but in turn a lower specificity (72%). The overall accuracy was decreased which was represented by an AUC of 0.74 (95%CI: 0.62, 0.85) for ‘low’ and a AUC of 0.73 (95%CI: 0.63, 0.84) for ‘high’ cases. The profit averaged 0.397 with an overall error rate of 30.139% (Table 3.54; Figure 3.31).
ROOT 33 low / 56 HIGH
IP_10_1 <= 1697.9 4 low / 34 HIGH
IP_10_1 > 1697.9 29 LOW / 22 high
IL_8_1 > 3.34 20 LOW / 6 high
IL_8_1 <= 3.34 9 low / 16 HIGH ROOT
33 low / 56 HIGH
IP_10_1 <= 1697.9 4 low / 34 HIGH
IP_10_1 > 1697.9 29 LOW / 22 high
IL_8_1 > 3.34 20 LOW / 6 high
IL_8_1 <= 3.34 9 low / 16 HIGH
Figure 3.28: SEVERE_CYTOA_89_IL8: Decision tree for severity prediction calculated on 89 patients only including cytokine data and using interleukin-8 as the last splitting criteria. IP_10=interferon inducible protein 10; IL_8=interleukin-8.
ROOT 33 low / 56 HIGH
IP_10_1 <= 1697.9 4 low / 34 HIGH
IP_10_1 > 1697.9 29 LOW / 22 high
IL_10_1 > 6.66 21 LOW / 7 high
IL_10_1 <= 6.66 8 low / 15 HIGH ROOT
33 low / 56 HIGH
IP_10_1 <= 1697.9 4 low / 34 HIGH
IP_10_1 > 1697.9 29 LOW / 22 high
IL_10_1 > 6.66 21 LOW / 7 high
IL_10_1 <= 6.66 8 low / 15 HIGH
Figure 3.29: SEVERE_CYTOA_89_IL10: Decision tree for severity prediction calculated on 89 patients only including cytokine data and using interleukin-10 as the last splitting criteria.
IP_10=interferon inducible protein 10; IL_10=interleukin-10.
Table 3.51: SEVERE_CYTOA_89: Decision tree for severity prediction calculated on 89 patients only including cytokine data. Statistical analysis of splitting criteria performed on the whole dataset.
IP_10=interferon inducible protein 10; IL_10=interleukin-10; IL_8=interleukin-8; 1=1st visit data;
RR=relative risk; OR=odds ratio; CI=confidence interval.
p value Decision Node Feature RR OR 95% CI (OR)
IP_10_1 [pg/ml] 5.40 11.20 7.97, 14.44 < 0.001 Cut-off value > 1697.9
IL_8_1 [pg/ml] 2.57 4.48 1.95, 7.00 0.001 Cut-off value > 3.34
IL_10_1 [pg/ml] 1.50 1.86 -0.68, 4.40 0.004 Cut-off value > 6.66
Table 3.52: SEVERE_CYTOA_89: Decision tree for severity prediction calculated on 89 patients only including cytokine data. Statistical analysis of splitting criteria performed on each subgroup at the decision nodes. IP_10=interferon inducible protein 10; IL_10=interleukin-10; IL_8=interleukin-8. 1=1st visit data; RR=relative risk; OR=odds ratio; CI=confidence interval.
p value Decision Node Feature RR OR 95% CI (OR)
IP_10_1 [pg/ml] 5.40 11.20 7.97, 14.44 < 0.001 Cut-off value > 1697.9
IL_8_1 [pg/ml] 2.14 5.93 2.52, 9.33 0.003 Cut-off value > 3.34
IL_10_1 [pg/ml] 2.16 5.63 2.27, 8.98 0.005 Cut-off value > 6.66
Table 3.53: SEVERE_CYTOA_89_IL8: Summary of K-fold (k=10) cross validation for severity prediction based on 89 patients only including cytokine data and using interleukin-8 as the last splitting criteria.
Overall Evaluation Value (n=89) Confusion Matrix Total
misclassifications 26.0 Predicted Class
Overall error rate 29.028% high low
SE of error rate 12.838 42 14
high
Actual Class
(76%) 214%) Average profit 0.419
SE of profit 0.257 12 21
low
95%CI: 0.65,
0.85 (34%) (66%)
AUC high 0.7506
95%CI: 0.64, 0.86 AUC low 0.7531
Table 3.54: SEVERE_CYTOA_89_IL10: Summary of K-fold (k=10) cross validation for severity prediction based on 89 patients only including cytokine data and using IL_10_1 as the last splitting criteria.
Overall Evaluation Value (n=89) Confusion Matrix Total
misclassifications 27.0 Predicted Class
Overall error rate 30.139% high low
SE of error rate 18.026 40 16
high
Actual Class
(72%) (28%) Average profit 0.397
SE of profit 0.361 11 22
low
95%CI: 0.63,
0.84 (30%) (70%)
AUC high 0.7338
95%CI: 0.62, 0.85 AUC low 0.7354
Figure 3.31: SEVERE_CYTOA_89_IL10: Receiver operating characteristics (ROC) curve for severity