There were 1,384 patients enrolled in the study; 1,311 of these patients had a diagnosis of DHF, DF, or OFI (presumed viral, non-dengue illness). For 65 of these patients (55 with OFI, 6 with DF, and 4 with DHF), a day of defervescence could not be assigned, and they were excluded from the analysis. Some patients with OFI were discharged from the study before defervescence because of a negative PCR. Data obtained from the day of enrollment showed no significant differences between these 65 excluded patients and the remainder of the cohort with the following exceptions: (1) patients with DF included in the study were older than those with DF that were excluded (8.7 years versus 6.0 years; P = 0.03), and (2) patients with OFI included in the analysis had higher percent monocytes than patients with OFI that were excluded (3.8% versus 2.4%; P = 0.01). An additional 19 patients did not have information on all clinical laboratory variables used in the analysis and were excluded.
There were 1,227 patients included in the analysis (228 with DHF, 386 with DF, and 613 with OFI). Patients with DHF were classified by grade as follows: grade 1 (N = 59), grade 2 (N = 129), grade 3 (N = 39), and grade 4 (N = 1). There were 1,058 patients who completed the study at QSNICH and 169 patients from KPPPH (). The number of patients included in each model is shown in .
Study sample characteristics
Figure 1. Flow chart of study. Boxes show the total number of patients enrolled in the study, reasons for exclusion from the analysis, and number of patients from the training dataset (QSNICH) and test dataset (KPPPH) used in each model. DF = dengue fever; DHF (more ...)
We compared the characteristics of subjects enrolled at QSNICH and KPPPH. A higher proportion of patients with DHF were enrolled at KPPPH compared with QSNICH (30.8% versus 16.6%; P < 0.001). Patients presenting to KPPPH were also older than those presenting to QSNICH (8.9, 95% confidence interval [CI] = 8.4–9.3 versus 7.7, 95% CI = 7.5–7.9; P < 0.001). There were no differences between the hospitals with regards to gender and days ill at presentation. However, at QSNICH, patients with DHF presented later than patients with DF or OFI (2.3 versus 2.1 and 1.7, respectively; P < 0.001), and patients with DF presented later than patients with OFI (2.1 versus 1.7, respectively; P < 0.001). Additionally, at QSNICH, patients with DHF and patients with DF were older than patients with OFI (P < 0.001). At KPPPH, a lower percentage of males was seen among patients with DF compared with patients with DHF or OFI (χ2 P = 0.03 and χ2 P = 0.02, respectively).
shows the results of univariate logistic regression modeling using the QSNICH data. Clinical laboratory variables distinguishing each of the different diagnostic categories included minimum platelet count, maximum daily hematocrit (Hct), AST > 100, maximum ALT, and a positive tourniquet test (> 20 petechiae). Among these variables, unit decreases in minimum platelet count and increases in maximum Hct, AST, and ALT and having a positive tourniquet test were associated with having the more severe outcome in each of the models. The AUC for these variables ranged from 0.92 to 0.54. Additional variables of decreasing age and decreasing minimum WBC count distinguished between all categories except DHF versus DF and were associated with having the more severe outcome in each model. The maximum percent lymphocytes and percent neutrophils distinguished patients with DHF versus DF and those with dengue versus OFI.
Univariate analysis of maximum or minimum clinical laboratory variables using the QSNICH data
shows the results of multivariable analysis using QSNICH data. Maximum AST and ALT were found to be highly correlated (Pearson correlation = 0.86), and therefore, only maximum AST was used in the modeling. Incremental decreases in minimum platelet count (one unit = 25,000 cells/mm3) and a maximum AST > 100 were found to be associated with the more severe outcome in each of the final multivariable models. Additionally, increases in the maximum daily Hct were associated with having a diagnosis of DHF compared with DF or DF + OFI. Increases in maximum percent neutrophils, age, and incremental decreases in minimum WBC count (one unit = 500 cells/mm3) were associated with a diagnosis of DHF versus DF. All variables except maximum percent lymphocytes and percent neutrophils showed an association with having severe dengue illness compared with non-severe dengue or OFI. In addition to distinguishing patients with severe dengue, a tourniquet test of petechiae ≥ 20 was also included in the model for distinguishing patients with serologically confirmed dengue from those with OFI. Age is included in the models of DHF versus DF and severe dengue versus all others and shows that younger patients have higher odds of DHF or severe dengue after adjusting for all other variables in the model.
Multivariable models among QSNICH data
The probability cutoff for each multivariable model was defined as the cutoff that gave the highest percent of patients correctly classified where the sensitivity remained higher than the specificity. illustrates the tradeoff between sensitivity and specificity for each probability cutoff for all of the multivariable models and shows the optimal cutpoint.
Figure 2. Sensitivity and specificity of multivariable logistic regression models from the training dataset. Sensitivity is indicated by the solid blue line, specificity is indicated by the solid red line, and the optimal probability cutoff is indicated by a solid (more ...)
Validation of multivariable models.
The frequency distribution of each diagnosis (DHF, DF, and OFI) according to the optimal probability cutoff for each model is given in , showing the distribution of diagnoses in both the training and test datasets. The sensitivity/specificity analysis, including AUC, positive and negative predictive values, and percent correctly classified, and validations with the test dataset are indicated in . When applying each model to the test dataset, the sensitivity decreased by 2.7% (DHF versus DF) to 8.9% (DHF versus all others). The specificity increased for each validation except severe dengue versus all others, where specificity decreased and sensitivity increased. The percent correctly classified decreased by 0.7% (DHF versus DF) to 4.3% (severe dengue versus all others). The final model distinguishing between patients with dengue and patients with OFI performed the best, giving the highest AUC, specificity, and percent correctly classified in both the training and test datasets.
Figure 3. Distribution of calculated probabilities among each diagnosis for each model. Blue = DHF; red = DF; gold = OFI. In each section, the top panel represents the distribution of probabilities in the training dataset (QSNICH), and the bottom panel represents (more ...)
Validation of QSNICH multivariable models to KPPPH data using optimal probability cutoff
Outside of a research setting, the entire spectrum of a patient's illness may not be available. Therefore, each model was applied to the test dataset using data only at defervescence and/or at 24 hours after defervescence, the period of greatest risk for plasma leakage when patients are most likely to require hospitalization. When using data from the test dataset obtained in these later stages of illness, the sensitivity for most models decreased. However, using data at 24 hours after defervescence, it remained moderately high for DHF versus DF, increasing from 76.9% to 86%, and DHF versus all others, at 71%.
shows the four validated multivariable models. Coefficients for the calculation of probability for each model are obtained by taking the natural logarithm of the odds ratio for each variable in .
Figure 4. Validated multivariable probability models for classifying patients with dengue. (A) DHF versus DF. (B) DHF versus all others. (C) Dengue versus OFI. (D) Severe dengue versus all others. Note that categorical variables (percent neutrophils, percent monocytes, (more ...)
Classification from models compared with WHO and expert-physician diagnosis of DHF.
shows the percent agreement and κ statistics between the model's classification of DHF, the physician's diagnosis of DHF, and the WHO diagnosis of DHF. When applying the optimal probability cutoffs, the percent agreement between the model and the WHO classification of DHF compared with DF was 80.0% (κ = 0.58; P < 0.001). The percent agreement between the model and the WHO definition of DHF improved to 86.2% compared with all others (DF + OFI; κ = 0.60; P < 0.001). In both cases, the model had a higher percent agreement with a WHO diagnosis of DHF than the physician's diagnosis of DHF.
Comparison of percent agreement and κ statistics between final models, WHO DHF criteria, and physician's final diagnosis
The model of DHF versus DF classified 42.0% (258/614) of patients with dengue as having DHF. The physician diagnosed 37.1% (228/614) of patients with dengue as DHF, and strict adherence to the WHO definition would have diagnosed 33.1% (203/614) patients with dengue as DHF. The model of DHF versus all others classified 51.6% (317/614) of patients with dengue as DHF. The model of severe dengue versus all others classified 49.2% (302/614) of patients with dengue as having a severe dengue illness, including 203 patients diagnosed by the physician as DF. Only 6.5% (40/614) of patients with dengue were diagnosed by the physician as DHF grade 3 or 4.