|
nif:isString
|
-
We conducted a retrospective cohort study in which we extracted from chart review demographic, epidemiological, co-morbidity, serial clinical and laboratory, radiological, treatment and outcome data from all confirmed adult dengue patients treated using a standardized dengue care path in the Department of Infectious Diseases, Tan Tock Seng Hospital, from 2006 to 2008. Prior to admission, patients were mainly referred from primary care centers where the standard is to monitor with full blood count, give symptomatic medication for fever, pain, nausea or diarrhea, and tell the patients to rest and drink copious fluid. Two groups of laboratory-confirmed dengue cases were considered: (i) had positive dengue polymerase chain reaction (PCR) (definite dengue) [28]; (ii) were positive to dengue immunoglobulin-M (IgM) or immunoglobulin-G (IgG) (probable dengue) (Dengue Duo IgM & IgG Rapid Strip, Panbio Diagnostic, Queensland, Australia). Patients who were diagnosed with severe dengue at the time of hospital presentation were excluded from further analysis; patients without SD but with warning signs were included (Figure 1). SD cases fulfill at least one of three criteria: (1) severe plasma leakage: associated with shock (tachycardia >100 or narrow pulse pressure <20 or blood pressure <90 mmHg) or respiratory distress, (2) severe hemorrhage: gastrointestinal tract bleeding such as haemetemesis or melaena or bleeding per rectum, pack cell transfusion or blood transfusion requirement or menorrhagia (3) severe organ impairment: defined as aspartate or alanine transaminase ≥1000units/l, acute renal impairment (creatinine ×2 upper limit of normal for age and sex (based on modification of diet in renal disease equation for Glomerular filtration rate = 75 ml/min) or baseline estimated from minimum creatinine recorded), encephalopathy and myocarditis [14]. In addition, the WHO proposed warning signs suggestive of high risk for SD were recorded: abdominal pain or tenderness, persistent vomiting (vomiting during two or more consecutive days), clinical fluid accumulation, mucosal bleed, lethargy or restlessness, hepatomegaly and concurrent increase in hematocrit and rapid decrease in platelet count (hematocrit change > = 20% concurrent with platelet <50×109/l occurring in the same day) [14].
Figure data removed from full text. Figure identifier and caption: 10.1371/journal.pntd.0002972.g001 Study inclusion criteria.PCR: polymerase chain reaction, WHO: world health organization, DHF: dengue hemorrhagic fever, SD: severe dengue. Variables presenting missing data in a proportion greater than 10% were removed from the analysis. Those variables with fewer missing data had their missing values replaced by an imputed value being the mean over all individuals in the study. For dichotomous variables, missing values were imputed by noes, the assumption being that if the sign or symptom were present, it would have been noted by the attending clinician. Variables that were not clinically relevant to severe dengue (e.g. runny nose, sore throat) or that conveyed similar information or information that was a linear combination of other variables were further removed from the analysis. Variables corresponded to values at presentation with the exception of hematocrit change > = 20% and drop in platelet count was obtained comparing the value at presentation with values during the course of the disease. In total, 66 explanatory variables extracted from chart review at presentation were employed and classified in groups: demographic and epidemiological characteristics; co-morbidities; symptoms and signs at presentation; laboratory results at presentation; and miscellaneous variables (see the Supporting Information Table S1 for the list of variables considered). Co-morbidities were scored for: congestive heart failure, cerebrovascular disease, chronic pulmonary disease, liver disease, diabetes, hemiplegia, renal disease, malignancy and AIDS. The scores were combined into the Charlson score and two variables indicating whether the individual had any comorbidity and whether individuals had Charlson score greater than three were respectively created. The large number of variables allowed us to minimize potential confounders while retaining variables that were clinically relevant. Prior to the analysis, we further tested the presence of collinear variables using variance inflation factors. Variables with a variance inflation factor greater than 3 were removed from the analysis stepwise using the package car in the R environment [29]. Predictive tools for SD development after presentation were constructed using logistic regression or its equivalent generalized linear models (GLM) with a logit link function and binomial errors [30]. Logistic regression or the GLMs used relate a group of explanatory variables to the log of the odds of a patient developing severe dengue versus not developing it. The GLMs were constructed and simplified using a stepwise procedure that started with the full model and removed one variable at a time using a chi-square test [31], [32]. The receiver operating characteristic (ROC) curve, the area under the ROC curve, sensitivity and specificity of the GLM models were estimated. Because our aim was to minimize the number of SD cases triaged for outpatient care, while allowing practitioners to decide on the tradeoff between sensitivity and specificity, we reported the specificity of the models given high sensitivity rates of 90%, 95% and 100% (see Figure 2, for a comparison of the ROC curves). Given the large number of potential predictors and to avoid the risk of overfitting or finding relationships by chance alone, we corrected the estimates of sensitivity and specificity for optimism—decrease in model performance in new patients compared to those in the sample studied [33]. The bootstrapping method was used to create samples with replacements from the dataset. A new model repeating all the stepwise model simplification process was constructed. The performance of the model on the prediction of the bootstrapped sample and the original sample was estimated. The optimism estimate was the difference between both performances and deducted to the original estimates of sensitivity and specificity [33]. 100 simulations were conducted per model to obtain stable estimates of optimism. In addition, the models were further validated using 5-fold validation by which the data are partitioned in 5 subsets, 4 subsets – the training dataset – are used to fit the GLM model and then employed to predict the remaining subset of the data – validation dataset. The procedure is then repeated changing the subset used as the validation dataset. The 5-fold validation process was repeated 1000 times and the mean of the root mean squared prediction error (RMSPE) was estimated for each model. The estimation of RMSPE was also subjected to optimism correction. Lower RMSPE indicated a better performance in the prediction of new cases and robustness of the model to potential associations found by chance alone. The R statistical environment was used for all analyses [34].
Figure data removed from full text. Figure identifier and caption: 10.1371/journal.pntd.0002972.g002 Comparison of the receiver operator characteristic curve in resourced and resource-limited settings for the GLM models fitted to predict SD in PCR confirmed cases (A) and PCR and serology confirmed cases (B). This study received ethical approval from the Domain Specific Review Board, National Healthcare Group, Singapore (DSRB-E/08/567) and all data analyzed were anonymized. Dengue cases were analyzed with and without laboratory information so that appropriate tools could be developed for both well-resourced and resource-limited settings. GLMs were fitted to six combinations of data and variables: GLM fitted to only PCR confirmed cases including laboratory variables to predict any form of SD.GLM fitted to only PCR confirmed cases excluding laboratory variables to predict any form of SD.GLM fitted to PCR and serology confirmed cases including laboratory variables to predict any form of SD.GLM fitted to PCR and serology confirmed cases excluding laboratory variables to predict any form of SD.GLM fitted to only PCR confirmed cases including laboratory variables to predict severe hemorrhage.GLM fitted to only PCR confirmed cases including laboratory variables to predict severe plasma leakage.
|
![[This page as RDF]](http://data.gesis.org/somesci/static/rdf-icon.gif){kind=icon}