Prediction of mortality in severe dengue cases

Dengue infection has occurred in Malaysia for over a century [1]. Since the 1980s, outbreaks have been increasing in frequency [2]. Dengue is now accepted as endemic in Malaysia for the last 20 years. Despite massive efforts by various central government departments managing public health and many other agencies of the local government to curb it, the number of dengue cases reported continues to increase, and a rise in mortality has paralleled that increase. A national key performance indicator was established; a case fatality rate of less than 0.2% was targeted.

An epidemiological analysis in Malaysia revealed that all four DENV serotypes were found to be co-circulating during the period 2000–2012, although the predominant serotypes varied over time, both nationally and within the individual states in Malaysia [3]. An example of this variation occurred in the year 2014 when the predominant serotype had switched twice from DENV-2 to DENV-1, in February and June [4]. Prior to 2013, the case fatality rates were less than 0.2%, which was within the national target. However, from 2013 onwards, dengue incidence and case fatality rate started to increase and in 2015 had almost doubled [5].

Along with this increase in burden of disease the management of dengue, especially severe dengue, continues to baffle physicians. Despite continuous efforts to upgrade and improve clinical practice guidelines, physicians are confronted with patients who do not fall neatly into typical description of the disease, hence pose a clinical conundrum. The spectrum of dengue infection ranges from asymptomatic to mild febrile illness, dengue with warning signs and severe dengue. Severe dengue is defined as end organ involvement, severe bleeding and/or severe plasma leakage. Although only a small percentage of patients went on to develop severe dengue, with a study quoting a proportion of 5% of all dengue infections studied [6], these severe dengue carries significant mortality. Mortality in dengue is believed to be potentially preventable. Dengue is a dynamic, systemic illness and physicians have always found themselves summing up best estimates based on personal experience.

In that respect, additional knowledge would be indispensable in the management of dengue patients. Many studies have attempted to predict those that will develop severe dengue using clinical features and laboratory findings. There have also been studies that examined inflammatory markers, cytokines and genetic markers in predicting severity of dengue [7]. However, we believe that there is a need to refine this further - to predict mortality among those diagnosed to have severe dengue. We therefore aimed to identify factors at diagnosis of severe dengue that predict mortality and, built and assess predictive models based on these identified factors.

The study was approved by the Medical Research and Ethics Committee (MREC), Ministry of Health of Malaysia (Research ID NMRR-15-2023-24,849). As the study involved data collection from case notes only, the MREC granted a waiver of informed consent. Our report is based on the STrengthening the Reporting of OBservational studies in Epidemiology (STROBE) 2015 guideline.

The study was a retrospective study of adult patients with severe dengue who were admitted into Hospital Kuala Lumpur throughout 2014. Sample recruitment was conducted from 1 January 2014 until 31 December 2014. Data were extracted from case notes of selected patients who fulfilled the study’s inclusion criteria. A data collection pro forma was used to ensure integrity of data. As the information in the case notes were information used in patients’ management, and thus would have been corrected for errors by managing clinicians, therefore we considered the information in the case notes as accurate.

Patients were selected for inclusion if, 1) they were ≥ 18 years old, 2) their presentation satisfied our criteria for suspected dengue, 3) the case fulfilled our definition of severe dengue, and 4) the presence of dengue viral infection was confirmed via NS1 antigen, high-titre level of IgG or positive IgM, from an admission serum sample. Our criteria for suspected dengue, based on the World Health Organization (WHO) 2009 criteria, were fever plus any two of: 1) aches and pain, 2) nausea and/or vomiting, 3) rash, 4) leucopenia, or 5) presence of any warning signs. The definition of severe dengue in our study was also based on WHO 2009 definition. We defined severe dengue by presence of any one of the following: 1) decompensated shock due to severe plasma leakage, 2) compensated shock due to severe plasma leakage, 3) respiratory compromise due to severe plasma leakage, 4) severe hepatitis, 5) severe bleeding that required intervention, or 6) severe organ involvement such as acute kidney injury defined by elevated serum creatinine (according to gender-specific levels), myocarditis or encephalopathy. Decompensated shock was defined by presence of systolic blood pressure (SBP) less than 90 mmHg, or mean arterial pressure (MAP) of less than 65 mmHg, or a drop in systolic blood pressure of more than 40 mmHg from patient’s known usual baseline readings. Compensated shock required signs of impaired peripheral perfusion, occurring in combination rather than singly, in presence of systolic blood pressure of ≥90 mmHg. Severe hepatitis was defined as AST level ≥ 1000 IU/L, or ALT level ≥ 1000 IU/L.

Clinical management at Hospital Kuala Lumpur followed the Malaysian Clinical Practice Guideline on Management of Dengue Infection in Adults 2010 and the WHO 2009 clinical practice guideline for dengue.

Our outcome of interest was mortality among patients diagnosed with severe dengue. Patients were grouped into those who died and those who survived. We reviewed case notes to collect data. Data collected were baseline characteristics, clinical parameters, laboratory findings and time of events. Times of events were: dates and time of fever onset, time of admission, time of diagnosis of severe dengue, time of defervescence (start of temperature persistently < 38 °C) and time of occurrence of outcome. Clinical and laboratory parameters that were closest to the time of diagnosis of severe dengue were collected. These variables formed as candidate predictors for our prediction model. We also collected data on nadir and highest levels of relevant laboratory parameters, along with their timings, for descriptive purposes. Data on types of severe dengue diagnoses that occurred in a patient were also collected.

We calculated that a sample size of 13 deaths was needed for this study for an AUROC of 70%, at confidence level of 95% and power of 80%. Analyses were performed in SAS University Edition (Copyright © 2012–2016, SAS Institute Inc., Cary, NC, USA). Continuous variables were tested for uniformity using the Kolmogorov-Smirnov test and normality with the Shapiro-Wilk test. As our data were mostly non-parametric, we used non-parametric analyses for data interrogation. Categorical variables were expressed as frequencies and percentages. Continuous variables with non-normal distribution were summarised as median and inter-quartile range (IQR).

Identification of all factors that were significantly associated with mortality was made using the Chi-square test or Fisher’s exact test for categorical variables and Mann–Whitney U test for continuous variables. We then performed multivariate analysis, with age and gender adjustments, to determine factors independently associated with mortality.

In order to build a predictive model, we used variables at the time of diagnosis of severe dengue as base for variable selection in model-building. Variable selection was performed using five-fold cross-validated Lasso regression. Selected variables were then used to build logistic regression models with cross-validation. Cross-validation and Lasso addressed overfitting that is known to occur in logistic regression. Logistic regression models built were composed of a combination of 2 of the Lasso-selected variables; and were adjusted for age and gender. All possible combinations were built. Finally, their AUROC were computed to allow comparison between models.

All tests of significance were 2-sided, and we took p-value < 0.05 indicating statistical significance.

Our study made two notable findings: the first being description of aspects of severity of a severe dengue cohort; and the second, the rigorous building of a predictive model, which was selected by assessment and comparison of many models, that predicts death early at the recognition of severe dengue.

Our cohort bears some characteristics of Malaysia’s population. The National Health Morbidity Survey 2015 determined that 30.6% of the population were obese and 17.5% were diabetics, figures similar to this study [8].

The fatality rate of this study was 10.1%. Previous studies similar to ours had rates of 3.8% [9] and 18.6% [10] in Brazil. Amâncio et al. [10] non-survivors had median (IQR) AST 151 (43–474) U/L and mean (SD) platelet count 83.3 × 10³/μL (76.2 × 10³/μL); the other study [9] however, had no similar parameters for comparison. These levels were worse in our study. Other studies, which used ICU admission as the defining criterion of their cohorts, had case fatality rates of 23.1% [11] in Taiwan whilst in India rates were 11.1% [12] and 6.1% [13]. Chen et al. [11] had non-survivors with mean (SD) AST 3444.4 (4191.9) U/L and mean (SD) nadir platelet 3.5 × 10³/μL (4.3 × 10³/μL). In contrast, the non-survivors of our study had even higher peak AST level, suggesting a more severe cohort in ours.

Our study showed that about a third of patients presented as severe dengue upon admission to hospital whilst the remaining developed severe dengue after admission. Duration of fever onset to severe dengue diagnosis were shorter in those who presented as severe dengue as compared to those who developed severe dengue after admission, median(IQR) 4.38 [2] days vs median(IQR) 5.02(1.78) days, p = 0.003, respectively. Interestingly, there were no difference in mortality and in the duration of fever onset to admission between those who presented as severe dengue as compared to those who developed severe dengue after admission, p = 0.95 vs p = 0.29, respectively.

Guidelines in dengue have repeatedly highlighted that the timing of deterioration is around the time of defervescence and within the critical phase that ensues. It has also been noted that organ impairment follows the same timing. Intriguingly, we found that almost 60% of patients were still febrile at diagnosis of severe dengue; and the proportion of patients who were still febrile at diagnosis of severe dengue was statistically significantly more in those who survived as compared to those who died (Table 2). In fact, those who were febrile at diagnosis of severe dengue were less likely to die as compared to those who had defervesced, OR 0.29 (95% CI: 0.09–0.92, p = 0.03). The importance of timing of development of severe dengue in the course of illness needs to be further explored.

Our findings that nadir platelet occurred about a day after nadir WBC is consistent with a previous study [14]. However, another intriguing finding of our study is that the timing of peak serum creatinine coincided with nadir WBC in those who survived whereas in those who died it coincided with the later nadir platelet.

Based on our cohort, independent predictors of death at the time when the diagnosis of severe dengue was made were: lethargy, bleeding, pulse rate, serum bicarbonate and serum lactate. There have been only 2 studies so far, that examined factors associated with death among severe dengue patients, which used WHO 2009 classification. A study utilising notification database from Brazil [9] of mixed age groups showed age > 55 years (OR 4.98), gastrointestinal bleeding (OR 10.26), haematuria (OR 5.07), and thrombocytopenia (OR 2.55) were factors associated with death. However, exact timings of these parameters were not included. A second study [10] was of severe dengue patients admitted to ICU in Brazil. In this study of 97 patients admitted to ICU, parameters taken within 24 h of ICU admission that were found to be associated with death were: comorbidity of chronic renal disease (OR 15.6), presence of persistent vomiting (OR 4.25), lethargy (OR 3.23), dyspnoea (OR 3.27), elevated WBC, higher serum creatinine and lower serum albumin.

We found that the best prediction model to predict death at the time when the diagnosis of severe dengue was made is a model that incorporated serum bicarbonate and ALT levels taken at that time. This is rather fortuitous for a few reasons. Firstly, both serum bicarbonate and ALT are objective measures as opposed to subjective warning signs such as lethargy and bleeding which have inherent variability in establishing their presence and severity. Secondly, both are currently accessible laboratory tests.

A recent study on serum lactate in dengue [15] found that this biomarker is a good predictor of severe dengue (AUROC for peripheral venous lactate at admission was 0.84 [95% CI: 0.72–0.97]). In perspective, our study revealed that in predicting death, lactate-incorporated models had lower AUROCs as compared to those of bicarbonate-based models (Table 4). Additionally, we showed that nadir serum bicarbonate occurred earlier than highest serum lactate level (Table 2) in patients with severe dengue. In fact, at diagnosis of severe dengue, more patients had abnormal serum bicarbonate than abnormal serum lactate (46.7% vs 27.6%, respectively). These are important aspects to consider in clinical practice as earlier management will be more advantageous. Therefore, whilst lactate could be an additional alternative criterion to establish the diagnosis of severe dengue, our study suggests that lactate alone is not sufficient in prognostication of patients with severe dengue.

Unexpectedly, serum creatinine was not found to be an independent predictor of death by multivariate analysis. However, its incorporation into our models led to good AUROC performances.

Under the assumption that the time of recognition of severe dengue is approximately the actual time of development of severe dengue, we believe that prognosticating mortality in patients with severe dengue at the time of its recognition provides a sensible approach. We postulate that at this time, underlying pathophysiological processes which determine outcome would most likely have reach significance. Prognosticating prematurely before this moment may have little specificity, as the ultimate outcome determining processes may yet to occur. Prognosticating too late however is obviously futile. With our model management decisions may be better informed in terms of resource allocation, especially in conditions of high volume care. It has to be noted however that the underlying outcome determining pathophysiology has yet to be clearly elucidated. Further investigation into the kinetics of biochemistry with respect to timing of events in dengue, in particular the time of development of severe dengue, is needed and may perhaps fill this gap.

The main limitation of our study was the retrospective design. However, data accuracy was reasonable as management of patients followed standard management guidelines for dengue which have clear specifications of timing of blood investigations. Though only a single centre study, we believe the sample size was adequate as illustrated by the results of our findings. Finally, though we have rigorously built a predictive model and took steps to address overfitting, the actual performance of any model will vary according to the population it is applied on. As we have mentioned, our cohort is different from similar studies, hence our model will require population specific external validation and assessment.

In conclusion, the degree of severity observed in this study was more serious than those reported in similar studies. We showed that a large proportion of severe dengue occurred early, whilst patients were still febrile. Finally, the regression equation to our model is: Log odds of death among severe dengue cases = − 1.021 - 0.220(Serum bicarbonate) + 0.001(ALT) + 0.067(Age) - 0.190(Gender).