Prognostic Models in Patients with Dengue: A Systematic Review

ABSTRACT.

There is uncertainty regarding the usefulness of predictive models for dengue prognosis. We performed a systematic review to identify and evaluate prognostic models in patients with dengue. We conducted a literature search in PubMed, Embase, and Literatura Latinoamericana y del Caribe en Ciencias de la Salud (LILACS) up to May 24, 2023. We included case–control and cohort studies that developed or validated multivariable prognostic models related to severity, hospitalization, intensive care unit (ICU) admission, or mortality in patients of any age with a laboratory-based diagnosis of dengue. A narrative synthesis of the performance measures of the prognostic models evaluated in each study was performed. Of the 4,211 articles, a total of 35 studies reporting information on 43 prognostic models were included. Among these, 35 were developmental and 8 were for external validation. Most models were designed to predict severity (n = 30), followed by mortality (n = 10), hospitalization (n = 2), and ICU admission (n = 1). The reported C-statistic in the models ranged from 0.70 to 0.95 for severity, 0.83 to 0.99 for mortality, 0.87 for hospitalization, and 0.92 for ICU admission. Calibration measures were poorly reported in the vast majority of models. According to the Prediction Study Risk of Bias Assessment Tool, the risk of bias was considered high for all included models, and applicability was of low concern for most models. Our study identified multiple prognostic models, particularly for predicting severity and mortality in patients with dengue. Although most models demonstrated acceptable discriminative ability, calibration measures were poorly reported, and the overall methodological design was poor.

INTRODUCTION

Dengue continues to be a major public health problem in developing countries.¹ Although most cases are mild, some progress to a severe form,² which can lead to serious complications, including death. The early identification of patients at increased risk of developing adverse events is critical for improving clinical outcomes and reducing the burden of disease.³

In recent years, there has been increasing interest in the development of accurate and reliable prognostic models capable of predicting the likelihood of severe dengue and its complications in affected patients.⁴ These models use clinical, epidemiological, and laboratory variables to estimate the individual risk for each patient. However, because of the heterogeneity in the characteristics of the population and uncertainty regarding which variables are the most relevant for predicting patient prognosis, it is necessary to conduct a thorough assessment of the existing literature. The information generated will be of great use to health professionals, patients, and public policymakers. The ability to predict early and accurately the risk of poor prognosis will facilitate better management of healthcare resources, more efficient triage of patients, and improved decision-making in the management of this disease, especially during outbreak or epidemic situations.¹

Our systematic review aimed to evaluate prognostic models in patients with dengue, as well as their applicability in different clinical scenarios. In addition, we sought to identify the most relevant clinical and laboratory characteristics for predicting of clinical outcomes, which could help guide the selection of predictors in future prognostic models.

MATERIALS AND METHODS

This systematic review was conducted in accordance with the recommendations of the Transparent Reporting of Multivariable Prediction Models for Individual Prognosis or Diagnosis: Checklist for Systematic Reviews and Meta-analyses (TRIPOD-SRMA),⁵ and was registered in the international repository PROSPERO (www.crd.york.ac.uk/prospero) under code CRD42023434595. The TRIPOD-SRMA checklist is presented in Supplemental Table 1.

Search strategy.

The search for studies was conducted in the electronic databases PubMed, Embase, and LILACS using search strategies specific to each database. The search was conducted from the creation of each database until January 2024, with no language or publication date restrictions. Free and controlled vocabulary terms related to dengue and prognostic models were used. The detailed search strategy for each database is provided in Supplemental Table 2. Additionally, a manual search was performed in the reference lists of the included articles, as well as in the articles that cited them.

Eligibility criteria.

Inclusion criteria were as follows: 1) case–control and cohort studies that developed or validated multivariable prognostic models based on individual patient information, without restrictions on the predictors included; 2) studies that included patients of any age with a diagnosis of dengue confirmed using real-time reverse transcriptase polymerase chain reaction (RT-PCR), XX (NS1) antigen, or serology attended in outpatient or inpatient care; 3) studies that considered any timeframe and prediction horizon; and 4) studies that evaluated outcomes related to severity, hospitalization, intensive care unit (ICU) admission, or mortality. We excluded conference abstracts, studies focused on prognostic factors, cross-sectional studies, case reports, case series, editorials, and reviews.

Selection of studies.

All articles from the electronic search were downloaded to EndNote 20 (EndNote, Berkeley, CA), and duplicate records were removed. Subsequently, all unique articles were exported to the Rayyan platform (Rayyan Systems, Boston, MA; https://rayyan.qcri.org/) for the study selection process. Two reviewers independently screened the titles and abstracts to identify relevant studies. In addition, two reviewers independently assessed the full text of the selected studies and recorded reasons for study exclusion. Any disagreements regarding selection based on the title and abstract and full text were resolved via consensus.

Outcomes.

The outcomes in this review included severity, hospitalization, ICU admission, and mortality at any follow-up time. The definitions reported in each study were taken into account for all outcomes.

Data extraction.

Information from each selected study was extracted independently by two reviewers, and any disagreements regarding extraction were resolved via consensus. A standardized data extraction form was used in Microsoft Excel (Microsoft Corp., Redmond, WA), based on the Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modeling Studies, which had been previously piloted.⁶

Assessment of the risk of bias and applicability.

The risk of bias and applicability of each study were assessed using the Prediction Study Risk of Bias Assessment Tool (PROBAST) by two independent reviewers, with discrepancies resolved via consensus.⁷ After classifying each study into one of three categories (model development study with or without external validation in the same publication, and external validation study of a previously developed model), the risk of bias was assessed according to the following four PROBAST domains (once per developed or validated model and per outcome): participants, predictors, outcome, and analysis. Signaling questions within each domain were answered with one of five options (“yes,” “probably yes,” “probably no,” “no,” or “no information”). Domain-level risk of bias assessments were scored using one of the following three options: low risk of bias, high risk of bias, or uncertain risk of bias. Additionally, the applicability of the model to the research question was evaluated. This assessment was performed per domain (for the first three domains only) with the following response options: low concern, high concern, and uncertain concern. Based on these domain-level assessments, an overall assessment per study was established based on the PROBAST tool for risk of bias (low, high, and unclear).

STATISTICAL ANALYSES

Only a narrative synthesis of the performance measures of the prognostic models evaluated in each study was performed because no more than five external validation studies were available for the same prognostic model and had evaluated the outcome for which they were originally developed. This threshold is based on expert recommendations because fewer studies may compromise statistical robustness and the meaningful interpretation of pooled estimates. To assess the discrimination of the models, the C-statistic or area under the curve was used. To assess calibration, calibration plots, calibration slope, or the ratio of observed to expected events were used. Publication bias was assessed through the funnel plot only if 10 or more external validation studies were available for each predicted model. The most frequent predictors of the prognostic models for severity and mortality were presented in bar charts, and the C-statistics with their 95% CIs were presented in forest plots. The statistical software R 4.3.1 (R Foundation, Vienna, Austria; www.r-project.com) was used for all analyses.

RESULTS

Identification and selection of studies.

Our electronic search identified 4,211 articles. After removing 654 duplicates, 3,557 articles were screened by title and abstract, from which articles were selected for full-text evaluation. A total of 153 articles were excluded for the following reasons: no assessment of prognostic model (n = 128), other outcomes (n = 16), other populations (n = 3), editorials (n = 3), and abstracts (n = 3). Finally, 35 studies met the eligibility criteria (Supplemental Figure 1).^8–42

Description of included studies.

The characteristics of the studies are described in Table 1. The 35 included studies presented 43 multivariable prognostic models in patients with dengue, comprising 35 developmental models and 8 external validation models. The majority of the studies used a retrospective cohort design (49%), followed by a prospective cohort design (37%). Forty-six percent of the studies included only adult patients, 31% included only pediatric patients, and 17% included both adult and pediatric patients. The proportion of women varied between 21% and 55% across studies. The countries where most studies were conducted were India (17%), Taiwan (17%), Singapore (11%), and Thailand (11%). The enrollment period for participants varied between 1994 and 2021 between studies. The most commonly used dengue diagnostic methods were serological (71%), RT-PCR (51%), and NS1 antigen (46%). Most studies (74%) included patients admitted to the hospital, whereas 11% enrolled outpatients. Only two studies enrolled patients admitted to the ICU. The median number of study participants was 383 (range: 46–2,358; Table 2). The most evaluated outcome was severity (70%), followed by mortality (22%), hospitalization (5%), and ICU admission (3%). The median number of variables included in the prognostic models was five (range: 2–14; Table 2). In almost all models, the predictors were assessed at hospital admission. In addition, 38 models used regression techniques as their modeling method, whereas the remainder used machine learning techniques.

Table 1.

Characteristics of the included studies

Author, Year	Country of Study	Study Design	Type of Population	Recruitment Period	Eligibility Criteria	Diagnostic Method for Dengue	Age	Female Sex
Fernandez, 2017	Honduras	Retrospective cohort	Adults and children	2009–2010	Patients presenting with a diagnosis of dengue fever to the health facility	Serology, viral isolation	Average 22.4 years	43%
Tamibmaniam, 2016	Malaysia	Retrospective cohort	Adults	2014	Hospitalized patients with dengue fever	Serology	≤60 years (93%)	–
Huy, 2013	Vietnam	Prospective cohort	Children	2002–2007	Patients between 6 months and 15 years who presented first episode of shock during hospitalization for dengue fever	RT-PCR, serology	Recurrent shock: 10 (0.5–15) years, single shock: 10 (0.25–15) years	55%
Pang, 2014	Singapore	Cases and paired controls	Adults	2004–2008	Adult patients hospitalized with a diagnosis of dengue fever	RT-PCR, serology	ICU: 44 (36–53) years old, non-ICU: 34 (25–44) years old	35%
Huang, 2020	Taiwan	Prospective cohort	Adults	2015	Adult patients who come to the hospital with a diagnosis of dengue fever	RT-PCR, NS1 antigen, IgM, IgG	Severe dengue: 75 (68–79) years, non-severe dengue: 55 (35–71) years	48%
Djossou, 2016	French Guyana	Prospective cohort	Adults and children	2013–2013	Patients presenting to the hospital with a diagnosis of dengue fever	RT-PCR, NS1 antigen, IgM	<1 year: 3%, 1–15 years: 36%, 15–65 years: 59%, >65 years: 2%	50%
Hsieh, 2017	Taiwan	Retrospective cohort	Adults	2015–2015	Adult patients who present hospitalized in the ICU with a diagnosis of dengue fever	RT-PCR, NS1 antigen, IgM	72.3 ± 9.3 years	39%
Hsieh, 2017	Taiwan	Retrospective cohort	Adults	2015–2015	Adult patients who present hospitalized in the ICU with a diagnosis of dengue fever	RT-PCR, NS1 antigen, IgM	–	–
Tuan, 2017	Vietnam	Prospective cohort	Children	2010–2013	Pediatric patients between 1 and 15 years of age who presented for outpatient consultation with ≤72 hours of fever with probable diagnosis of dengue fever	RT-PCR, NS1 antigen, IgM	Severe dengue: 9 (7–11), non-severe dengue: 9 (6–11)	–
Issop, 2023	France	Retrospective cohort	Adults	2019	Adult diabetic patients with a diagnosis of dengue fever admitted to the hospital	RT-PCR, NS1 antigen	70 (59.4–77.7) years old	55%
Vuong, 2021	Vietnam, Cambodia, Malaysia and El Salvador	Cases and nested controls	Children	2011–2016	Patients aged 5–15 years with fever or history of fever <72 hours and symptoms consistent with dengue fever	RT-PCR, NS1 antigen	Moderate or severe dengue: 10 (7–12), uncomplicated dengue: 10 (8–12)	45%
Vuong, 2021	Vietnam, Cambodia, Malaysia and El Salvador	Cases and nested controls	Adults	2011–2016	Patients >15 years with fever or history of fever <72 hours and symptoms consistent with dengue fever	RT-PCR, NS1 antigen	Moderate or severe dengue: 22 (18–30), uncomplicated dengue: 26 (20–34)	42%
Lam, 2017	Vietnam	Prospective cohort	Children	2001–2009	Patients aged 5–15 years with a diagnosis of dengue fever attending the hospital	RT-PCR, IgM, IgG	12 (10–13) years	41%
Lee, 2008	Singapore	Retrospective cohort	Adults	2004	Adult patients with dengue fever admitted to the hospital	RT-PCR, IgM, IgG	Severe dengue: 34 (18–54) years, non-severe dengue: 32 (17–58) years	37%
Lee, 2018	Taiwan	Retrospective cohort	Adults	2002–2015 years 2009–2013	Adult patients ≥18 years with dengue presenting to the hospital	RT-PCR, IgG, NS1 antigen	52 (range: 18–91) years	53%
Lee, 2016	Taiwan	Retrospective cohort	Adults	2002–2015 years 2009–2013	Adult patients ≥18 years with dengue presenting to hospital with time of illness ≤4 days	RT-PCR, IgG, NS1 antigen	51 (range: 18–91) years	53%
Lee, 2016	Taiwan	Retrospective cohort	Adults	2002–2015 years 2009–2013	Adult patients ≥18 years with dengue presenting to hospital with time of illness >4 days	RT-PCR, IgG, NS1 antigen	–	–
Carrasco, 2014	Singapore	Retrospective cohort	Adults	2006–2008	Adult patients presented to the hospital without a diagnosis of severe dengue fever	RT-PCR	37 (27–43) years	25%
Marois, 2021	New Caledonia	Retrospective cohort	Adults and children	2017	Female patients hospitalized with a diagnosis of dengue fever	RT-PCR	–	100%
Marois, 2021	New Caledonia	Retrospective cohort	Adults and children	2018	Female patients hospitalized with a diagnosis of dengue fever	RT-PCR	–	100%
Pang, 2016	Singapore	Prospective cohort	Adults	2005–2008	Patients ≥18 years with fever ≥38°C within 72 hours without clinically evident alternative diagnosis	RT-PCR	Hospitalization: 37 (19–77) years, non-hospitalization: 41 (21–63) years	49%
Pang, 2016	Singapore	Prospective cohort	Adults	2009–2012	Patients ≥18 years with fever ≥38°C within 72 hours without clinically evident alternative diagnosis	RT-PCR	Hospitalization: 41 (25–52) years, non-hospitalization: 33 (25–42) years	21%
Tanner, 2008	Singapore and Vietnam	Prospective cohort	Adults	2005	Adult patients (>18 years) presenting for outpatient consultation within 72 hours of onset of fever without rhinitis or clinically evident alternate diagnosis	RT-PCR	–	–
Md-Sani, 2018	Malaysia	Retrospective cohort	Adults	2014	Adult patients ≥18 years with a diagnosis of severe dengue fever	NS1, IgM, IgG antigen	30.8 (24.7–41.3) years	36%
Srisuphanunt, 2022	Thailand	Retrospective cohort	Adults and children	2017–2019	Patients diagnosed with dengue fever who attended the hospital	NS1, IgM, IgG antigen	Severe dengue: 26.7 ± 16 years, non-severe dengue: 21.8 ± 18 years	55%
Huang, 2017	Taiwan	Cases and controls	Adults	2015–2015	Patients ≥65 years with dengue who attended the hospital	NS1 antigen, IgM, IgG, epidemiological criterion, clinical criterion	74.1 ± 6.3 years	52%
Huang, 2017	Taiwan	Cases and controls	Adults	2015–2015	Adult patients who come to the hospital with a diagnosis of dengue fever	NS1 antigen, IgM, IgG, epidemiological criterion, clinical criterion	47.8 ± 21.9 years	49%
Bhaskar, 2022	India	Case cohort	Children	2016–2020	Patients from 0 to 18 years of age presenting to the hospital with a diagnosis of dengue fever	NS1 antigen, IgM	≤6 years: 20%, >6 years: 80%	47%
Gayathri, 2023	India	Retrospective cohort	Children	2019	Patients between 2 months and 12 years hospitalized due to dengue fever	NS1 antigen, IgM	6.4 ± 3.4 years	–
Gupta, 2020	India	Retrospective cohort	Children	2013–2015	Patients between 1 and 12 years of age admitted to the hospital with a diagnosis of dengue fever	NS1 antigen, IgM	Severe dengue: 69.8 ± 37.7 months, non-severe dengue: 78.4 ± 37.7 months	43%
Jain, 2017	India	Prospective cohort	Adults	2015	Adult patients hospitalized with a diagnosis of dengue shock syndrome	NS1 antigen, IgM	32 ± 16.8 years	43%
Juneja, 2011	India	Prospective cohort	Adults	2008–2010	Adult patients diagnosed with dengue admitted to the ICU	NS1 antigen, IgM	39.6 ± 17.1 year	39%
Juneja, 2011	India	Prospective cohort	Adults	2008–2010	Adult patients diagnosed with dengue admitted to the ICU	NS1 antigen, IgM	–	–
Phakhounthong, 2018	Cambodia	Retrospective cohort	Children	2009–2010	Patients <16 years with fever ≥38°C within 48 hours of admission	NS1 antigen, IgM	28 days to 1 year (28%), ≥1 year to <5 years (30%), ≥5 years to <16 years (42%)	46%
Sreenivasan, 2018	India	Prospective cohort	Children	2015–2016	Patients aged 1 month to 12 years admitted to the hospital with a serological diagnosis of dengue fever	NS1 antigen, IgM	Median 7.7 years	46%
Low, 2018	Malaysia	Prospective cohort	Adults	2016–2017	Patients ≥15 years with a diagnosis of dengue who presented within 72 hours of symptoms	NS1 antigen	Severe dengue: 34.4 ± 15.9 years, non-severe dengue: 30.2 ± 13.3 years	43%
Yang, 2023	Bangladesh	Prospective cohort	Adults and children	2019	Hospitalized patients with a serologic diagnosis of dengue fever	NS1 antigen	<18 years old (29%), 18–39 years old (51%), ≥40 (20%)	40%
Potts, 2010	Thailand	Prospective cohort	Children	1994–1997, 1999–2002 years 2004–2007	Patients aged 6 months to 15 years presenting with fever ≥38°C <72 hours without localizing symptoms for outpatient or inpatient consultation	IgM, IgG, hemagglutination and inhibition assay, viral isolation, RT-PCR	Average 8.4–9.1 year	–
Pinto, 2016	Brazil	Retrospective cohort	Adults and children	2001–2013	Patients with a diagnosis of severe dengue with laboratory confirmation	Ig M, NS1 antigen, viral isolation, RT-PCR, immunohistochemistry, immunohistochemistry	<15 years old (47%), 15–55 years old (49%), >55 years old (4%)	53%
Pongpan, 2014	Thailand	Retrospective cohort	Children	2007–2010	Hospitalized patients from 1 to 15 years of age diagnosed with dengue fever	ICD-10 codes (A90, A91 and A910)	10.3 ± 3.4 years	44%
Pongpan, 2013	Thailand	Retrospective cohort	Children	2007–2010	Hospitalized patients from 1 to 15 years of age diagnosed with dengue fever	ICD-10 codes (A90, A91 and A910)	9.6 ± 3.3 years	52%
Marois, 2021	New Caledonia	Retrospective cohort	Adults and children	2017	Male patients hospitalized with a diagnosis of dengue fever	–	–	0%
Marois, 2021	New Caledonia	Retrospective cohort	Adults and children	2018	Male patients hospitalized with a diagnosis of dengue fever	–	–	0%

ICD-10 = International Classification of Diseases, 10th Revision; ICU = intensive care unit; NS1 = non-structural protein 1; RT-PCR = reverse transcription polymerase chain reaction.

Table 2.

Characteristics of the prognostic models evaluated in the studies

Author, Year	Number of Participants	Number of Events	Type of Model	Modeling Method	Internal Validation Method	Outcome	Definition of Outcome	Number of Predictors	Predictors Included in the Final Model
Bhaskar, 2022	303	63	Development	Logistic regression	–	Severity	Severe dengue according to WHO classification 2009	5	Platelets, packed cell volume, leukocytes, alanine aminotransferase, hypotension
Carrasco, 2014	596	96	Development	Logistic regression	Bootstrapping	Severity	Severe dengue according to WHO classification 2009	8	Age, leukopenia, hematocrit, sex, duration of fever, fever on admission, vomiting, abdominal distention
Djossou, 2016	806	78	Development	Cox regression	Bootstrapping	Severity	Hypotension/shock	10	Body ache, extensive purpura, minor bleeding, rash, serous effusion, age, hematocrit, lymphocyte count, proteinemia, serum sodium
Fernandez, 2017	320	34	Development	Logistic regression	Bootstrapping	Severity	Manifestations of plasma extravasation and hemorrhagic symptoms	4	Headache, petechiae, ascites, platelets
Gayathri, 2023	312	83	Development	Logistic regression	–	Severity	Severe dengue according to WHO classification 2009	3	Pulse pressure, minor mucosal bleeding, fluid accumulation in third space
Gupta, 2020	135	89	Development	Logistic regression	–	Severity	Severe dengue according to WHO classification 2009	3	Aspartate aminotransferase, capillary refilling, lactate
Huang, 2020	798	138	Development	Artificial neural networks	k-fold cross-validation	Severity	Severe dengue according to WHO classification 2009	5	Age, sex, NS1 antigen, IgG, IgM
Huang, 2017	627	27	Development	Logistic regression	–	Mortality	Death from any cause within 30 days	4	Glasgow scale, prostration, aspartate aminotransferase, creatinine
Huang, 2017	2,358	34	Development	Logistic regression	Bootstrapping	Mortality	Death from any cause within 30 days	5	Age, systolic blood pressure, hemoptysis, diabetes, chronic prostration
Huy, 2013	444	126	Development	Logistic regression	k-fold cross-validation	Severity	Recurrent shock after first episode of shock during dengue hospitalization	5	Day of admission, purpura/ecchymosis, pleural effusion/ascites, platelets, pulse pressure
Issop, 2023	184	60	Development	Poisson regression with robust variance	–	Severity	Severe dengue according to WHO classification 2009	4	Diabetic complications, non-severe bleeding, altered mental status, coughing
Jain, 2017	46	22	Development	Logistic regression	–	Mortality	In-hospital death from any cause	3	Age, dyspnea at rest, altered sensorium
Lam, 2017	2,301	143	Development	Logistic regression	–	Severity	Severe dengue was defined as dengue shock	7	Age, sex, time of illness, vomiting, temperature, hepatomegaly, platelets, etc.
Lee, 2018	1,078	31	Development	Logistic regression	–	Mortality	All-cause mortality ≤7 days after onset of dengue and ≤3 days after hospital admission	4	Gastrointestinal bleeding ≤72 hours post-admission, platelets, leukocytes, hemoconcentration
Lee, 2016	630	37	Development	Logistic regression	Temporary splitting	Severity	Severe dengue according to WHO classification 2009	4	Age, minor gastrointestinal bleeding, white blood cells, platelets
Lee, 2016	433	18	Development	Logistic regression	Temporary splitting	Severity	Severe dengue according to WHO classification 2009	2	Age, leukocytes
Lee, 2008	1973	118	Development	Logistic regression	–	Severity	Severe dengue was defined as dengue hemorrhagic fever plus dengue shock	4	Bleeding history, total protein, urea, lymphocytes
Low, 2018	82	29	Development	Logistic regression	–	Severity	Severe dengue according to WHO classification 2009	5	Sex, VEGF, leukocytes, hematocrit, alanine aminotransferase
Marois, 2021	209	65	Development	Logistic regression	k-fold cross-validation	Severity	Severe dengue according to WHO classification 2009	6	Age, hypertension, rash, mucosal bleeding, platelets, alanine aminotransferase
Marois, 2021	174	65	Development	Logistic regression	k-fold cross-validation	Severity	Severe dengue according to WHO classification 2009	5	Age, alcohol consumption, mucosal bleeding, platelets, alanine aminotransferase
Md-Sani, 2018	199	20	Development	Logistic regression	k-fold cross-validation	Mortality	Mortality from any cause during hospitalization	4	Age, sex, serum bicarbonate, alanine aminotransferase
Pang, 2016	92	47	Development	Logistic regression	–	Hospitalization	Hospitalization according to attending physician’s decision + alarm signs	3	CCL8, VPS13C, platelets
Pang, 2014	135	27	Development	Conditional logistic regression	–	Admission to ICU	Admission to ICU during hospitalization	3	Neutrophils, alanine aminotransferase, urea
Phakhounthong, 2018	198	38	Development	CART Analysis	k-fold cross-validation	Severity	Severe dengue according to WHO classification 2009	5	Hematocrit, Glasgow scale, protein in urine, creatinine, platelets
Pinto, 2016	1605	61	Development	Logistic regression	–	Mortality	Dengue-related mortality according to monitoring system records	4	Age, hematuria, gastrointestinal bleeding, platelets
Pongpan, 2013	777	90	Development	Ordinal logistic regression	–	Severity	Dengue shock syndrome according to WHO 1997 classification	6	Age, hepatomegaly, hematocrit, systolic blood pressure, leukocytes, platelets
Potts, 2010	1,230	37	Development	CART Analysis	k-fold cross-validation	Severity	Dengue shock syndrome according to WHO 1997 classification	4	Leukocytes, monocytes, platelets, hematocrit
Sreenivasan, 2018	359	93	Development	Logistic regression	–	Severity	Severe dengue according to WHO classification 2009	3	Clinical accumulation of fluids, packed cell volume/platelets, persistent vomiting
Srisuphanunt, 2022	302	195	Development	Ordinal logistic regression	–	Severity	Severe dengue according to WHO classification 2009	6	Albumin, aspartate aminotransferase, alanine aminotransferase, platelets, partial thromboplastin time, IgM positive
Tamibmaniam, 2016	657	44	Development	Logistic regression	–	Severity	Severe dengue according to WHO classification 2009	3	Vomiting, pleural effusion, systolic blood pressure
Tanner, 2008	161	23	Development	C4.5 decision tree classifier	k-fold cross-validation	Severity	Dengue hemorrhagic fever according to WHO 1997	3	Platelets, Ct, IgG positive
Tuan, 2017	2,060	117	Development	Logistic regression	k-fold cross-validation	Severity	Severe dengue according to WHO classification 2009	4	Vomiting, platelets, aspartate aminotransferase, NS1 antigen
Vuong, 2021	464	127	Development	Logistic regression	Bootstrapping	Severity	Moderate or severe dengue according to recommended classification for dengue clinical trials	6	IL-1RA, Ang-2, IL-8, ferritin, IP-10, SDC-1
Vuong, 2021	373	154	Development	Logistic regression	Bootstrapping	Severity	Moderate or severe dengue according to recommended classification for dengue clinical trials	7	SDC-1, IL-8, ferritin, sTREM-1, IL-1RA, IP-10, sCD163
Yang, 2023	1,090	158	Development	Random forest	–	Severity	Severe dengue according to WHO classification 2009	4	Age, educational level, plasma leakage, platelets, dyspnea, dyspnea, etc.
Hsieh, 2017	75	31	External validation	Cox regression	–	Mortality	In-hospital death from any cause	14	Age, temperature, mean arterial pressure, heart rate, respiratory rate, PaO2/FiO2, PHa, Na+, K+, creatinine, hematocrit, leukocytes, Glasgow scale, chronic disease
Hsieh, 2017	75	31	External validation	Cox regression	–	Mortality	In-hospital death from any cause	8	PaO2/FiO2, platelets, bilirubin, Glasgow scale, mean arterial blood pressure, vasopressor use, creatinine, diuresis
Juneja, 2011	198	12	External validation	Logistic regression	–	Mortality	Death from any cause 28 days after discharge from ICU	14	Age, temperature, mean arterial pressure, heart rate, respiratory rate, PaO2/FiO2, PHa, Na+, K+, creatinine, hematocrit, leukocytes, Glasgow scale, chronic disease
Juneja, 2011	198	12	External validation	Logistic regression	–	Mortality	Death from any cause 28 days after discharge from ICU	8	PaO2/FiO2, platelets, bilirubin, Glasgow scale, mean arterial blood pressure, vasopressor use, creatinine, diuresis
Marois, 2021	66	15	External validation	Logistic regression	–	Severity	Severe dengue according to WHO classification 2009	6	Age, hypertension, rash, mucosal bleeding, platelets, alanine aminotransferase
Marois, 2021	64	19	External validation	Logistic regression	–	Severity	Severe dengue according to WHO classification 2009	5	Age, alcohol consumption, mucosal bleeding, platelets, alanine aminotransferase
Pang, 2016	80	25	External validation	Logistic regression	–	Hospitalization	Hospitalization according to treating physician’s decision + alarm signs	3	CCL8, VPS13C, platelets
Pongpan, 2014	400	35	External validation	Ordinal logistic regression	–	Severity	Dengue shock syndrome according to WHO 1997 classification	5	Age, hepatomegaly, systolic blood pressure, white blood cells, platelets

Ang-2 = angiopoietin-2; CART = classification and regression trees; CCL8 = chemokine (C-C motif) ligand 8; Ct = cycle threshold; FiO₂ = fraction of inspired oxygen; ICU = intensive care unit; IL-1RA = interleukin-1 receptor antagonist; IL-8 = Interleukin-8; IP-10 = interferon gamma-induced protein-10; K+ = serum potassium; Na+ = serum sodium; NS1 = non-structural protein 1; PaO₂ = arterial partial pressure of oxygen; PHa = arterial PH; sCD163 = soluble cluster of differentiation 163; SDC-1 = syndecan-1; sTREM-1 = soluble triggering receptor expressed on myeloid cells-1; VEGF = vascular endothelial growth factor; VPS13C = vacuolar protein sorting-associated protein 13.

Assessment of the risk of bias and applicability.

All prognostic models were judged to be at high risk of bias in the overall assessment, primarily due to the outcome and analysis domains (Supplemental Figure 2). Within the outcome domain, the main reason for the high risk of bias was that in most models (58%), predictors were included in the outcome definition. In addition, it was not adequately reported whether the outcome was determined without knowledge of the predictor information. In the analysis domain, several important factors contributed to the high risk of bias: an insufficient number of participants with the outcome, inappropriate handling of continuous and categorical predictors, a lack of consideration of data complexity in the analysis, and failure to account for overfitting, underfitting, and optimism in model performance. In addition, most models did not provide information on the weights assigned in the final model and their correspondence with the reported multivariable analysis. In terms of applicability, the majority of models (77%) were considered to be of low concern. However, in 21% of models, applicability was judged as a high concern, primarily because of the predictor domain because some biomarkers not routinely available in clinical practice were included.

Description of prognostic models.

Severity.

Dengue severity was evaluated in 25 studies, corresponding to 30 prognostic models. Of these, 27 were new development models and 3 were external validation models. In 19 models, severity was defined according to the 2009 WHO classification. The predictors most frequently included in the models were platelet count (n = 17 models), age (n = 13 models), presence of minor or major bleeding (n = 9 models), blood leukocyte level (n = 8 models), and presence of serous effusion (pleural effusion or ascites; n = 7 models; Supplemental Figure 3A; Supplemental Table 3). Regarding the performance of the models, the discriminative ability was evaluated in 19 models, with the C-statistic values varying between 0.70 and 0.95 (Supplemental Table 4). The evaluation of model calibration was poorly reported in the studies; only two studies provided the calibration plots, calibration slope, and Hosmer–Lemeshow test. In terms of classification measures, sensitivity and specificity were reported in 17 models, with values ranging from 60% to 100% and 29% to 98%, respectively (Supplemental Figure 4A).

Mortality.

Mortality was evaluated in eight studies, corresponding to ten prognostic models, of which six were new developments and four were external validations. In five models, mortality was defined as in-hospital, whereas in the remaining models, the follow-up time extended beyond hospitalization, with a maximum of 30 days. The four external validation models evaluated acute physiology and chronic health disease classification system II and sequential organ failure assessment scores. The predictors most commonly included in the models were age (n = 6 models), Glasgow scale (n = 6 models), blood pressure (n = 5 models), serum creatinine (n = 5 models), and platelet count (n = 4 models; Supplemental Figure 3B; Supplemental Table 3). Regarding the discriminative performance of the models, the C-statistic ranged from 0.83 to 0.99 (Supplemental Table 4). One study reported the Hosmer–Lemeshow test as a calibration measure. Meanwhile, sensitivity and specificity were reported in only four studies, ranging from 92% to 95% and 69% to 88%, respectively (Supplemental Figure 4B).

Hospitalization.

Hospitalization was evaluated in only one study, which reported two models (one developmental and one for external validation). These models included the following predictors: platelet count and serum levels of the biomarkers chemokine (C-C motif) ligand 8 and vacuolar protein sorting-associated protein 13. In terms of discriminative ability, the developmental model yielded a C-statistic of 0.87, and no calibration measure was reported (Supplemental Table 4). The sensitivity and specificity were 81% and 84% for the developmental model and 60% and 78% for the validation model.

Intensive care unit admission.

Intensive care unit admission was evaluated in only one study that included one developmental model. The predictors considered were the percentage of neutrophils, serum alanine transpeptidase level, and serum urea. The C-statistic was 0.92, and the Hosmer–Lemeshow test result was 0.52 (Supplemental Table 4). In addition, the sensitivity was 88%, and the specificity was 89%.

DISCUSSION

In this systematic review of predictive prognostic models in patients with dengue, 43 models described in 35 studies were identified and critically evaluated. The prognostic models can be divided into those predicting dengue severity (n = 30), mortality (n = 10), hospitalization (n = 2), and ICU admission (n = 1). Most studies reported moderate to good predictive performance; however, none of the models had a low risk of bias due to a combination of poor reporting and poor methodological quality. In addition, both the sample sizes and the number of events were limited. This is a common problem when building prediction models, resulting in an increased risk of model overfitting.⁴³ Other common causes of bias included the inclusion of some predictors in the definition of outcomes, the use of techniques that do not account for optimism in performance estimates, and a lack of model calibration. A high risk of bias implies that the performance of these models in new samples is likely to be worse than that reported in the primary studies.⁷ Therefore, the estimated C-statistics, which often indicated good discrimination, are probably optimistic. In addition, most models were not independently externally validated, which limits their applicability in different clinical settings.

The implementation of prognostic models in patients with dengue is crucial for the proper management of the disease.^1,2,44 These models, based on clinical and laboratory parameters, allow for the risk stratification of each individual, aiding in decision-making for triage and the application of early interventions, especially for the most vulnerable groups.^44,45 However, it is important to consider that the presentation and progression of dengue can be heterogeneous, encompassing a wide spectrum of severity, from mild, self-limited disease to severe forms such as dengue shock syndrome.⁴⁶ Another aspect to consider is the dynamic nature of dengue outbreaks and epidemics, which are constantly changing due to factors such as climate change, urbanization, the availability of diagnostic tests, and vector control measures.⁴⁷ Consequently, predictive models must account for these variations, incorporating real-time data and adapting to the specific characteristics of different populations and geographic regions. The availability and reliability of clinical information pose significant challenges in the development and validation of prognostic models for patients with dengue. Several models have been based on retrospective data,^11,12,14 which carries the inherent risk of measurement errors and missing data. Therefore, prospective studies with standardized data collection protocols are essential to overcome these limitations and ensure adequate generalizability and performance. In this regard, collaboration within a region is important to obtain appropriate databases for model development and validation.

Recently, a systematic review of prognostic factors for progression to severe dengue was published.⁴⁸ We found that extremes of age, female sex, comorbidities such as hypertension, diabetes, renal disease, and cardiovascular disease, as well as the presence of vomiting, abdominal pain, bleeding, clinical fluid accumulation, secondary infection, low platelet count, low serum albumin, and high serum alanine aminotransferase and aspartate aminotransferase levels measured during the febrile phase of illness, were the most important factors for the early prediction of severe dengue. Similarly, our review identified that several of these predictors were the most frequently included in prognostic models for severity. Therefore, we suggest that incorporating these features be prioritized for potential use in clinical practice,⁴⁵ considering that they are readily available in both high- and low-resource areas.

Our systematic review identified several sources of bias in dengue prognostic models that limit their reliability and applicability.⁴⁹ A critical issue was the inclusion of predictors that overlapped with outcome definitions, particularly in models using variables such as platelet counts, which are also part of severity classifications. However, our analysis revealed that most studies did not provide sufficient detail to determine if the temporal relationship between predictors and outcomes was explicitly considered during model development. It was often unclear whether the predictors were measured before the outcome occurred or whether they were part of the same assessment period. To address this lack of clarity, future models should define predictors and outcomes to ensure temporal and conceptual independence. Small sample sizes and limited events per predictor variable were common and often resulted in overfitting, especially in models employing advanced techniques such as machine learning. Adhering to a minimum ratio of 10 events per predictor variable may help reduce this risk.⁵⁰ Furthermore, the transparent reporting of predictor selection, handling, and model performance is critical. Although most dengue models reported good discrimination, calibration measures were frequently missing, thereby limiting the assessment of their clinical utility. Researchers should adopt standardized reporting frameworks like transparent reporting of multivariable prediction models for individual prognosis or diagnosis to improve the transparency and comparability of studies.⁵

The lack of external validation represents another significant limitation. External validation in diverse populations, including under-resourced settings, is crucial for assessing the generalizability of dengue prognostic models.⁵¹ Internal validation methods, such as cross-validation or bootstrapping, should complement, but not replace, independent validation efforts. Additionally, handling missing data remains a challenge in dengue research. Multiple imputation techniques could be used to manage missing predictors effectively and preserve dataset integrity.⁴⁹ By addressing these issues, future studies can reduce bias and enhance the quality and applicability of prognostic models for dengue, ensuring they provide reliable tools for improving patient outcomes in endemic regions.

The identified prognostic models for predicting dengue severity offer potential for clinical application, particularly in settings in which early risk stratification can significantly impact outcomes. Models that rely on widely available clinical and laboratory predictors, such as platelet count, hematocrit, and signs of plasma leakage, are particularly suited for use in resource-limited settings. For example, the “Early Severe Dengue Identifier” serves as a practical tool with good discrimination and calibration for identifying patients at risk of severe disease.⁴⁰ More complex models that incorporate advanced biomarkers or machine learning techniques may offer enhanced predictive accuracy but require further external validation before routine implementation. These models should be prioritized for use in settings with the necessary resources and laboratory infrastructure. To ensure reliability and clinical relevance, selected models should undergo external validation and calibration to the local population.⁵¹ Implementation efforts must also address training for healthcare providers and integration into existing clinical workflows. Future research should focus on refining and adapting these models to diverse settings, maximizing their utility in improving outcomes for patients with dengue.

CONCLUSION

Our systematic review identified a significant number of prognostic models for predicting severity and mortality in patients with dengue in both adult and pediatric populations. However, although most models demonstrated good discriminatory ability, calibration measures were rarely reported, and most of the models were developmental. In addition, all models exhibited a high risk of bias, suggesting that their performance may be overestimated. Therefore, further prospective cohort studies are required to perform external validation (e.g., in different populations or geographical regions) and possibly update the identified models using appropriate methodology before their application in clinical practice.

Note: Supplemental materials appear at www.ajtmh.org.