Multicenter validation of a simple clinical risk score for early prediction of severe dengue among adults in Thailand

Abstract

Background

Dengue remains a major public health challenge in Thailand. Although only a minority of patients progress to severe dengue, the condition is potentially life-threatening and requires early recognition. Simple clinical prediction tools may support early risk stratification in routine care, particularly in resource-limited settings. This study aimed to externally validate a simple clinical risk score for predicting severe dengue among adult patients using multicenter hospital data in Thailand.

Methods

A retrospective multicenter validation study was conducted using clinical data from six hospitals in Thailand (2019–2022). Model discrimination was assessed using the area under the receiver operating characteristic curve (AuROC). Patients were categorized into three predefined risk groups. Likelihood ratios were calculated to evaluate clinical utility across risk strata. Calibration was assessed using calibration plots, slope, and intercept.

Results

The model demonstrated good discriminative performance, with an AuROC of 0.83 (95% CI 0.77–0.90) in the validation cohort, compared with 0.88 in the development cohort (p-value = 0.249). Likelihood ratios indicated strong rule-in ability in the high-risk group (LR + = 52.02), while the low-risk group was associated with reduced probability of severe disease. Calibration was acceptable overall, although systematic underestimation was observed in higher-risk patients.

Conclusions

This simple clinical risk score demonstrated good and stable performance across diverse hospital settings. The model may support early in-hospital risk stratification and clinical monitoring decisions among adult dengue patients. Further prospective validation and recalibration studies are warranted.

Clinical trial number

Not applicable.

Background

Dengue remains one of the most significant mosquito-borne viral diseases worldwide, with an estimated 100–400 million infections occurring annually. The burden is particularly concentrated in tropical and subtropical regions, where rapid urbanization, climate variability, and high vector density sustain transmission. Southeast Asia continues to report among the highest incidence rates globally, and Thailand is one of the most affected countries. National surveillance data indicate a marked resurgence of dengue in Thailand in 2023, followed by persistently high transmission in early 2024, placing substantial pressure on healthcare services during epidemic periods [1].

Traditionally regarded as a predominantly pediatric disease in Asia, dengue has undergone a notable epidemiological shift in Thailand, with adults now comprising an increasing proportion of hospitalized cases. Evidence suggests that adult patients may experience more severe plasma leakage, organ involvement, and higher mortality compared with children [2, 3]. This shift has important clinical implications, as adults often present with comorbidities, atypical symptoms, or delayed healthcare-seeking behavior, complicating early recognition of severe disease. Strengthening risk stratification for adult dengue patients has therefore become an important priority for health systems in endemic settings.

Although only a small proportion of dengue infections progress to severe disease, complications such as shock, respiratory distress, and multiorgan failure can be rapidly fatal if not recognized early [4]. Timely identification of patients at risk of clinical deterioration is essential for guiding monitoring intensity, clinical management, and referral decisions. While the 2009 World Health Organization dengue classification introduced warning signs to support early detection, studies have reported variable predictive accuracy across populations and clinical contexts [4, 5]. This variability suggests that existing classification approaches may be insufficient for consistent individual-level risk stratification, particularly in high-volume or resource-constrained settings.

To address these limitations, numerous clinical prediction rules and risk-scoring models have been proposed. However, many rely on multiple laboratory tests or specialized assessments that may not be readily available in routine practice. Moreover, despite favorable performance in development studies, relatively few models have undergone validation beyond their original development cohorts across diverse hospital settings, especially among adult populations. This raises concerns regarding model transportability and real-world applicability, as predictive performance often declines when tools are applied outside their original development context [6, 7].

A simplified clinical severity score for adult dengue was recently developed in Thailand using four routinely available clinical parameters [8]. The model demonstrated good discriminative performance in the development cohort, suggesting potential utility for early risk stratification in routine clinical care. However, its performance has not yet been validated across diverse hospital settings and epidemic periods within Thailand.

This study aimed to validate the performance of this simple clinical risk score for predicting severe dengue among adults using multicenter retrospective cohort data from hospitals in Thailand.

Methods

Study design and setting

A retrospective multicenter cohort validation study was conducted to evaluate the performance of a previously developed dengue severity risk scoring system among adult patients. The validation cohort comprised patients from six hospitals in Thailand (Lopburi Hospital, Samut Sakhon Hospital, Roi-et Hospital, Buriram Hospital, Krabi Hospital, and Songkhla Hospital) during the period from 2019 to 2022.

Study participants

Eligible participants were adult patients aged > 15 years with laboratory-confirmed dengue infection, diagnosed by NS1 antigen testing, reverse transcription–polymerase chain reaction (RT–PCR), or dengue-specific IgM or IgG antibody assays. Patients were identified using dengue-related ICD-10 codes (A90, A91, A91.0, A97.0, A97.1, and A97.2).

All eligible adult dengue patients identified from hospital records using dengue-related ICD-10 codes during the study period constituted the source cohort. From this source cohort, patients were selected using computer-generated random sampling across participating centers. Sampling was not based on outcome status and was not conducted using a case-control design.

This study represents geographic external validation, as the development and validation datasets were collected during the same calendar period but derived from different hospitals, thereby representing distinct patient populations defined a priori rather than a random split of a single dataset.

Therefore, this validation should be interpreted as an assessment of geographic transportability across different hospitals and healthcare settings in Thailand, rather than temporal transportability across different epidemic periods.

The age threshold of > 15 years is consistent with clinical practice definitions for adult dengue patients in Thailand.

Predictor variables included in the risk score were extracted from the initial clinical assessment recorded at first hospital presentation, prior to formal classification of severe dengue. Severe dengue status was determined through retrospective review of clinical records using standardized criteria based on the World Health Organization 2009 classification.

Outcome definition

Patients were classified into two outcome groups according to a modified version of the World Health Organization (WHO) 2009 dengue classification [4].

Severe dengue was defined by the presence of at least one of the following: (1) severe plasma leakage leading to shock or respiratory distress; or (2) severe organ involvement, including pulmonary edema, disseminated intravascular coagulation, encephalopathy, liver failure, or renal failure.

Non-severe dengue was defined as dengue infection not meeting any of the above criteria.

Outcome classification was determined through retrospective review of clinical records using standardized criteria applied uniformly across all participating centers to minimize outcome misclassification and ensure consistency of classification. All outcome assessments were based on clinical information documented during the hospital course, independent of the predictor variables extracted at initial presentation.

Sample size determination

The sample size required for model validation was estimated based on the expected discriminative performance of the model, expressed as the area under the receiver operating characteristic curve (AuROC). Based on previous findings reported by Lee et al. [9], an expected AuROC of 0.80 was assumed, with a one-sided α of 0.05 and 80% power. An event-to-non-event ratio of approximately 1:3 was used to inform the sample size estimation; however, patient selection was not based on outcome status. Instead, eligible patients were randomly sampled from the source cohort of adult dengue patients across participating centers using computer-generated random sampling. A total of 351 patients were included in the validation cohort, comprising 68 severe and 283 non-severe dengue cases. Although the number of severe cases was slightly lower than the initially estimated target (n = 83), the number of outcome events remained sufficient for model validation. With four predictors included in the risk score, the events per variable (EPV) was 17, exceeding the commonly recommended minimum threshold of 10 and supporting the adequacy of the validation sample size [10, 11]. This was further supported by the number of outcome events available for validation, which was considered adequate according to current methodological recommendations for external validation studies [10].

Data collection

Clinical data were retrospectively extracted from hospital medical records using standardized data collection forms. Variables included demographic characteristics, underlying comorbidities, clinical manifestations, hemodynamic parameters, and laboratory findings. Quantitative variables were recorded as continuous measures and, where appropriate, categorized using clinically relevant thresholds consistent with standard clinical practice and the original model development.

Pleural effusion was defined based on clinical examination or radiographic findings documented at the initial patient assessment, prior to formal classification of severe dengue.

The development dataset consisted of 684 adult dengue patients from nine hospitals in the previously published derivation study, whereas the present validation cohort comprised 351 adult patients from six different hospitals. Although both datasets were collected during the same calendar period (2019–2022), they represented distinct cohorts defined a priori and were not created by randomly splitting a single pooled dataset.

All variables included in the risk score were extracted from information available at first hospital presentation to reflect the intended clinical use of the model in early in-hospital risk stratification.

Risk score and statistical analysis

Risk scores were calculated according to the previously developed scoring scheme. Continuous predictors were categorized using predefined clinically relevant cut-off values consistent with the original model. Specifically, age was categorized as < 55 and ≥ 55 years, and platelet count as < 100,000/µL and ≥ 100,000/µL (Table 1). Patients were classified into three predefined risk groups based on the total score: low risk (≤ 2.0 points), moderate risk (2.5–5.0 points), and high risk (≥ 5.5 points).

Table 1.

Risk score components for classification of severe dengue

Predictors	Category	Score
Age < 55 years	yes	1.5
	no	0
Severe bleeding*	yes	2.5
	no	0
Pleural effusion	yes	2.0
	no	0
Platelet count (/µL)	< 100,000	1.0
	≥ 100,000	0

Notes: *Bleeding from the gastrointestinal tract, hematemesis or melena, menorrhagia, or hematuria

Baseline characteristics were compared between the development and validation cohorts. Categorical variables were compared using Fisher’s exact test, while continuous variables were compared using the independent t-test for approximately normally distributed data and the Mann–Whitney U test for skewed data.

Model discrimination was assessed using the area under the receiver operating characteristic curve (AuROC). To assess the potential effect of incorporation bias, an exploratory sensitivity analysis was performed by recalculating model discrimination after excluding severe bleeding and pleural effusion from the score. The reduced score retained only age < 55 years and platelet count < 100,000/µL. Discrimination of the reduced score was evaluated using the same binormal ROC approach applied in the primary analysis. As an exploratory assessment of clinical utility, model performance was evaluated at a clinically plausible predicted probability threshold of 10% for identifying patients who may require closer monitoring or escalation of care. Estimated net benefit was compared with treat-all and treat-none strategies.

Differences in AuROC between the development and validation cohorts were formally compared using a nonparametric approach.

Calibration was assessed by comparing predicted probabilities with observed event rates across grouped levels of predicted risk [12]. Predicted probabilities were categorized into quintiles, and calibration plots were constructed by plotting the mean predicted probability against the corresponding observed event rate within each group. In addition, calibration-in-the-large (intercept) and calibration slope were estimated by fitting a logistic regression model with the observed outcome as the dependent variable and the logit of predicted probabilities as the independent variable. A calibration slope close to 1 and an intercept close to 0 indicate good agreement between predicted and observed risks.

Missing data were assessed descriptively. Missingness was low (< 5%) for core variables included in the final model, including platelet count (3.9%), white blood cell count (4.0%), hematocrit (4.0%), and hemoglobin (4.0%), with missingness observed primarily in AST, ALT, and hematocrit variables. Given the small proportion of missing data, complete-case analysis was applied.

Statistical analyses were performed using Stata version 15.1 (StataCorp, College Station, TX, USA). All tests were two-sided, and a p-value < 0.050 was considered statistically significant. This study was conducted and reported in accordance with the TRIPOD guidelines [13].

Results

Validation cohort characteristics

The validation cohort comprised 351 adult patients with dengue recruited from six hospitals in Thailand. Overall, baseline demographic characteristics, clinical manifestations, hemodynamic parameters, and laboratory findings were broadly comparable between the development cohort (n = 684) and the validation cohort (n = 351). Some statistically significant differences were observed in sex distribution, age, systolic and diastolic blood pressure, hematocrit, and aspartate aminotransferase levels. However, the majority of clinical and laboratory parameters were similar between the two cohorts (Table 2). The proportion of missing data was low (< 5% for all variables), and complete-case analysis was therefore applied. These findings suggest that, despite some differences, the validation cohort was broadly representative of the population used in the model development.

Table 2.

Clinical characteristics of adult dengue patients in the development and validation datasets

Patient characteristics	Development (n = 684)	Validation (n = 351)	p-value
Demographic
Female, n (%)	314 (45.9)	192 (54.7)	0.009
Age (years), Mean ± SD	28.6 ± 12.8	23.4 ± 5.5	0.001
BMI (kg/m2), Mean ± SD	23.2 ± 5.1	23.3 ± 5.1	0.585
Underlying disease, n (%)	50 (7.3)	36 (10.3)	0.122
Hypertension	30 (4.4)	24 (6.8)	0.105
Diabetes Mellitus	22 (3.2)	13 (3.7)	0.718
Dyslipidemia	17 (2.5)	12 (3.4)	0.428
Asthma	8 (1.2)	5 (1.4)	0.771
Clinical presentation, n (%)
fever	684 (100)	351 (100)	-
Headache	444 (64.9)	248 (70.7)	0.070
Myalgia	487 (71.2)	265 (75.5)	0.162
Retro-orbital pain	92 (13.5)	35 (10.0)	0.111
Bone pain	12 (1.8)	4 (1.1)	0.598
Joint pain	45 (6.6)	32 (9.1)	0.168
Abdominal pain	144 (21.1)	64 (18.2)	0.326
Vomiting	311 (45.5)	158 (45.0)	0.895
Cough	162 (23.7)	100 (28.5)	0.097
Diarrhea	141 (20.6)	73 (20.8)	0.936
Petechiae	106 (15.5)	49 (14.0)	0.581
Rash	48 (7.0)	33 (9.4)	0.180
Epistaxis	30 (4.4)	17 (4.8)	0.754
Bleeding from gums	64 (9.4)	21 (6.0)	0.072
Severe bleeding*	69 (10.0)	37 (10.5)	0.829
Hepatomegaly	9 (1.3)	6 (1.7)	0.594
Pleural effusion	18 (2.6)	4 (1.1)	0.170
Rapid weak pulse	40 (5.9)	27 (7.7)	0.286
Hemodynamic, Mean ± SD
Systolic blood pressure (mmHg)	110.8 ± 15.9	108.1 ± 17.8	0.013
Diastolic blood pressure (mmHg)	69.5 ± 12.2	67.1 ± 12.7	0.003
Pulse pressure (mmHg)	41.3 ± 10.7	40.9 ± 12.4	0.673
Hematological
Hematocrit (%), Mean ± SD	42.3 ± 5.8	41.5 ± 5.2	0.025
White blood cell (/µL), Median (IQR)	3520 (2600–5045)	3500 (2670–4800)	0.930
Platelet count (/µL), Median (IQR)	94 (58–140)	90 (64–127)	0.494
Neutrophils (%), Mean ± SD	60.9 ± 17.3	59.8 ± 17.7	0.345
Lymphocyte (%), Median (IQR)	25 (17–37)	25 (17–36.6)	0.935
Biochemical
AST (U/L), Median (IQR)	86 (45–155)	74 (43–131)	0.042
ALT (U/L), Median (IQR)	48.5 (27–99)	43.5 (25–86)	0.064
Duration of fever (days), Mean ± SD	3.9 ± 1.4	3.8 ± 1.3	0.158
Duration of admission (days), Median (IQR)	3 (2–4)	3 (2–4)	0.878
In-hospital mortality, n (%)	6 (0.9)	7 (2.0)	0.111

Notes: *Bleeding from the gastrointestinal tract, hematemesis or melena, menorrhagia, or hematuria

Distribution of risk categories

Patients in the validation cohort were classified into three predefined risk categories based on total risk scores: low risk (≤ 2.0 points), moderate risk (2.5–5.0 points), and high risk (≥ 5.5 points). The positive likelihood ratios (LR+) for severe dengue were 0.30 (95% CI 0.17–0.55) in the low-risk group, 0.95 (95% CI 0.73–1.25) in the moderate-risk group, and 52.02 (95% CI 12.63–214.31) in the high-risk group (Table 3). The low-risk group was associated with a substantially reduced probability of severe dengue, whereas the high-risk group demonstrated strong rule-in performance. In contrast, the moderate-risk group included 177 of 351 patients (50.4%) and had an LR+ close to 1.0, indicating minimal change in post-test probability and limited discriminatory value for more than half of the validation cohort. These findings support the use of the score as a risk stratification tool to identify low- and high-risk patients and to guide clinical triage and monitoring intensity, while patients in the moderate-risk category require continued clinical assessment and close monitoring.

Table 3.

Distribution of severe and non-severe patients according to risk categories, likelihood ratios, and 95% confidence intervals

Risk categories	Score	severe (n = 68)		non-severe (n = 283)		LR+	95%CI	p-value
		n	%	n	%
Low	≤ 2	10	14.7	137	48.4	0.30	0.17–0.55	< 0.001
Moderate	2.5–5	33	48.5	144	50.9	0.95	0.73–1.25	0.727
High	≥ 5.5	25	36.8	2	0.7	52.02	12.63–214.31	< 0.001

Model discrimination

The scoring system demonstrated acceptable discriminative performance in both the development and validation cohorts. The area under the receiver operating characteristic curve (AuROC) was 0.88 (95% CI 0.84–0.92) in the development cohort and 0.83 (95% CI 0.77–0.90) in the validation cohort. Although the difference was not statistically significant (p-value = 0.249), the decline in AuROC from 0.88 to 0.83 may still be clinically relevant and should be interpreted cautiously (Table 4; Fig. 1).

Table 4.

Score-classified severity levels and model performance in the development and validation cohorts

Score classification	Development (n = 684)	Validation (n = 351)	p-value
Mean score ± SD	2.8 ± 1.5	2.8 ± 1.4	0.662
Risk categories, n (%)
Low risk	309 (45.2)	147 (41.9)	0.493*
Moderate risk	318 (46.5)	177 (50.4)
High risk	57 (8.3)	27 (7.7)
AuROC (95%CI)	0.88 (0.84–0.92)	0.83 (0.77–0.90)	0.249

Notes:* Exact test comparing the distribution of patients across risk categories between cohorts

Abbreviations: SD, standard deviation; AuROC, area under the receiver operating characteristic curve

Fig. 1.

Receiver operating characteristic (ROC) curves of the dengue severity risk score in the development and validation cohorts. (A) Development cohort (n = 684). (B) Validation cohort (n = 351)

Sensitivity analysis for incorporation bias

To explore the potential effect of incorporation bias, an exploratory sensitivity analysis was performed excluding severe bleeding and pleural effusion from the score.

The reduced score, which retained only age < 55 years and platelet count < 100,000/µL, demonstrated substantially lower discrimination than the original score. Using the same binormal ROC approach applied in the primary analysis, the reduced score yielded an ROC area of 0.67 (95% CI 0.57–0.78), compared with approximately 0.83 for the original score.

These findings suggest that severe bleeding and pleural effusion contributed substantially to the discriminative performance of the original model.

Exploratory clinical utility assessment

In the exploratory clinical utility assessment using a 10% predicted probability threshold, the model identified 58 of 68 severe dengue cases (85.3%). The estimated net benefit of the model was 0.119, compared with 0.104 for a treat-all strategy and 0 for a treat-none strategy. These findings suggest potential clinical utility for early in-hospital risk stratification.

Model calibration

Calibration in the validation cohort was assessed by comparing predicted probabilities with observed event rates. The calibration slope was 0.90 (95% CI 0.66–1.15), and the calibration intercept was 0.38 (95% CI 0.07–0.70).

The positive calibration intercept indicated systematic underestimation of severe dengue risk in the validation cohort. The calibration plot showed that underestimation was most evident in the highest-risk group, where the observed event rate exceeded the predicted probability (Fig. 2). Given this pattern, recalibration-in-the-large may be required before broader implementation in settings where systematic underestimation of severe dengue risk is observed. This pattern may reflect differences in case-mix between the development and validation cohorts.

Fig. 2.

Calibration plot of predicted versus observed probabilities of severe dengue in the validation cohort

Clinical interpretation / summary of model performance

Overall, the model demonstrated good and stable performance in the external validation cohort. Discrimination remained acceptable, with an AuROC of 0.83, and calibration showed reasonable agreement between predicted and observed risks, although some underestimation was observed in higher-risk patients.

Risk stratification analysis further supported the clinical utility of the model. The high-risk group demonstrated strong rule-in performance, while the low-risk group was associated with a reduced probability of severe dengue. In contrast, the moderate-risk group showed limited discriminatory value, reflecting a clinically relevant intermediate-risk category in which further monitoring and clinical judgment are required. Taken together, these findings suggest that the model may be useful as a risk stratification tool to support early in-hospital clinical decision-making, rather than as a definitive diagnostic classifier. This is particularly relevant in resource-limited settings, where simple clinical tools may assist in prioritizing monitoring and resource allocation.

Discussion

In this multicenter external validation study, the simple dengue severity risk score demonstrated acceptable discriminative performance when applied to adult patients across six hospitals in Thailand. Although the AuROC was modestly lower in the validation cohort than in the development cohort, the difference was not statistically significant, which is consistent with the expected decline in performance when prediction models are applied to new populations. These findings support the overall stability and transportability of the model across different clinical settings. Nevertheless, the reduction in AuROC from 0.88 to 0.83, although not statistically significant, may still be clinically relevant, and differences in baseline characteristics between the derivation and validation cohorts may partly explain this decline. In particular, differences in age, sex distribution, and systolic blood pressure may reflect case-mix variation between cohorts and may have contributed to differences in calibration and discrimination.

Risk stratification into low-, moderate-, and high-risk categories provided clinically meaningful information for early triage. The high-risk group showed a markedly elevated likelihood of severe dengue, supporting its potential use for identifying patients who may benefit from intensified monitoring or earlier escalation of care. In contrast, the low-risk group was associated with a reduced likelihood of severe outcomes and may support standard monitoring with lower immediate concern for progression. The moderate-risk group accounted for approximately half of the validation cohort and demonstrated a likelihood ratio close to 1.0, indicating limited discriminatory contribution and minimal change in post-test probability. This intermediate category therefore represents a clinically important “grey zone,” in which the score should not be used alone for rule-in or rule-out decisions. Serial assessment, close monitoring, and clinical judgment remain essential for these patients. In practice, the score may be particularly useful for structuring and standardizing risk assessment across clinicians, especially in high-volume or resource-limited settings.

The discriminative performance observed in this study is consistent with previous reports of parsimonious dengue prediction models, which have achieved AuROC values in the mid-to-high 0.80s [14–17]. These findings reinforce the value of simple models based on routinely available clinical variables. In line with the derivation study [8], the present model relies on a limited number of readily obtainable parameters while maintaining performance that is broadly comparable to more complex approaches.

From a calibration perspective, the model showed moderate calibration performance in the validation cohort. The positive calibration intercept indicates that the model tended to underestimate the overall risk of severe dengue. This pattern was most evident in the highest-risk group, where observed event rates exceeded predicted probabilities. Such discrepancies may reflect the relatively small number of patients in the high-risk subgroup, as well as differences in case-mix between the derivation and validation cohorts. These findings underscore the importance of cautious interpretation of absolute risk estimates, particularly at the upper end of the risk spectrum. Recalibration-in-the-large may therefore be required before broader implementation in settings where systematic underestimation is observed.

In this study, age < 55 years was included as a predictor of severe dengue within an adult-only cohort. This finding should not be interpreted as evidence that younger age universally increases the biological risk of severe dengue. Rather, it may represent a cohort-specific association observed within an adult Thai hospital population. The score was developed to stratify risk among adults and was not intended to compare pediatric and adult patients. The observed association may reflect epidemiological characteristics of the derivation cohort, healthcare-seeking behavior, exposure patterns, or residual confounding, and should therefore be interpreted cautiously.

Variability in model performance across studies may be attributable to differences in patient populations, circulating dengue serotypes, and healthcare practices. The inclusion of multiple hospitals in this study enhances the generalizability of the findings and supports the applicability of the model in routine clinical care. From a clinical and public health perspective, this risk score offers practical advantages in resource-constrained settings, as it relies solely on routinely available clinical parameters. Its use may facilitate early identification of high-risk patients, optimize resource allocation, and reduce unnecessary hospital admissions among low-risk individuals.

Several predictors in the score, particularly severe bleeding, pleural effusion, and thrombocytopenia, are closely related to established manifestations of dengue severity. These variables should therefore be interpreted not as distant upstream risk factors, but as pragmatic early in-hospital clinical markers that may identify patients already entering a higher-risk disease trajectory. The exploratory sensitivity analysis demonstrated a marked reduction in discrimination after excluding severe bleeding and pleural effusion from the score. This finding supports the concern that part of the model’s discriminative performance may be attributable to predictors closely related to the severe dengue outcome definition. In particular, pleural effusion reflects plasma leakage that may already prompt clinical escalation and overlaps conceptually with the definition of severe plasma leakage. Accordingly, the score is best viewed as a structured early in-hospital risk stratification tool rather than a purely pre-complication prediction model. Although all predictors were assessed before formal classification of severe dengue in the present analytic framework, the incremental value of the score beyond clinicians’ recognition of these findings may be limited in some cases. The score is intended to complement, not replace, clinical diagnosis by helping structure early recognition and risk communication in routine hospital care.

In Thai hospitals, adult patients with suspected dengue are typically triaged using routine clinical assessment, hemodynamic status, laboratory parameters, and warning signs. In this context, the present score is intended to support early in-hospital risk stratification and monitoring rather than replace standard clinical judgment or guideline-based management. Such a framework may be particularly helpful for standardizing risk assessment across clinicians and supporting decisions in settings with limited experience, heavy workload, or high patient volume. The exploratory clinical utility assessment further suggested that the score may provide practical benefit for early in-hospital triage decisions at a clinically plausible risk threshold. However, because this assessment was exploratory, retrospective, and based on a decision threshold that was not prospectively predefined, these findings should be interpreted cautiously.

Several limitations should be acknowledged. First, the retrospective design may introduce information bias. Second, because both cohorts were assembled during the same calendar period, this study did not evaluate temporal transportability. Similar circulating dengue serotypes, epidemic conditions, and clinical practice patterns across sites during 2019–2022 may have influenced model performance and may have led to an overoptimistic assessment of model stability. However, because serotype data were not available, this remains a plausible but unconfirmed explanation. Third, although the events per variable exceeded recommended thresholds, the sample size may still limit the precision of calibration assessment, particularly across grouped risk strata and in the highest-risk subgroup. Although hospital-specific or hospital-tier-specific performance assessment would be informative, severe dengue events were unevenly distributed across centers and sparse within individual hospitals. Therefore, site-specific estimates would likely be statistically unstable and potentially misleading.

Fourth, predictor variables were extracted from the initial hospital assessment documented in the medical record, prior to formal classification of severe dengue; however, because this was a retrospective multicenter study, the interval between predictor assessment and subsequent clinical deterioration was not uniformly available across all centers. Therefore, the score should be interpreted as supporting early in-hospital risk stratification rather than as a purely prospective prediction tool for future clinical deterioration.

Future prospective studies should therefore collect standardized temporal data on predictor appearance relative to disease onset and progression. Finally, although an exploratory clinical utility assessment was performed, the decision threshold was not prospectively predefined and the analysis was based on retrospective data. Therefore, these findings should be interpreted cautiously and confirmed in prospective implementation studies with predefined clinical decision thresholds.

Overall, this study supports the use of a simple, clinically based dengue severity risk score with good discriminative performance and acceptable calibration. Further prospective studies are needed to confirm its clinical impact, evaluate potential recalibration, and assess its performance in broader populations and practice settings.

Conclusions

This multicenter external validation study showed that a simple dengue severity risk score demonstrated acceptable discriminative performance with moderate calibration across diverse hospital settings in Thailand. Using only routinely available clinical parameters, the score was able to stratify adult dengue patients into clinically meaningful risk categories and may support early in-hospital triage and monitoring decisions, particularly in resource-limited settings.

Although calibration was generally acceptable, some underestimation of risk was observed in higher-risk patients, indicating that further recalibration may be needed before broader implementation. The score should therefore be viewed as a practical tool to complement, rather than replace, routine clinical assessment and judgment.

Further prospective studies are warranted to confirm the clinical impact of the score, evaluate its performance in other epidemiological settings, and assess whether recalibration or implementation-focused refinement could improve its applicability in routine practice.

Abbreviations

ALT

Alanine aminotransferase

AST

Aspartate aminotransferase

AuROC

Area under the receiver operating characteristic curve

BMI

Body mass index

CI

Confidence interval

EPV

Events per variable

ICD-10

International Classification of Diseases, 10th Revision

LR+

Positive likelihood ratio

RT-PCR

Reverse transcription polymerase chain reaction

TRIPOD

Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis

WHO

World Health Organization

NS1

Non-structural protein 1

IgM

Immunoglobulin M

IgG

Immunoglobulin G

Author contributions

Conceptualization, S.P. and P.K.; methodology, S.P. and P.K.; formal analysis, S.P. and P.K.; investigation, S.P. and P.D.; resources, P.D.; data curation, S.P.; writing—original draft preparation, S.P.; writing—review and editing, P.K., P.T., P.D. and S.V.; visualization, P.K.; supervision, S.P.; project administration, P.T. and S.V.; funding acquisition, S.P.

Funding

This work was supported by the Thailand Science Research and Innovation Fundamental Fund in 2022 (Contract No. TUFF 56/2565) and by the Thammasat University Research Fund (Contract No. TUFT-FF 45/2565). The funding bodies had no role in the design of the study; collection, analysis, and interpretation of data; or in writing the manuscript.

Data availability

The datasets generated and/or analyzed during the current study are not publicly available due to ethical restrictions and patient confidentiality but are available from the corresponding author on reasonable request with approval from the relevant ethics committee.

Declarations

Ethics approval and consent to participate

The study was conducted in accordance with the Declaration of Helsinki and was approved by the Human Research Ethics Committee of Thammasat University (Science) (Project No. 011/2565; COA No. 053/2565) and by the ethics committees of all six participating hospitals. As this study used retrospectively collected anonymized clinical data without direct patient contact, the requirement for informed consent was waived by the respective ethics committees. All data were fully anonymized prior to analysis.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.