Development and validation of an interpretable machine learning model for early hospital-based differentiation of chikungunya and dengue fever using routine clinical data

Abstract

Background

Chikungunya fever (CHIKF) and dengue fever are mosquito-borne viral diseases. These infections often circulate in the same regions at the same time. Early symptoms can look very similar between the two diseases. This overlap makes early and accurate diagnosis difficult. Many endemic clinics do not have easy access to molecular tests such as RT-PCR [1, 2]. Clinicians therefore need other practical tools for early decision-making. In this study, we aimed to develop an interpretable machine learning model. The model uses routine clinical signs and standard laboratory results. We also aimed to validate the model for differentiation of CHIKF from dengue fever at initial hospital-based assessment.

Methods

This retrospective observational study analyzed 1,058 laboratory-confirmed arboviral infections, including 366 patients with CHIKF and 692 patients with dengue fever. The dataset was stratified by diagnosis and randomly divided into a training set (n = 742) and a held-out test set (n = 316) at a 7:3 ratio. The team collected clinical symptoms, complete blood count (CBC) results, and inflammatory marker data. The team then used these variables to build eight machine learning models. The study evaluated model performance with discrimination metrics. The study also assessed calibration. The study further used decision curve analysis to estimate clinical usefulness. The team examined feature importance with Shapley Additive Explanations (SHAP). The team deployed the best-performing model as a web-based clinical decision-support tool.

Results

Among the tested approaches, the gradient boosting model (GBM) showed the best and most consistent performance. The GBM achieved a high area under the ROC curve (AUC) in both the training and test sets. The GBM also delivered strong sensitivity and specificity across both cohorts. The SHAP analysis repeatedly highlighted platelet count (PLT) and rash as the most important predictors of CHIKF. These findings match well with known clinical patterns. The online deployment integrated the final model into a simple platform. The platform provides automated, real-time risk estimates using only a small set of routinely available variables.

Conclusion

This study shows that interpretable machine learning models can help clinicians distinguish CHIKF from dengue fever early. The models rely on routine clinical information and standard laboratory tests. These inputs are widely available in many settings. The study also presents a web-based tool that applies the best model at the bedside. The tool may be especially useful in resource-limited clinics. However, the tool still needs external validation. Future studies should test the model in multicenter, prospective cohorts.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12879-026-13778-6.

Introduction

Chikungunya fever (CHIKF) is an acute viral infectious disease caused by the Chikungunya virus (CHIKV) and transmitted primarily by Aedes mosquitoes, notably Aedes albopictus and Aedes aegypti. Typical clinical manifestations include high fever, myalgia, polyarthralgia, and maculopapular rash [1, 3].

In recent years, researchers have reported more CHIKV outbreaks in southern China. Guangdong Province recorded a local outbreak in 2010. Imported cases triggered that outbreak. The event showed that local ecology and climate can support ongoing transmission after the virus enters the region [4]. Foshan sits within the Guangdong-Hong Kong-Macao Greater Bay Area. The city has a high population density. The city also has frequent population movement. These conditions increase the risk of viral importation and local spread [5, 6]. Public health teams therefore need continuous surveillance in this area. Investigators also need detailed studies of the epidemiology and clinical features of CHIKV in Foshan [7].

Dengue fever and chikungunya fever are two major mosquito-borne viral infections. These diseases often co-circulate in tropical and subtropical regions [8]. Both infections share similar mosquito vectors. Outbreaks can also occur at the same time [9]. Clinicians therefore often see patients with acute fever in whom early diagnosis is not straightforward. Dengue fever often presents with high fever, leukopenia, and thrombocytopenia. CHIKF more often presents with rash and severe joint pain. However, many patients show mixed or non-specific symptoms early on. The early phase of illness and the time of initial clinical assessment often show substantial overlap between the two diseases. This window also creates the most diagnostic uncertainty. Clinicians still need to make timely decisions during this period [10–12].

Clinicians need accurate early differentiation between these two infections. Dengue fever can progress to plasma leakage, bleeding, or shock. These complications often require close monitoring and careful fluid management [13]. CHIKF is rarely fatal, but it can lead to chronic inflammatory arthritis. This arthritis can greatly reduce quality of life. Early recognition supports appropriate symptom control. Early recognition also guides follow-up planning. Many endemic settings still lack easy access to molecular tests such as RT-PCR. Serology can also create confusion because arboviruses may cross-react in antibody tests [14]. Clinicians therefore need dependable approaches that rely on routine clinical evaluation and basic laboratory data.

Previous studies have explored single clinical signs or individual laboratory markers to separate CHIKF from dengue fever [15]. However, few studies have combined clinical features and hematological results in the same patient cohort. Many reports also rely on univariable comparisons rather than integrated models. Only a limited number of studies have used multivariable methods or machine learning to support real-world differential diagnosis [16]. Arboviruses are now co-circulating more widely across the globe. Clinicians therefore need robust and evidence-based tools for early triage [10]. Clinicians also need tools that can guide early clinical management.

Therefore, this study aimed to identify independent clinical and hematological predictors that can distinguish dengue fever from CHIKF. The study used a large cohort of laboratory-confirmed cases [17]. We combined clinical symptoms with CBC indices and inflammatory markers. We then used these data to build a practical and interpretable diagnostic model. We designed the model for point-of-care use in regions where both viruses co-circulate.

Methods

Study design and participants

This retrospective observational study included patients with laboratory-confirmed CHIKF or dengue fever treated at The Affiliated Foshan Women and Children Hospital, Guangdong Medical University (Foshan, Guangdong, China) between 2024 and 2025. The final cohort comprised 1,058 patients (366 CHIKF and 692 dengue fever). Laboratory confirmation criteria for CHIKF and dengue fever are provided in the Supplementary Methods (or should be specified according to local laboratory standards).

The final cohort comprised 1,058 patients, including 366 patients with CHIKF and 692 patients with dengue fever. The full dataset was stratified according to the outcome variable and randomly divided into a training set and a held-out test set at a 7:3 ratio. The training set included 742 patients, comprising 257 CHIKF cases and 485 dengue fever cases. The test set included 316 patients, comprising 109 CHIKF cases and 207 dengue fever cases. The training set was used for model development, feature selection, hyperparameter tuning, and cross-validation, whereas the test set was used only for performance evaluation. Missing CRP values were handled using multiple imputation by chained equations, as described below.

Data collection and variable definition

Clinical symptoms and laboratory data were collected from the first inpatient clinical assessment after hospital admission. For laboratory parameters, such as CBC indices and CRP, we used only the first available results from admission or the initial inpatient evaluation. For clinical symptoms, we included only symptoms recorded during the same initial assessment. We did not use the lowest, highest, or most severe laboratory values measured during hospitalization. We also did not include follow-up laboratory results after admission. Therefore, platelet count referred to the baseline platelet count at the first hospital-based assessment. It did not represent the nadir platelet count during the disease course.

Clinical symptoms and laboratory data were extracted from the Laboratory Information System (LIS) and electronic medical records, including age, sex, CBC indices, CRP, and key clinical manifestations (fever, rash, arthralgia, and headache). Binary symptom variables were encoded as 0 (absent) and 1 (present). The outcome variable was CHIKF (1) versus dengue fever (0).

Missing data and imputation

We assessed missing values before imputation. We calculated the number and percentage of missing values for each variable. We performed this assessment in the overall cohort, the training set, and the internal validation set. Supplementary Table S1 presents the missing-rate information. CRP showed most of the missing data. The missing rate for CRP was 23.06% in the overall cohort, 22.78% in the training set, and 23.73% in the internal validation set. The remaining variables had no missing values. These variables included demographic characteristics, clinical symptoms, and complete blood count parameters.

We used multiple imputation by chained equations to handle missing CRP values [18]. We chose MICE because it can impute missing values based on other observed clinical and laboratory variables. This method is also suitable for multivariable clinical datasets. This approach preserved the sample size compared with complete-case analysis. It may also reduce bias under the missing-at-random assumption [19]. We used predictive mean matching for CRP. This method imputes values based on observed data. It also helps avoid clinically implausible laboratory values. We generated five imputed datasets for the subsequent analyses.

Missingness was mainly observed in CRP. Values of CRP reported as below the detection limit, including values beginning with “<”, were treated as missing and recoded as NA before imputation. Predictive mean matching was used to impute CRP because this method preserves the observed distribution and avoids generating implausible negative or extreme CRP values.

Reporting guideline

This study was conducted and reported in accordance with the TRIPOD statement for prediction model development and validation [20].

Feature selection

We also performed LASSO logistic regression in the training set [21]. This analysis aimed to improve the robustness of predictor selection. It also helped reduce the potential influence of multicollinearity among clinical and laboratory variables. We entered all candidate demographic, clinical, inflammatory, and complete blood count variables into the LASSO model. The outcome was chikungunya fever versus dengue fever. We selected the optimal penalty parameter using 10-fold cross-validation. We retained variables with non-zero coefficients at lambda.1se. This approach produced a more parsimonious feature set. It also helped reduce the risk of overfitting. We used multivariable logistic regression to identify clinically interpretable independent predictors. We used LASSO regression as a supplementary feature-selection analysis. This analysis helped assess the robustness of predictor selection.

Machine learning model development

The full dataset included 1,058 laboratory-confirmed patients. We used the createDataPartition function in the R caret package to perform a stratified random split according to the outcome variable. The dataset was divided into a training set and a held-out test set at a 7:3 ratio. The training set included 742 patients and was used for model development, feature selection, hyperparameter tuning, and cross-validation. The test set included 316 patients and was used only for model performance evaluation. No separate external test set was used in this study.

We evaluated eight machine-learning algorithms. These algorithms included logistic regression, support vector machine, gradient boosting machine, neural network, random forest, XGBoost, k-nearest neighbors, and AdaBoost. We trained all models in the training set. We evaluated all models in the test set. We used the same predictor set for each model. We performed hyperparameter tuning within the training set. We used repeated cross-validation for this process. We assessed model performance in both the training set and the test set. We used AUC, accuracy, sensitivity, specificity, precision, and F1 score as performance metrics. We also used calibration curves to assess model calibration. We used decision curve analysis to evaluate clinical utility [22]. We selected the final model based on several factors. These factors included test performance, calibration, clinical utility, and interpretability. We did not select the final model based only on training performance. We carefully reviewed models with perfect or near-perfect training performance. This review helped us assess the possible risk of overfitting.

Feature importance

The SHAP analysis provided an explanation of the crucial factors that played a significant role in generating model predictions and quantified their respective contributions to the overall performance of the model [23]. In this study, we calculated the feature importance using the SHAP value for eight models.

Statistical methods

All analyses were performed using R software version 4.4.3. Categorical variables are presented as number and percentage and were compared between groups using the chi-square test or Fisher’s exact test, as appropriate. The normality of continuous variables was assessed using the Shapiro-Wilk test, and the results are presented in Supplementary Table S2. Normally distributed continuous variables are presented as mean ± standard deviation and were compared using Student’s t-test. Non-normally distributed continuous variables are presented as median with interquartile range and were compared using the Mann-Whitney U test. A two-sided P value < 0.05 was considered statistically significant.

Results

Baseline characteristics of the training cohort

A total of 1,058 laboratory-confirmed patients were included in the final analysis. After stratified random sampling according to the outcome variable, 742 patients were assigned to the training cohort and 316 patients were assigned to the held-out test set. The training cohort included 257 CHIKF cases and 485 dengue fever cases, whereas the test set included 109 CHIKF cases and 207 dengue fever cases. Baseline characteristics of the training cohort are shown in Table 1. All clinical symptoms and laboratory predictors used for model development were obtained from the first hospital-based assessment after admission. Follow-up laboratory results, in-hospital platelet nadir values, and longitudinal disease-course information were not included in the model. Missing-data analysis showed that CRP was the only variable with missing values. The missing rate of CRP was 23.06% in the overall cohort, 22.78% in the training set, and 23.73% in the test set. All other variables had no missing values. Detailed missing-rate information is provided in Supplementary Table S1.

Table 1.

Baseline characteristics of the training cohort (n = 742)

	Category	Dengue fever, n = 485	Chikungunya fever, n = 257	p
Sex, n (%)	Female	284 (58.6)	147 (57.2)	0.781
	Male	201 (41.4)	110 (42.8)
Rash, n (%)	No	412 (84.9)	75 (29.2)	<0.001
	Yes	73 (15.1)	182 (70.8)
Fever, n (%)	No	23 (4.7)	81 (31.5)	<0.001
	Yes	462 (95.3)	176 (68.5)
Arthralgia, n (%)	No	457 (94.2)	105 (40.9)	<0.001
	Yes	28 (5.8)	152 (59.1)
Headache, n (%)	No	265 (54.6)	244 (94.9)	<0.001
	Yes	220 (45.4)	13 (5.1)
Age, median (IQR)		33.00 [12.00, 54.00]	33.00 [11.00, 55.00]	0.804
CRP, median (IQR)		4.50 [1.60, 10.30]	6.80 [2.40, 14.70]	0.001
WBC, median (IQR)		3.36 [2.46, 5.25]	5.96 [4.54, 7.66]	<0.001
RBC, median (IQR)		4.52 [4.21, 4.94]	4.53 [4.17, 4.93]	0.622
HCT, mean ± SD		0.39 (0.04)	0.38 (0.04)	0.122
RDW-CV, median (IQR)		12.70 [12.20, 13.30]	12.60 [12.30, 13.30]	0.555
RDW-SD, median (IQR)		40.00 [37.90, 42.00]	39.70 [37.40, 42.30]	0.684
P-LCR, median (IQR)		22.70 [17.40, 28.40]	18.80 [15.40, 23.90]	<0.001
MONO#, median (IQR)		10.80 [8.10, 13.50]	9.30 [7.10, 11.40]	<0.001
MONO%, median (IQR)		0.22 [0.02, 0.43]	0.52 [0.40, 0.71]	<0.001
LYM#, median (IQR)		29.60 [16.40, 42.90]	20.10 [12.90, 31.00]	<0.001
LYM%, median (IQR)		1.11 [0.73, 1.94]	1.16 [0.81, 1.68]	0.86
MCV, median (IQR)		86.30 [82.70, 89.80]	86.10 [81.80, 89.70]	0.468
MCH, median (IQR)		29.20 [27.70, 30.40]	29.20 [27.50, 30.40]	0.327
MCHC, median (IQR)		338.00 [331.00, 344.00]	336.00 [328.00, 342.00]	0.019
MPV, median (IQR)		339.00 [331.00, 344.00]	9.20 [8.80, 9.90]	<0.001
BASO#, median (IQR)		340.00 [331.00, 344.00]	0.30 [0.20, 0.40]	0.001
BASO%, median (IQR)		341.00 [331.00, 344.00]	0.02 [0.01, 0.03]	<0.001
EO%, median (IQR)		342.00 [331.00, 344.00]	0.06 [0.03, 0.13]	0.802
EO#, median (IQR)		343.00 [331.00, 344.00]	1.00 [0.40, 2.20]	<0.001
HGB, mean ± SD		344.00 [331.00, 344.00]	128.58 (15.48)	0.027
PCT, median (IQR)		345.00 [331.00, 344.00]	0.22 [0.19, 0.26]	<0.001
PDW, median (IQR)		346.00 [331.00, 344.00]	9.60 [8.70, 11.20]	0.001
PLT, median (IQR)		347.00 [331.00, 344.00]	248.00 [198.00, 287.00]	<0.001
NEUT#, median (IQR)		348.00 [331.00, 344.00]	68.90 [54.30, 76.90]	<0.001
NEUT%, median (IQR)		349.00 [331.00, 344.00]	3.86 [2.38, 5.64]	<0.001

Data are presented as n (%) for categorical variables, mean ± SD for normally distributed continuous variables, and median (IQR) for non-normally distributed continuous variables. Between-group comparisons were performed using the chi-square test or Fisher’s exact test for categorical variables, Student’s t-test for normally distributed continuous variables, and the Mann-Whitney U test for non-normally distributed continuous variables

The training cohort included 485 dengue fever cases and 257 chikungunya fever cases. Compared with patients with dengue fever, patients with chikungunya fever had higher rates of rash and arthralgia, but lower rates of fever and headache.

Laboratory parameters also differed significantly between the two groups. Compared with patients with dengue fever, patients with chikungunya fever had higher CRP levels, WBC counts, and platelet counts. Because these variables were non-normally distributed, they are presented as median (IQR) and were compared using the Mann-Whitney U test, as shown in Table 1.

Age distribution was similar across the two groups (p > 0.05; Table 1). Sex distribution was also comparable between groups (p > 0.05; Table 1). We then split the dataset into a training set (n = 742) and an independent test set (n = 316). The test set included 109 chikungunya cases and 207 dengue cases (Fig. 1).

Fig. 1.

Confusion matrices of the eight candidate models in the test set. Confusion matrices are shown for (A) logistic regression, (B) support vector machine, (C) gradient boosting machine, (D) neural network, (E) random forest, (F) XGBoost, (G) k-nearest neighbors, and (H) AdaBoost. Each panel summarizes true positives, false positives, false negatives, and true negatives using the model-specific optimal threshold. Percentages and absolute counts are shown within each cell

Predictor identification and feature selection

We ran a multivariable logistic regression analysis (Table 2). The model identified arthralgia as the strongest independent predictor of chikungunya (adjusted OR 71.87, 95% CI 23.45-220.25, p < 0.001). The model also highlighted rash as a strong predictor (adjusted OR 44.68, 95% CI 15.49-128.91, p < 0.001). The model identified inverse associations for two symptoms. The presence of fever was associated with significantly lower odds of being diagnosed with chikungunya fever rather than dengue fever (adjusted OR 0.10, 95% CI 0.03–0.35, p < 0.001). Similarly, the presence of headache was associated with significantly lower odds of being diagnosed with chikungunya fever rather than dengue fever (adjusted OR 0.03, 95% CI 0.01–0.10, p < 0.01).

Table 2.

Multivariable logistic regression analysis of independent predictors of chikungunya

Variable	Adjusted OR	95% CI	p-value
Rash (Yes)	44.68	15.49 – 128.91	<0.001
Arthralgia (Yes)	71.87	23.45 – 220.25	<0.001
Fever (Yes)	0.1	0.03 – 0.35	<0.001
Headache(Yes)	0.03	0.01–0.10	<0.01
CRP (per 1 mg/L)	1.03	1.01 – 1.06	0.013
Lymphocyte %	0.98	0.90 – 1.08	0.732
Neutrophil %	1.03	0.95 – 1.12	0.47
Platelet count	1.04	1.01 – 1.07	0.014
Age	1	0.99 – 1.01	0.859
WBC	1.29	0.68 – 2.24	0.438
RBC	0.97	0.75 – 1.25	0.822
MCV	0.99	0.96 – 1.01	0.184
MCH	1.36	1.13 – 1.65	0.002
MCHC	0.93	0.88 – 0.97	0.003
MPV	0.71	0.19 – 2.65	0.606
Monocytes (Abs)	0.84	0.68 –1.04	0.114
Hemoglobin	1.01	0.99 – 1.04	0.318
PDW	1.08	0.84 – 1.39	0.547

Several laboratory variables also showed significant associations. Higher CRP levels were associated with higher odds of being diagnosed with chikungunya fever rather than dengue fever (adjusted OR 1.03, 95% CI 1.01–1.06, p = 0.013). Higher platelet count showed a similar positive association (adjusted OR 1.04, 95% CI 1.01–1.07, p = 0.014), as did MCH (adjusted OR 1.36, 95% CI 1.13–1.65, p = 0.002). In contrast, MCHC was inversely associated with chikungunya fever rather than dengue fever (adjusted OR 0.93, 95% CI 0.88–0.97, p = 0.003). Other covariates, including age, WBC, and differential leukocyte counts, were not statistically significant in the adjusted model (Table 2).

To further evaluate the robustness of predictor selection, LASSO logistic regression was performed in the training cohort. Using 10-fold cross-validation, lambda.1se was selected to obtain a more parsimonious model. The variables retained with non-zero coefficients were rash, fever, arthralgia, headache, age, CRP, MONO%, EO, EO%, PLT, and NEUT. These LASSO-selected variables were broadly consistent with the multivariable logistic regression and model-interpretability results, particularly for rash, arthralgia, headache, fever, CRP, and PLT. The full LASSO-selected feature set is shown in Supplementary Table S3, and the cross-validation and coefficient-path plots are shown in Supplementary Figure S1.

Model discrimination performance

Eight algorithms were evaluated: logistic regression, support vector machine, gradient boosting machine, neural network, random forest, XGBoost, k-nearest neighbors, and AdaBoost. ROC analysis demonstrated excellent discrimination in both the training and internal validation sets (Fig. 2A-B). In the training set, AUCs ranged from 0.945 for AdaBoost to 1.000 for random forest (Fig. 2A). In the internal validation set, AUCs remained high, ranging from 0.944 for AdaBoost to 0.985 for the neural network (Fig. 2B). The highest test AUC was achieved by the neural network (AUC 0.985, 95% CI 0.974–0.995), while GBM also performed strongly (AUC 0.981, 95% CI 0.969–0.994). Random forest showed perfect apparent performance in the training set but lower performance in the internal validation set, suggesting possible overfitting rather than superior generalizability.

Fig. 2.

Discrimination, calibration, and clinical utility of the eight candidate models in the training and test sets. (A-B) Receiver operating characteristic (ROC) curves of logistic regression, support vector machine, gradient boosting machine, neural network, random forest, XGBoost, k-nearest neighbors, and AdaBoost in the training set (A) and internal validation set (B). The corresponding AUCs with 95% confidence intervals are shown in the legends. (C-D) Calibration plots of the eight candidate models in the training set (C) and internal validation set (D). The dashed diagonal line indicates perfect calibration; points represent observed event rates within predicted-risk bins. (E-F) Decision curve analysis (DCA) of the eight candidate models in the training set (E) and internal validation set (F), showing standardized net benefit across a range of risk thresholds. “All” and “None” denote strategies of treating all or no patients as high-risk, respectively

Using the prespecified optimal thresholds (Tables 3 and 4), overall classification performance in the internal validation set was robust across models (Table 3). GBM achieved the highest accuracy (0.953) with sensitivity 0.917, specificity 0.971, precision 0.943, and F1-score 0.930 (threshold = 0.476). The neural network provided the highest sensitivity (0.963) with accuracy 0.946, whereas AdaBoost showed reduced sensitivity (0.789) despite high specificity (0.961) (Table 3). Training set results are summarized in Table 4. Random forest achieved perfect apparent training performance, with training accuracy and AUC values of 1.000. This result was interpreted cautiously because perfect training performance may reflect overfitting rather than superior generalizability. Therefore, model selection was based primarily on test performance, calibration, decision curve analysis, and interpretability.

Table 3.

Prespecified optimal thresholds and classification performance of the eight candidate models in the internal validation set

Evaluation metrics for the test set
Model	Threshold	Accuracy	Sensitivity	Specificity	Precision	F1
Logistic	0.41229167	0.946	0.945	0.947	0.904	0.924
SVM	0.740484053	0.934	0.945	0.928	0.873	0.907
GBM	0.475918552	0.953	0.917	0.971	0.943	0.93
NeuralNetwork	0.246619134	0.946	0.963	0.937	0.89	0.925
RandomForest	0.378	0.918	0.936	0.908	0.843	0.887
Xgboost	0.49995324	0.949	0.936	0.957	0.919	0.927
KNN	0.513737515	0.934	0.927	0.937	0.886	0.906
Adaboost	0.770556835	0.902	0.789	0.961	0.915	0.847

Table 4.

Prespecified optimal thresholds and classification performance of the eight candidate models in the training set

Evaluation metrics for the train set
Model	Threshold	Accuracy	Sensitivity	Specificity	Precision	F1
SVM	0.862108572347181	0.923	0.879	0.946	0.897	0.888
GBM	0.37202860038231	0.961	0.965	0.959	0.925	0.945
NeuralNetwork	0.292487285056865	0.931	0.934	0.93	0.876	0.904
RandomForest	0.5	1	1	1	1	1
KNN	0.322413374617368	0.969	1	0.953	0.918	0.957
Adaboost	0.770556835341249	0.903	0.774	0.971	0.934	0.847
LightGBM	0.496435698789205	1	1	1	1	1
CatBoost	0.606769831383696	0.943	0.907	0.963	0.928	0.917

Confusion matrix results on the internal validation set

Confusion matrices are shown in Fig. 1. For GBM, the model correctly classified 100/109 chikungunya cases (true positives) and 201/207 dengue cases (true negatives), with 9 false negatives and 6 false positives (Fig. 1). The neural network produced 105 true positives and 194 true negatives, while AdaBoost produced 86 true positives and 199 true negatives, reflecting its lower sensitivity (Fig. 1). These findings are consistent with the summary metrics in Table 3.

Calibration and clinical utility

Calibration curves indicated generally good agreement between predicted and observed risks across models (Fig. 2C-D), with more noticeable deviations for some classifiers (e.g., AdaBoost) at intermediate probability bins. Decision curve analysis (DCA) demonstrated that the evaluated models provided higher standardized net benefit than “treat-all” and “treat-none” strategies across a wide range of clinically relevant threshold probabilities in both the training and internal validation sets (Fig. 2E-F), supporting their potential utility for early diagnostic risk stratification.

Model interpretability and key predictors

Feature-importance rankings were broadly consistent across models (Fig. 3). PLT, arthralgia, and rash were repeatedly identified as dominant contributors, followed by headache, fever, and inflammatory/erythrocyte indices (e.g., CRP, MCH, MCHC) depending on the algorithm.

Fig. 3.

Feature importance rankings across the eight machine-learning models. Bar plots show the relative importance of predictors in each model: (A) logistic regression, (B) support vector machine, (C) gradient boosting machine, (D) neural network, (E) random forest, (F) XGBoost, (G) k-nearest neighbors, and (H) AdaBoost. Across models, platelet count (PLT), arthralgia, rash, headache, fever, C-reactive protein (CRP), mean corpuscular hemoglobin (MCH), and mean corpuscular hemoglobin concentration (MCHC) contributed to model prediction to varying degrees

To provide a more comprehensive interpretation of the GBM model, SHAP dependence plots were generated for the top-ranked predictors, including PLT, rash, arthralgia, and headache (Fig. 4). The PLT dependence plot showed that higher platelet counts generally contributed positively to the prediction of chikungunya fever. This pattern is clinically plausible because dengue fever is more commonly associated with prominent thrombocytopenia, whereas platelet reduction is usually less characteristic in chikungunya fever [24]. Notably, the distribution of SHAP values differed according to rash status, suggesting an interaction between hematological status and clinical manifestation. Patients with both relatively preserved or elevated PLT and rash tended to receive a stronger positive SHAP contribution toward chikungunya classification, whereas lower PLT values weakened this contribution even when rash was present. This indicates that the model did not rely on rash alone, but interpreted rash in the context of platelet profile.

Fig. 4.

Model interpretability of the GBM classifier using SHAP. (A) SHAP force plot illustrating how individual features shift the model output from the baseline expected value to the final predicted risk for a representative patient. (B) Global feature importance ranked by mean absolute SHAP values. (C) SHAP summary plot showing the distribution of SHAP values for each feature and the corresponding feature values. (D) SHAP decision plot visualizing cumulative feature contributions to the predicted risk. (E) SHAP dependence plot for PLT, colored by rash status, showing the interaction between platelet count and rash. (F) SHAP dependence plot for rash. (G) SHAP dependence plot for arthralgia. (H) SHAP dependence plot for headache. Positive SHAP values indicate increased contribution toward chikungunya fever classification, whereas negative SHAP values indicate contribution toward dengue fever classification

The additional dependence plots for rash and arthralgia further supported their strong positive contribution to chikungunya prediction. The presence of rash and arthralgia shifted SHAP values toward chikungunya classification, consistent with the classical clinical presentation of chikungunya fever. In contrast, headache showed a generally negative contribution to chikungunya prediction, which was consistent with the multivariable logistic regression results showing that headache was inversely associated with chikungunya fever. Together, these findings indicate that the GBM model captured clinically meaningful patterns: preserved platelet count, rash, and arthralgia favored chikungunya fever, whereas prominent headache and lower platelet count were more compatible with dengue fever (Fig. 4).

Online platform for clinical application

To facilitate clinical application, we integrated the optimized GBM model into a web-based platform (Fig. 5). Users can obtain an automated risk prediction for CHIKF by simply entering a set of routinely available clinical and laboratory variables. This platform demonstrates the feasibility of translating the proposed model into a practical clinical decision-support tool. The Web server application is freely accessible via https://hongmenghui.shinyapps.io/make_web/.

Fig. 5.

Web-based clinical decision support tool for individualized risk prediction. Screenshot of the web-based platform integrating the optimized model. Users input routinely available clinical symptoms and laboratory variables to obtain an automated individualized risk probability and risk stratification output

Discussion

In this study, we developed and validated multiple machine learning models to enable the early differentiation of CHIKF from dengue fever using readily available clinical symptoms, CBC parameters, and inflammatory markers. Among the eight evaluated models, GBM demonstrated strong and balanced performance across discrimination, calibration, clinical utility, and interpretability metrics, achieving high AUC values as well as favorable sensitivity and specificity in both training and test sets [25]. Although the neural network achieved the highest test AUC, GBM was selected as the final model because it provided robust validation performance together with better interpretability and practical applicability [16].

Accurate early distinction between CHIKF and dengue fever remains a major clinical challenge, particularly during the initial phase of illness when symptoms are nonspecific and laboratory abnormalities overlap [26]. Both diseases are transmitted by Aedes mosquitoes and frequently co-circulate.

Previous studies have explored individual clinical features or laboratory parameters to distinguish CHIKF from dengue fever, often highlighting the diagnostic value of rash, arthralgia, leukopenia, or thrombocytopenia. However, most prior investigations relied on univariate analyses or conventional statistical models and did not comprehensively integrate multiple data domains within a single predictive framework. Moreover, relatively few studies have focused on true differential diagnosis between CHIKF and dengue fever using machine learning approaches [27], despite the increasing availability of electronic health records and advances in computational methods that enable the integration of heterogeneous clinical and laboratory data.

Compared with existing research, the present study offers several methodological and practical advantages. First, it integrates clinical symptoms, CBC indices, and inflammatory markers within a unified modeling framework, thereby capturing both clinical presentation and host hematologic response. Second, the use of gradient boosting-based models, particularly GBM enables effective handling of nonlinear relationships and complex feature interactions that may be overlooked by traditional regression-based approaches. Third, model interpretability was explicitly addressed through SHAP analysis, which revealed that PLT, rash, and arthralgia were among the most influential features driving model predictions, consistent with established clinical knowledge of CHIKF. The interaction between PLT and rash further provides clinically meaningful information. Although rash is a typical manifestation of chikungunya fever, it is not entirely specific and may also occur in dengue fever. The SHAP dependence analysis suggested that rash contributed more strongly to chikungunya prediction when accompanied by relatively preserved or elevated platelet counts. This is clinically plausible because dengue fever is commonly associated with more pronounced thrombocytopenia, whereas chikungunya fever more often presents with rash and arthralgia and may show less severe platelet reduction. Therefore, the PLT-rash interaction indicates that the GBM model integrated both clinical manifestations and hematological profiles rather than relying on a single symptom. In addition, LASSO logistic regression was performed to provide a more robust and parsimonious feature-selection strategy. Although some hematological variables selected by LASSO differed from those identified by multivariable logistic regression, this discrepancy may reflect correlations among CBC-derived indices. Importantly, the core clinical predictors, including rash, arthralgia, fever, headache, and PLT, remained consistent across methods. These findings suggest that the main predictor signals were stable across conventional regression, penalized feature selection, and machine-learning interpretability analyses.

Importantly, this study moves beyond model development by demonstrating clinical applicability through the deployment of an online risk prediction platform. The platform requires only a limited number of routinely collected variables. It provides automated and real-time risk assessment. This function can support early triage and diagnostic decision-making. It may be especially useful in resource-limited settings [28]. From a broader perspective, this work highlights the potential of interpretable machine learning models as adjunctive diagnostic tools for arboviral infections and offers a scalable framework that could be adapted to other co-circulating mosquito-borne diseases.

A critical methodological issue in early diagnostic model development is the timing of predictor measurement. Dengue fever may show dynamic hematological changes during the disease course, and the use of in-hospital nadir platelet counts or follow-up laboratory results could introduce temporal data leakage. In this study, all CBC indices, CRP values, and clinical symptoms were restricted to the first hospital-based assessment after admission. No subsequent laboratory measurements or in-hospital worst values were used. Therefore, the contribution of PLT in the model should be interpreted as the diagnostic value of baseline platelet status at initial inpatient evaluation rather than the platelet nadir during disease progression.

The perfect apparent performance of random forest in the training cohort should be interpreted with caution. Although tree-based ensemble models can capture nonlinear relationships and high-order interactions, they are also susceptible to overfitting, particularly in relatively small clinical datasets with a limited number of highly informative predictors. In the present study, the perfect training accuracy and AUC observed for random forest was considered signs of potential overfitting rather than evidence of superior clinical performance. Therefore, these models were not selected as the final model despite their excellent apparent training results. Instead, the final GBM model was chosen based on its more balanced performance across test discrimination, calibration, decision curve analysis, and interpretability.

However, because this study was conducted during a large outbreak, the included patients were mainly those admitted to hospital for more severe symptoms or closer clinical evaluation. Patients with mild symptoms were often managed in community healthcare settings or underwent home isolation. Therefore, the study population may not fully represent all patients with chikungunya fever or dengue fever in the community. This potential spectrum bias may affect model generalizability, and the model should be interpreted as a hospital-admission-based decision-support tool rather than a screening tool for all community cases. External validation in community-based and multicenter cohorts is needed before broader application.

Despite these strengths, several limitations should be acknowledged. First, this was a retrospective, single-center study based on hospital-admitted patients, which may limit the generalizability of the findings to other geographic regions, healthcare settings, and patients with mild disease managed in community clinics or home isolation. During the outbreak period, patients with mild symptoms were often diverted to community healthcare facilities or home-based management, whereas hospitalized patients generally had more severe symptoms or required closer clinical assessment. Therefore, the model should be interpreted as a tool for early differentiation among hospital-admitted patients at initial inpatient evaluation, rather than as a universal screening model for all community cases. External validation using multicenter, prospective cohorts is necessary to confirm the robustness and transportability of the proposed models. Second, the study focused on laboratory-confirmed cases of CHIKF and dengue fever; therefore, the model’s performance in patients with co-infections, atypical presentations, or other febrile illnesses remains to be determined. Third, although the online platform demonstrates feasibility, its impact on real-world clinical workflows and physician decision-making has not yet been formally evaluated.

Future studies should aim to externally validate the model across diverse populations, incorporate additional clinical or biomarker data when available, and assess prospective clinical utility. Such efforts will be essential to further refine machine learning-based diagnostic tools and to establish their role in improving early recognition and management of arboviral diseases [29].

Conclusion

This study demonstrates that GBM is effective machine learning models for the early differentiation of Chikungunya fever from dengue fever using routine clinical symptoms and basic laboratory parameters. By integrating key features such as rash, arthralgia, and complete blood count indices, the proposed model achieved high discriminatory performance and good clinical utility.

The deployment of the optimal model as a web-based tool further highlights its potential to support early clinical decision-making, particularly in resource-limited settings where access to molecular diagnostics is restricted. Nevertheless, external validation in multicenter and prospective cohorts is warranted to confirm the generalizability and real-world applicability of this approach. This model may serve as a practical hospital-based decision-support tool for patients requiring inpatient assessment in endemic regions.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

CHIKF

Chikungunya fever

CHIKV

Chikungunya virus

CBC

Complete Blood Count

CRP

C-reactive protein

DCA

Decision curve analysis

GBM

Gradient Boosting Model

PLT

Platelet count

SHAP

Shapley Additive Explanations

Author contributions

Meng-Hui Hong conceived the study. Meng-Hui Hong and Lin Zhang designed the experiments. Lin Zhang performed the experiments. Jing Liu analyzed the data. Jia-Liang Mai prepared the figures and tables. Lin Zhang wrote the main manuscript text. All authors reviewed and approved the final manuscript.

Funding

This study was supported by the Foshan self-funded science and technology innovation projects (2520001002503).

Data availability

The datasets generated and/or analyzed during the current study are not publicly available due to patient privacy and confidentiality restrictions. However, anonymized data can be made available from the corresponding author upon reasonable request and with appropriate ethics approval.

Declarations

Ethical approval

This study was approved by the Ethics Committee of The Affiliated Foshan Women and Children Hospital, Guangdong Medical University (approval number: FSFY-MEC-2026-062). The study was conducted in accordance with the ethical principles of the Declaration of Helsinki. The requirement for informed consent was waived by the Ethics Committee because this was a retrospective study using anonymized clinical data.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.