This retrospective multicentric study with one cohort accessed electronic health records of a network of seven hospitals in India for all confirmed dengue patients (with the NS-1 antigen test or serologically confirmed dengue infection) admitted during the study period from February 2012 to September 2017. All multiday stay inpatients, including emergency and elective, who were at least one year old were included in the study. A case was excluded if the patient had a negative result for dengue or serology for febrile illness and in cases of insufficient data.

The data included the demographic profile of patients, administrative characteristics, laboratory parameters and radiological parameters. A total of 1360 patients with disease-related diagnoses of dengue were admitted to this hospital network. The study was reviewed and approved by the institutional healthcare ethics committee. The standard WHO definitions were used to classify suspected dengue infection.23 Two methods were used to confirm infection:

• 1.

  An NS-1 antigen test (Panbio Dengue Early ELISA, Standard Diagnostics Inc., Republic of Korea) was performed if the patient presented within five days after the onset of the symptoms.

• 2.

  A test for dengue immunoglobulin M (IgM) serum (NIVDEN Immunoglobulin [IgM] Capture ELISA, National Institute of Virology, Pune, India) was performed when the patient presented more than five days after the onset of the symptoms.

Data for the study was extracted from the hospital digital information systems. Data was retrieved sequentially, beginning with the demographics of confirmed dengue patients. Subsequently, diagnostic investigations, administrative and clinical information was extracted. However, 212 cases were excluded from the study due to unavailability of complete datasets.

Initially, based on the aggregate data of discharged dengue patients, forty candidate risk factors such as age, gender, type of admission, ventilation assistance, laboratory, and radiological characteristics were included. The means, medians, percentages, and lower and upper quartiles of the features are given in Table 1.

The final dataset included 1148 patients for further analysis. Demographic and clinical data were recorded in a template pro-forma upon admission, whereas lab related findings were collected on daily basis until discharge or death. The dataset includes features with continuous and categorical values for demographic, administrative, laboratory, and radiological data. New fields were created based on existing features as shown in Table 1 of the Supplementary material. These features are transformed to binary features to improve the interpretation.

All the transfused blood components (red blood cells, plasma, cryoprecipitate, or platelets) were considered based on the research question and data availability and the need for any component (binary outcome) was chosen as a classification target. The eight state-of-the-art machine-learning methods Adaboost, ExtraTrees, gradient boosting decision tree (GDBT), k-nearest neighbour (KNN), logistic regression (LR), multilayer perceptron (MLP), random forests (RFs), and extremely gradient boosting (XGBoost) were considered in the classification of blood component needs (transfusion vs. non-transfusion).

Data cleansing and pre-processing are essential to achieve optimal results. The following steps were taken to prepare and clean the data:

• 1.

  Redundant records and fields that had more than 50 % of their data missing were eliminated.

• 2.

  Following that, all categorical features were one-hot encoded. A popular encoding method is one-hot encoding. It works by making a new variable for each categorical variable present in the feature by assigning a 1 or 0 to indicate the presence or absence of the category in the data.

• 3.

  The final data set had 389 columns.

• 4.

  Missing values in patient data are quite common in medical environments. Details for missing data handling are discussed in a later section.

Feature engineering is an important task in data preparation, but it is also a time-consuming component of machine-learning applications.24 Initially, Forty predictors were collected using the aggregate data of discharged patients, which included demographics, administrative, lab test results, and radiology related variables and subsequently new features were derived from these variables. The institution's standard lab reference ranges were used for laboratory data with features being coded based on whether they were below, within or above the reference range (Supplementary material Table 1) and other features were categorized into binary variables. Following the feature engineering process; one dependant and 389 independent features were identified. Supplementary material Table 1 presents a summary of the feature engineering techniques.

The techniques used by Ansari et al.25, Hachesu, et al.26 and Azari, et al.27 were used to handle missing values in this study. Features with more than 50 % of records of missing values were removed from the dataset; commonly these included prothrombin time, activated partial thromboplastin time, and basophil count (Table 1). The mean or median was used for continuous features and regression was used for nominal or ordinal type features with missing values below 50 %.

Variables for the blood component transfusion prediction model were selected following reviews of previous research on predicting and/or administering transfusion to dengue patients.21,22 Subsequently, these variables were reviewed by experts to ensure that relevant variables were being used as predictors. The voted method was used for feature selection (for details see Supplementary material) to select significant variables. This method applies a variety of techniques to select features. A vote is given when an algorithm selects a feature. In the end, this method calculates the total votes for each feature and then picks out the best ones based on votes with the selected feature being used to develop the machine-learning models.

After cleansing and pre-processing, 1148 complete records were extracted and obtained for classification tasks. A crucial step in assessing the effectiveness of data mining models is dividing the data into training and testing sets. In this study, the dataset was randomly split into training and testing datasets. Seventy percent of the data was used to train our models, while 30 % was used as a testing set. Five-fold cross validation was performed on the training set to tune the hyperparameter of the models. In the training set, 23.4 % patients received blood component transfusions during their hospital stay and in the testing set, 25.2 % patients received blood component transfusions.

AdaBoost28 is a machine-learning approach based on the idea of combining many relatively weak and inaccurate prediction rules to create a highly accurate prediction rule. AdaBoost, like all ensemble methods, works by generating a set of classifiers and then voting on them to classify test examples. The ExtraTrees algorithm29 is a supervised learning method for classification that works by training a number of randomized decision trees (also known as extra-trees) on different sub-samples of the data and then averaging them to enhance the predictive accuracy and control over-fitting. The GBDT algorithm30,31 is a popular machine-learning technique that involves implementing several models and aggregating their results. The KNN32 algorithm is a simple non-parametric, supervised machine-learning technique. It stores all the available data and a new data point is classified by searching for the K most similar data points (the neighbours) across the entire training dataset and then these K data points are summarized as the outcome variable. LR33 is a nonlinear regression model, where the target variable is binary. It solves binary classification problems by utilizing a linear combination of input data. Additionally it can predict the probability value between 0 and 1 of an event by fitting data objects in a logistic function. MLP34 is the classical type of neural network algorithm comprising of one or more layers of neurons. It is often applied to supervised learning problems: it learns to model the correlation between inputs and outputs by training on a set of input-output pairs. Training of the model involves adjusting the parameters, or the weights and biases, in order to minimize error. RF35 is an ensemble of multiple decision trees. Each non-leaf node in a decision tree represents a test on a feature, each branch denotes a test output, and each leaf node holds an outcome label. Not only is RF fast, easy to implement, and produces precise predictions, but it can also manage a large number of independent variables without overfitting.36 XGBoost37 is a widely used and highly effective decision-tree-based ensemble machine-learning algorithm that employs a gradient boosting framework. It implements several models and aggregates the results.

The dataset was randomly split into two groups, a training and a testing set. The training set of 803 (70 %) patients was used to create the models, while the testing set of 345 (30 %) patients was used to test the models. The bootstrap method was applied 1000 times to the internal validation set in order to determine the confidence interval (CI) for the area under the curve (AUC), accuracy, sensitivity, specificity, positive (PPV) and negative (NPV) predictive values, and F1-score. Discrimination was evaluated using area under the receiver operating characteristic curve (AUROC)38 of predicted probabilities in dengue patients for both transfusion and non-transfusion groups. The calibration of the predictive model was assessed using a calibration curve.

In the initial data analysis, independent continuous variables between transfused and non-transfused groups were compared either using the Student t-test or Mann-Whitney U test as considered suitable. Either chi-square test or Fisher's exact test was used to examine associations between the categorical variables.

The proposed prediction model was developed based on the most important features in the “XGBoost” package in Python language. The parameter values of the models were automatically adjusted. The performances of eight models were compared to select the best model. Five-fold cross-validation was used to train and tune the hyperparameter of the model. Specifically, the training set was split into five data sets, four data sets were used to train the model, and the remaining set was utilized to evaluate the performance/generalizability of the model.

The performance of each model was assessed using a set of performance metrics such as AUC, accuracy, sensitivity, precision, PPV and NPV. The decision threshold value was set to maximize the AUC value which was selected here due to its robustness to outcome class imbalance (if all patients were assigned to the larger class, the accuracy would be quite high but the sensitivity would be very low). A confusion matrix was used to compute the PPV or precision, sensitivity or recall specificity, and NPV, The F1 score and accuracy are shown in Eqs. (1)-(6).

Moreover, the area under the AUROC was plotted as a measure of the model performance, with larger AUCs indicating better performance. After the model was developed, the Shapley Additive exPlanations (SHAP) package was used to explain an instance prediction by computing the importance of each feature in the model prediction. The SHAP package interpreted the output of the machine-learning model using a coalitional game-theoretic approach.39

As shown in Figure 1, final enrolment in the study included 1148 patients; information on the cohort is shown in Table 1. The mean age of the patients was 30.34 years; 60.3 % of them were men. The mean ages of the non-transfusion and transfusion group patients were 28.20 (SD: ± 18.29) and 37.14 (SD: ± 16.75) years, respectively. The transfusion group patients were statistically significantly older than those of the non-transfusion group (p-value <0.01). Patients in the transfusion group had significantly longer hospital stays (5.1 days; interquartile range [IQR]: 4.04–6.88; p-value: <0.01) than those in the non-transfusion group (3.85 days; IQR: 2.9–5.0). The 10- to 20-year age group, with a male predominance, had the highest proportion of dengue cases (21.1 %) followed by the 20- to 30-year age group (19.4 %) and the 30- to 40-year age group (19.2 %). The over 70-year-old age group had the fewest cases. Blood component transfusions were given in 24 % (275/1148) of patients, while 76 % (873/1148) of the patients received no blood component transfusions. Table 1 compares the characteristics of patients who received blood component transfusions versus those who did not.

Figure 1.

. Modelling process and methodology flow of the study. (A) Data was acquired from an electronic health records of a network of hospitals; the included features were demographic, administrative, radiological, and lab related parameters. After feature engineering, 389 features were itemized and 16 features were screened thus 16 features were used to develop the model. (B) The methodology flow of the study.

(0.21MB).

Amongst all confirmed dengue cases, three patients (0.26 %) had platelet counts ≤ 20,000 (severe thrombocytopenia), 114 patients (9.93 %) had moderate thrombocytopenia (range: 20,000–50,000), 439 patients (38.24 %) had mild thrombocytopenia (range: 50,000–1,000,000) while the remaining 592 patients (51.57 %) had values equal to or more than 1000,000 (Figure 1 of the Supplementary material). Thrombocytopenia had a significant association with blood component transfusion (p-value <0.05).

Sixteen features were selected using the voting method (for details see Supplementary material) and used in each classifier to predict blood component transfusion during hospital stay. In the validation set, XGBoost and the random forest model achieved the same AUC (0.793) with the KNN classifier had the lowest AUC (0.703) (Figure 2). The XGBoost model significantly outperformed the seven other machine-learning models. The relatively balanced sensitivity, specificity, PPV, NPV and F1-score of the XGBoost model were 0.632, 0.953, 0.821, 0.885, and 0.714, respectively (Table 2). Confusion matrices, ROC curve, and the Precision-Recall curve for all the models are available in Figures 5-9 of the Supplementary material. Figure 3 shows the calibration plot for the prediction of blood component transfusion needs amongst the cases using the XGBoost model.

Figure 2.

ROC curves for the ML (XGBoost, GBDT, KNN, MLP, LR, and RF) models.

(0.25MB).

Figure 3.

Calibration plot of internal validation set for the XGBoost model. The predicted probability of blood component transfusion cases is compared to the actual proportion of blood component transfusion cases amongst subjects in the 30 % holdout set. Perfect predictive calibration ability is represented by the dashed diagonal line.

(0.09MB).

Of the 345 patients in the validation set, 301 patients were correctly classified. The SHAP summary plot demonstrated that severe thrombocytopenia, abdominal free fluid, longer hospital stays, lower sodium, and lower albumin ranked as the five most important features (Figure 4). This plot also revealed that severe thrombocytopenia, abdominal free fluid, longer hospital stays, lower sodium, and lower albumin were linked with higher SHAP value outputs in the XGBoost model indicating more likely to receive blood component transfusions (Figure 4). The SHAP summary plot of the random forest, GBDT, and ExtraTrees models also demonstrated that severe thrombocytopenia, abdominal free fluid, and longer length of hospital stay ranked amongst the three most important features, respectively (Supplementary material Figures 2-4).

Figure 4.

SHAP analysis of the test set using the XGBoost model. Each circular dot represents one patient. Features are listed on the y-axis in rank order with the uppermost being the highest contributor to the prediction model and the lowermost being the lowest contributor. The x-axis shows the range of SHAP values, a value of zero represents no contribution. Transfusion risk is encoded by its horizontal coordinates, where positive values indicate higher transfusion risks and negative values indicate lower transfusion risks.

(0.23MB).

Five examples of correctly classified cases (# 246, 261, 264, 268 and 297) are demonstrated as SHAP force and decision plots in Figure 5. The primary contributing variables for an individual's final model output are mostly shown in the force plot. The bars represent the SHAP value. The colours of the bars depict the consequence of the SHAP value (Positive or Negative). The prediction value is highlighted in black. The line graph shows the baseline. The instance value is also shown for each patient. The SHAP decision plots represent the decision path for every feature. The units shown on the x-axis are probabilities. The importance of each feature is listed in descending order. Each line on the plot strikes the x-axis at the predicted value of its corresponding observation. This shows how each feature contributes to the overall prediction.

Figure 5.

A: SHAP force plot for five patients (#: 246, 261, 264, 268 and 297). Features in pink tend to push the probability of need of blood component transfusion to the right, while features in green tend to push it to the left. Doctors can use this plot to easily identify the most important features with high decision power at the individual level. B: SHAP decision plot for the same five patients. This provides a better visualization of how all predictors affect each decision in terms of their importance. The decision path changes direction drastically with the highly important features and achieves an estimated probability of blood component transfusion needs. An individual physician can choose whether or not to trust the output based on how the features affect the direction.

(0.71MB).