A machine learning-based risk score for prediction of mechanical ventilation in children with dengue shock syndrome: A retrospective cohort study Authors: Nguyen Tat Thanh, Vo Thanh Luan, Do Chau Viet, Trinh Huu Tung, Vu Thien, Nilanka Perera, Nilanka Perera, Nilanka Perera Published: 2024 ABSTRACT BackgroundPatients with severe dengue who develop severe respiratory failure requiring mechanical ventilation (MV) support have significantly increased mortality rates. This study aimed to develop a robust machine learning-based risk score to predict the need for MV in children with dengue shock syndrome (DSS) who developed acute respiratory failure.MethodsThis single-institution retrospective study was conducted at a tertiary pediatric hospital in Vietnam between 2013 and 2022. The primary outcome was severe respiratory failure requiring MV in the children with DSS. Key covariables were predetermined by the LASSO method, literature review, and clinical expertise, including age (< 5 years), female patients, early onset day of DSS (≤ day 4), large cumulative fluid infusion, higher colloid-to-crystalloid fluid infusion ratio, severe bleeding, severe transaminitis, low platelet counts (< 20 x 109/L), elevated hematocrit, and high vasoactive-inotropic score. These covariables were analyzed using supervised models, including Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), k-Nearest Neighbor (KNN), and eXtreme Gradient Boosting (XGBoost). Shapley Additive Explanations (SHAP) analysis was used to assess feature contribution.ResultsA total of 1,278 patients were included, with a median patient age of 8.1 years (IQR: 5.4–10.7). Among them, 170 patients (13.3%) with DSS required mechanical ventilation. A significantly higher fatality rate was observed in the MV group than that in the non-MV group (22.4% vs. 0.1%). The RF and SVM models showed the highest model discrimination. The SHAP model explained the significant predictors. Internal validation of the predictive model showed high consistency between the predicted and observed data, with a good slope calibration in training (test) sets 1.0 (0.934), and a low Brier score of 0.04. Complete-case analysis was used to construct the risk score.ConclusionsWe developed a robust machine learning-based risk score to estimate the need for MV in hospitalized children with DSS. Introduction According to a recent report by the World Health Organization (WHO) in 2023, there has been an unprecedented increase in dengue cases across all WHO-defined regions [1]. Since the beginning of 2023, approximately 6.5 million dengue cases and over 7,300 dengue-related fatalities have been reported, marking a historic surge [1]. Critical complications associated with dengue have contributed to high fatality rates, varying from approximately 5% to over 20% [2–8]. Severe complications among hospitalized children with dengue include significant plasma leakage, dengue shock syndrome (DSS), severe bleeding, acute liver failure, and respiratory failure [2–8]. Most notably, extensive plasma leakage causes excessive fluid accumulation in pleural and abdominal cavities, resulting in mild-to-severe respiratory failure. Significantly higher mortality rates have been reported in severe dengue patients with acute respiratory failure and MV [8,9]. Our research group recently conducted a pediatric cohort study in which approximately a quarter of hospitalized children with DSS developed severe respiratory failure requiring MV [4]. Another adult dengue cohort in Taiwan showed that 35% of patients with severe dengue required MV support [10]. In addition, cumulative fluid infusion > 15% and obesity have been shown to be important predictors of severe respiratory failure and the need for MV in severe dengue cases [3,8]. Despite the high prevalence and associated mortality of mechanical ventilation in severe dengue, there remains a paucity of data on the prognostic factors for MV among children with DSS [8,9,11]. Therefore, identifying the determinants of MV in children with dengue shock syndrome upon admission to the pediatric intensive care unit (PICU) is crucial for this patient population. Notably, previous studies predicting the risk of death in hospitalized children with severe dengue have been limited by small sample sizes and a lack of robust statistical methodologies to validate predictive models [3–5]. Although our recent study, involving a cohort of 492 children with DSS, aimed to predict fatality, it was constrained by the small number of patients with fatal outcomes (26 deaths, 5.3%). Clinical evidence and literature review indicate that severe respiratory failure requiring MV is a strong independent predictor of mortality among children hospitalized with severe dengue [4,8–11]. Therefore, we used the need for mechanical ventilation as a surrogate outcome for mortality in children with DSS. Machine learning has recently gained prominence for its high performance in prediction and classification across various medical fields, particularly in tropical diseases [12–16]. This study aimed to identify the most significant predictors of MV and develop a robust machine learning-based risk-scoring system to predict the need for MV in hospitalized children with DSS. The identification of prognosticators can aid in optimizing management protocols, thereby reducing the need for MV support and improving survival outcomes in patients with DSS. Methods Ethics statement This study originated from the primary research titled "Prognostic model predicting mortality among patients presenting with dengue shock syndrome at Children’s Hospital No. 2, during 2013–2022," which was approved by the Scientific Committee and Institutional Review Board (IRB) of Children’s Hospital No. 2, Ho Chi Minh City, Vietnam (IRB No. 893/QD-BVND2 signed on 06-June-2022). A secondary dataset from the primary study was considered to pose less than minimal risk to participants. The requirement for informed consent was waived, and participants’ identities were de-identified to protect patient confidentiality. Data source, study setting and population This study utilized a dataset from the primary study, titled “Prognostic model predicting mortality among patients presenting with dengue shock syndrome at Children’s Hospital No.2, during 2013–2022”, which was approved by the Scientific Committee and Institutional Review Board (IRB) of the Children’s Hospital No.2, Ho Chi Minh City, Vietnam (IRB. No. 893/QD-BVND2 signed on 06-June-2022) [4]. The requirement for informed consent was waived, and participants’ identities were anonymized to ensure patient confidentiality. The study was performed in accordance with the principles of Good Clinical Practice and ethical guidelines of the Declaration of Helsinki. This study is reported in accordance with the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) guidelines (S1 Checklist). This single-institution retrospective study, known as the Vietnam Dengue-Infected Study (VNDIS), was conducted at the Children’s Hospital No. 2, a large tertiary pediatric hospital in Ho Chi Minh City, Vietnam. The hospital has a capacity of 1,400 beds and serves patients from central and southern Vietnam. The study period spanned from 2013 to 2022. The primary aims of this study were to identify the determinants of severe respiratory failure requiring MV and develop a machine learning-based risk scoring system for children with DSS. The eligibility criteria were age < 18 years, laboratory-confirmed dengue infection, and the presence of dengue shock syndrome [2]. The exclusion criteria were a lack of serologically confirmed dengue infection and missing data for variables of interest (≥ 50%). The indications for mechanical ventilation in children with severe DSS were based on the national guidelines for dengue diagnosis and treatment, as outlined in S1 File. Study definitions According to the World Health Organization (WHO) dengue guidelines (2009), laboratory-confirmed dengue infection was defined by the presence of dengue-IgM antibodies or a positive nonstructural 1 antigen test [2]. Dengue shock syndrome and severe bleeding were diagnosed according to the WHO dengue 2009 guidelines [2]. Severe transaminitis was defined as elevated aspartate aminotransferase (AST) and/or alanine aminotransferase (ALT) levels of ≥ 1,000 IU/L [2]. Study outcome The primary outcome of the study was the requirement for mechanical ventilation in children with dengue shock syndrome who developed acute respiratory failure during their PICU stay based on data collected upon PICU admission. Candidate variables From the original dataset, we identified 1,278 eligible participants with 82 variables [4]. We extracted 28 variables of interest that had a missing rate of < 50% and were consistent with the disease pathogenesis (S1 Table). These variables were considered the initial set of candidate variables and sorted into the following main categories: Demographic data Patient’s age, sex, presence of underlying diseases. Clinical data Severity of dengue shock, time of onset of dengue shock, severe bleeding, platelet transfusion, neurological status (AVPU scale), respiratory rate, systolic and diastolic shock indices, vasoactive inotropic score (VIS), cumulative volume of fluid infused (from the referral hospital and during 24h of PICU admission), and colloid-to-crystalloid fluid infusion ratio. Laboratory data Hematological parameters, liver enzymes, and organ dysfunction biomarkers. The percentages of the missing values for these variables are presented (S1 Fig). Feature selection and data preprocessing A predetermined set of clinical and laboratory covariables was determined using clinical knowledge and the Least Absolute Shrinkage and Selection Operator (LASSO) method. Handling missing data Missing data patterns were analyzed using the naniar, rms, VIM, and dlookr packages in R. Visualizations such as missing data patterns, Pareto charts, and hierarchical clustering were utilized to comprehend and manage missing values. Variables with more than 50% missing data were excluded from the statistical analysis. Multiple imputations were performed using the missRanger package to handle the missing values [17]. Data cleaning and transformation Categorical variables were converted into factors. Binary variables, such as the MV outcome, were converted into a numeric (0/1) format. Outliers and irrelevant categories were removed where applicable. Statistical analysis All the steps involved in developing the prediction model and risk score are detailed in Fig 1. The study flowchart for developing predictive models of mechanical ventilation in children with dengue shock syndrome. Model selection This study employed various supervised machine learning models, including Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), k-Nearest Neighbors (KNN), and eXtreme Gradient Boosting (XGBoost). Data splitting and preparation The dataset was randomly split into training (70%) and testing (30%) sets using stratified sampling to ensure a representative distribution of the outcome variable. One-hot encoding was applied to categorical variables and normalization was performed for numerical variables. Model training and hyperparameter tuning Ten-fold cross-validation was used to train the models, ensuring robustness, and preventing overfitting. Hyperparameter tuning was conducted using a grid search to optimize the model parameters. The best model for predicting severe respiratory failure requiring MV was selected based on performance metrics evaluated on the test set. Performance metrics Model performance was evaluated using accuracy, sensitivity, specificity, precision, area under the receiver operating characteristic curve (AUC), and F1 score. Model explanation The Shapley Additive Explanations (SHAP) method, based on the best performance supervised model, was employed to interpret the contribution of each feature to the model predictions. The SHAP values provided a unified measure for interpreting features of importance and their impact on the model’s output. Model validation and calibration, and risk scoring system We utilized the Hmisc and rms packages in the R library to develop the LR model to predict the need for mechanical ventilation in pediatric patients with DSS. Variables included in the validation and calibration were selected based on the SHAP results and clinical expertise. These variables comprised age, sex, severe bleeding, low platelet counts, severe transaminitis, peak hematocrit, cumulative volume of fluid from the referral hospital and 24h of PICU admission, the colloid-crystal fluid infusion ratio, early onset day of DSS, and vasoactive inotropic score. The predictive model was internally validated using bootstrap sampling (n = 500) and its performance was assessed using a calibration plot to visualize the consistency between the predicted and observed values. A risk-scoring system (Dengue MV score) was developed for predicting mechanical ventilation in DSS children using the ten most important features identified from the best-performing supervised models based on complete-case data analysis. Cutoff values for the final covariables in the risk score were determined using our clinical knowledge and the existing medical literature. All statistical analyses were performed using R (version 4.3.2) with packages, including rms, naniar, VIM, dlookr, missRanger, glmnet, caret, xgboost, SHAPforxgboost, and cutpointr. Statistical significance was set at P < 0.05. Results Baseline characteristics of study participants upon PICU admission Between 2013 and 2022, approximately 2,000 children with DSS were admitted to the PICU, of whom 1,278 met the eligibility criteria and were included in the analysis. Of these, 170 patients (13.3%) experienced severe respiratory failure requiring MV, while the remaining 1,108 patients (86.7%) did not require MV. The baseline clinical and laboratory characteristics of the study participants are presented in Table 1. Patients requiring MV were generally younger, with a higher proportion of female patients. The onset of dengue shock occurred earlier in the MV group, with a median of 4 days from the onset of illness. These patients also exhibited a greater DSS severity than those who did not require MV. Severe bleeding complications and elevated INR levels were more prevalent in the MV group. Patients in the MV group had higher systolic and diastolic shock indices and required more vasopressor support, as indicated by a significantly higher vasoactive inotropic score. Liver injury was more pronounced in the MV group, as evidenced by the significantly elevated levels of AST, ALT, and serum lactate. Severe transaminitis was more common in the MV group. Hematological parameters also differed between the groups; patients requiring MV had lower hemoglobin levels, peak hematocrit, nadir hematocrit, and significantly lower platelet counts. Respiratory distress was more pronounced in the MV group, as indicated by a higher respiratory rate, with marked differences in PCO2, PO2, and serum bicarbonate levels compared with the non-MV group. In terms of fluid management, the MV group received a greater volume of intravenous fluids both from the referral hospital and during the first 24 h of PICU admission. They also had higher colloid-to-crystalloid fluid infusion ratios. Notably, patients with DSS who required MV had worse outcomes, including significantly higher fatality rates and longer hospital stays, than those who did not require MV. 10.1371/journal.pone.0315281.t001Table 1Baseline clinical and laboratory characteristics of study participants on PICU admission and clinical outcomes at discharge (N = 1,278).CharacteristicsMechanical ventilation(n = 170)Non-mechanical ventilation(n = 1,108)p-valueAge (years)6.9 (4–9)8.4 (5.7–10.9)< 0.001Female patients (%)95 (56)524 (47)0.045Body mass index (kg/m2)18.4 (16–22)18.1 (15.5–21.4)< 0.01Underlying diseases (%)13 (7.7)67 (6.1)0.323Day of occurrence of dengue shocksince onset of fever (days)4 (4–5)5 (4–5)Grading of DSS severity (%) Compensated DSS Decompensated DSS137 (81)33 (19)1,034 (93)74 (7)< 0.001Severe bleeding (%)77 (45)18 (1.6)< 0.001Respiratory rate (/min)30 (25–35)25 (22–30)Systolic shock index (bpm/mmHg)1.45 (1.2–1.71)1.29 (1.11–1.44)< 0.001Diastolic shock index (bpm/mmHg)2.07 (1.71–2.67)1.65 (1.43–2.0)< 0.001White blood cell count (x 109/L)6.35 (4.2–10.3)4.6 (3.3–6.5)0.764Hemoglobin (g/dL)13.5 (11.8–15)15 (13.7–16.2)0.015Peak hematocrit (%)47 (41–50)48 (45–51)< 0.001Nadir hematocrit (%)36 (30–41)38 (35–41)< 0.001Platelet counts (x 109/L)30 (17.6–47)37 (24–56.8)0.014Aspartate aminotransferase (IU/L)588 (162–2020)140 (84–275)< 0.001Alanine aminotransferase (IU/L)253 (56–768)62 (33–132)< 0.001Severe transaminitis (%)68 (40)62 (5.6)< 0.001International normalized ratio (INR)1.85 (1.44–2.49)1.19 (1.09–1.37)< 0.001Serum lactate (mmol/L)3.2 (1.9–6.5)2.3 (1.6–3.0)< 0.001Serum creatinine (mmol/L)52 (42–66)52 (45–59)0.001Arterial blood gas analysis pH PCO2 (mmHg) PO2 with oxygen support (mmHg) Bicarbonate (mEq/L)7.39 (7.31–7.45)27.4 (21.7–33)138 (93–174)16 (13.8–18.6)7.44 (7.41–7.48)26 (22.1–29.7)106 (61–144)17.5 (15–19.7)0.320.01< 0.0010.03Cumulative fluid infused from referral hospitals and 24h PICU admission (mL/kg)258 (187–357)127 (101–163)< 0.001Colloid-to-crystalloid infusion ratio3.1 (1.1–6.9)0.6 (0–1.7)< 0.001Vasopressor support during first 24h (%)92 (54)10 (0.9)< 0.001Vasoactive inotropic score10 (0–30)0 (0–0)< 0.001Length of hospital stay (days)12 (9–18)4 (3–5)< 0.001Fatal outcome (%)38 (22.4)1 (0.1)< 0.001Summary statistics are presented as median (interquartile range, IQR) for continuous variables and frequency (%) for categorical variables. Two-sided, two-sample t-tests were used for the comparison of continuous variables, and the chi-square test (adjusted Fisher exact test) was used for categorical group comparisons. Abbreviations: DSS, dengue shock syndrome; PICU, pediatric intensive care unit. Baseline clinical and laboratory characteristics of study participants on PICU admission and clinical outcomes at discharge (N = 1,278). Summary statistics are presented as median (interquartile range, IQR) for continuous variables and frequency (%) for categorical variables. Two-sided, two-sample t-tests were used for the comparison of continuous variables, and the chi-square test (adjusted Fisher exact test) was used for categorical group comparisons. Abbreviations: DSS, dengue shock syndrome; PICU, pediatric intensive care unit. Performance of supervised models for mechanical ventilation in DSS patients Table 2 summarizes the performance of various supervised machine learning models employed to estimate the risk of MV in patients with dengue shock syndrome. The LR achieved an area under the curve of 0.949, with a sensitivity of 0.995, modest specificity of 0.597, precision of 0.941, F1 score of 0.732, and accuracy of 0.942. This LR model showed high sensitivity and accuracy, making it reliable for identifying true positives. However, its markedly lower specificity indicated a higher rate of false positives. RF had an AUC of 0.959, a sensitivity of 0.690, a specificity of 0.979, a precision of 0.841, an F1 score of 0.755, and an accuracy of 0.938. The RF model showed strong performance with high AUC, specificity, and accuracy. The SVM model had an AUC of 0.963, a sensitivity of 0.860, a specificity of 0.959, a precision of 0.725, an F1 score of 0.787, and an accuracy of 0.948. The SVM exhibited excellent performance with high sensitivity, specificity, and precision, making it a robust model for prediction. The highest AUC indicates that the SVM model has a strong discrimination power. The KNN model had an AUC of 0.794, a sensitivity of 0.686, a specificity of 0.902, a precision of 0.875, an F1 score of 0.769, and an accuracy of 0.794. The XGBoost achieved an AUC of 0.824, sensitivity, specificity, and accuracy of 0.824, precision and F1 score of 0.823. The XGBoost has demonstrated potential for predicting MV in pediatric patients with DSS. 10.1371/journal.pone.0315281.t002Table 2Performance of supervised models to estimate the risk of mechanical ventilation in patients with dengue shock syndrome (N = 1,278).ModelsAUCSensitivitySpecificityPrecisionF1 scoreAccuracyLR0.9490.9950.5970.9410.7320.942RF0.9590.6900.9790.8410.7550.938SVM0.9630.8600.9590.7250.7870.948KNN0.7940.6860.9020.8750.7690.794XGBoost0.8240.8240.8240.8230.8230.824AUC, Area under curve; LR, Logistic Regression; RF, Random Forest; KNN, K-Nearest Neighbors; SVM, Support Vector Machine; XGBoost, eXtreme Gradient Boosting. Performance of supervised models to estimate the risk of mechanical ventilation in patients with dengue shock syndrome (N = 1,278). AUC, Area under curve; LR, Logistic Regression; RF, Random Forest; KNN, K-Nearest Neighbors; SVM, Support Vector Machine; XGBoost, eXtreme Gradient Boosting. Variables of importance from the Random Forest model The RF model identified the most significant features contributing to the prediction of mechanical ventilation (MV), as shown in Fig 2. These features included the cumulative volume of fluid infused from the referral hospital and during the first 24 hours of PICU admission, a higher colloid-to-crystalloid fluid infusion ratio, low platelet counts (< 20 x 109/L), severe bleeding, vasoactive inotropic score (VIS) > 30, age < 5 years, female patients, and early onset of DSS. The ten most significant features of the Random Forest model based on classification accuracy. The Shapley Additive Explanations (SHAP) model The SHAP plot highlights the most important variables contributing to the risk of mechanical ventilation in hospitalized children with dengue shock syndrome (Fig 3). Most importantly, all predictors from the SVM model demonstrated the utmost significance in the SHAP model. Notably, the cumulative amount of infused fluid had the highest SHAP value, indicating that it was the most influential predictor of mechanical ventilation. A higher colloidal-to-crystalloid fluid infusion ratio was associated with an increased need for mechanical ventilation in patients with DSS. The vasoactive inotropic score reflects the intensity of vasoactive medications used to support hemodynamics and cardiac function. DSS patients with a high VIS (> 30) exhibited a markedly increased risk of requiring MV, indicating critical cardiovascular dysfunction. Severe bleeding, a serious complication of dengue, was also a significant predictor of MV. Other important contributors to the SHAP model included severe transaminitis, low platelet counts (< 20 x 109/L), and elevated INR, which were strongly associated with the need for mechanical ventilation. Increased serum lactate levels, indicative of severe metabolic acidosis and liver dysfunction, also predicted the requirement for MV. In addition, age was another significant factor, with younger children showing a higher risk of MV. Female patients were more likely to require MV support than male patients, making sex an important predictive factor. An increased respiratory rate upon admission, indicating respiratory distress, was predictive of MV. Additionally, early onset of DSS was identified as a risk factor for mechanical ventilation. Elevated hematocrit levels were found to be an important predictor in the SHAP model and were demonstrated to be a protective factor against mechanical ventilation. The SHAP interpretation of the clinically important variables in patients with dengue shock syndrome on mechanical ventilation. All ten features identified by the SVM model showed high significance in the SHAP model. However, the remaining variables had less impact on mechanical ventilation. Final predictive model, internal validation, and calibration The best-performing model for predicting MV, incorporating the ten most significant features from the SVM and SHAP models, was internally validated using complete-case analysis and the bootstrap method. As shown in Fig 4, the final predictive model demonstrated strong consistency between the predicted values and actual observations, indicating excellent calibration of both the slope and intercept, as well as a low Brier score (0.04). These results confirm that the final predictive model is robust in predicting severe respiratory failure requiring MV in pediatric patients with dengue shock syndrome. Calibration plot showing high consistency between the predicted values (x-axis) and actual observation data (y-axis). Risk score for mechanical ventilation among patients with DSS A complete-case analysis multivariable logistic model (n = 1,268 participants) utilizing the ten most significant features identified by the supervised models was used to construct a risk-scoring system (Table 3). The risk score was developed based on the subcoefficients of these ten predictors. The total points ranged from -2 to 46, corresponding to a probability of mechanical ventilation ranging from 0% to 100%. As shown in Fig 5, the risk of mechanical ventilation in hospitalized children with DSS increased proportionately with an elevated risk score. A simplified risk score chart for estimating the risk of mechanical ventilation in children with dengue shock syndrome. The total points ranged from -2 to 46, corresponding to a probability of mechanical ventilation ranging from 0% to 100%. 10.1371/journal.pone.0315281.t003Table 3Risk score to predict mechanical ventilation among children with dengue shock syndrome during the first 24 hours of PICU admission.PredictorsSubcategoryReference valuesSubcoefficientsSERisk scoreAge (years)≥ 5Reference0-0< 5-1.1430.3214Female patientsNoReference0-0Yes-0.8070.2903Cumulative amount of fluid infused (mL/kg)≤ 180Reference0-0> 180-2.6180.32010Colloid-to-crystalloid fluid infusion ratio< 1.6Reference0-0≥ 1.6-0.8730.2973Platelet counts(< 20 x 109/L)NoReference0-0Yes-0.2640.3291Severe bleedingNoReference0-0Yes-3.1710.41412+ 01-log2-increase in VIS during first 24h of admissionNoReference0-0Yes-1.1380.1484Severe transaminitisNoReference0-0Yes-1.7890.3857Onset day of shock> day 4Reference0-0≤ day 4-0.5100.2902Peak hematocrit (%)≤ 48%Reference0-0> 48%-- 0.4470.287-2 Range of risk score From -2 to 46Sub-coefficients derived from the slopes of the multivariable logistic model (complete-case analysis); Abbreviations: PICU, pediatric intensive care unit; SE, standard errors of coefficients; VIS, vasoactive inotropic score. Risk score to predict mechanical ventilation among children with dengue shock syndrome during the first 24 hours of PICU admission. Sub-coefficients derived from the slopes of the multivariable logistic model (complete-case analysis); Abbreviations: PICU, pediatric intensive care unit; SE, standard errors of coefficients; VIS, vasoactive inotropic score. Discussion Patients with severe dengue who experience respiratory failure and require mechanical ventilation have a significantly higher mortality rate [8–10]. High proportions of severe dengue patients with critical respiratory distress requiring mechanical ventilation support have been reported, ranging from approximately 25% to 35% in pediatric and adult cohorts, respectively [4,10]. In this study, 13.3% of the children with DSS required MV support, and a dramatically higher fatality rate was observed in the MV group than in the non-MV group (22.4% vs. 0.1%). Recognizing the prognostic indicators associated with MV in patients with DSS is crucial for optimizing the management protocols and improving survival outcomes. Although several studies have identified respiratory failure as a critical predictor of fatality in severe dengue [7,8,15,16], data comprehensively examining the determinants of MV support in DSS patients remain limited [3,8,18]. Previous research has identified several risk factors for MV in DSS, including female sex, obesity, large volume of intravenous fluid infusion, severe transaminitis and bleeding [3,8,18]. However, previous studies were constrained by small sample sizes or lacked robust statistical methods to determine the definitive prognosticators of MV [3,8,18–20]. Given the increasing popularity of machine learning owing to its robust modeling capabilities in various diseases [12–16], we employed supervised models to identify the key prognosticators of MV in children with DSS. Using predefined predictive covariates obtained from clinical knowledge and medical literature, we applied the LASSO method to prevent overfitting. The ten most significant features identified showed strong performance in the supervised models and SHAP analysis, aligning with previous reports [4,8,18–20]. The predictive factors identified in this study, including a large cumulative fluid infusion from the referral hospital and during the first 24 h of PICU admission, a high colloid-to-crystalloid infusion ratio, severe bleeding, and low platelet counts (< 20 x 109/L), particularly in patients requiring platelet transfusion, severe transaminitis, and an increased vasoactive inotropic score, were strong predictors of MV in patients with DSS. Additionally, younger age and female sex are associated with severe dengue, as reported in systematic reviews and meta-analyses [21,22]. Most importantly, dengue-related complications, including acute respiratory disease, severe transaminitis and bleeding, and a high VIS (> 30) have been shown to have high prognostic values for dengue-associated fatality [4,19,20]. Therefore, these data provide the basis for our predetermined covariates and the use of mechanical ventilation as a surrogate outcome for dengue-related mortality in patients with dengue shock syndrome. Notably, other covariates, including obesity, underlying diseases, and serum lactate levels, which have been reported as important risk factors for adverse clinical outcomes in patients with DSS, were not significant in LASSO and machine learning analyses [4,22]. Most significantly, excessive fluid resuscitation has a negative impact on clinical outcomes in PICU-admitted patients as well as in patients with severe dengue [5,23–25]. Each 1% excess in fluid resuscitation increases the risk of mortality by 6%, primarily by prolonging mechanical ventilation and inducing acute kidney injury [23]. Excess fluid resuscitation results in increased pressure in multiple body compartments, particularly the abdominal and thoracic cavities. This can lead to multiorgan failure due to prolonged, hypovolemic, and obstructive dengue shock when patients with severe DSS are resuscitated with a large volume of intravenous fluid [9,11]. Owing to the multiple adverse effects, macromolecular solutions are restrictively used in shock resuscitation, except for patients with DSS [26,27]. The use of colloidal solutions is crucial for patients with DSS when the initial resuscitation with crystalloid fluid shows a poor response. The judicious use of colloids is recommended, with a minimal amount of fluid infused to maintain hemodynamic stability. Notably, low serum albumin levels have been shown to be associated with progression to severe dengue; therefore, albumin solution should be administered to patients with severe DSS accompanied by unstable hemodynamic conditions [2,28,29]. In this study, an elevated colloid-to-crystalloid fluid infusion ratio (> 1.6) significantly predicted the need for mechanical ventilation. In clinical practice, a cut-off ratio of > 1.6 is a good indicator of the appropriate time point for switching from colloid solutions to more beneficial fluids such as albumin and fresh frozen plasma. Markedly, severe bleeding is a well-recognized predictor of adverse outcomes in DSS patients [4,30]. Severe transaminitis and coagulation dysfunction are frequently observed in patients with DSS. Therefore, colloid infusion in patients with severe DSS can potentially aggravate coagulation disorders and bleeding, thereby increasing the demand for blood product transfusions [27]. Furthermore, iatrogenic hemodilution in patients receiving large volumes of colloid solutions may result in a paradoxical decrease in oxygen delivery owing to a decrease in hemoglobin concentration, which can increase unnecessary transfusion of blood products [31]. Another point to be highlighted is that elevated hematocrit is associated with hemoconcentration in patients with DSS, and fluid supplementation is necessary in this regard [2]. Notably, the most important goal of treating DSS is to infuse a minimum amount of intravenous fluid to ensure adequate tissue perfusion and urine output of > 0.5 ml/kg/h [32]. When patients with elevated hematocrit levels present with good hemoperfusion, the excessive administration of fluid targeted at reducing and/or maintaining low hematocrit levels may lead to poor clinical outcomes. Importantly, increased hematocrit levels have been shown to be a protective factor in reducing the risk of MV [3]. Therefore, emphasis is placed on sustaining a minimal amount of fluid to achieve good hemoperfusion of vital organs and maintain urine output > 0.5 ml/kg/h. In the presence of signs of adequate hemoperfusion, reducing the rate of intravenous fluid infusion is recommended [32]. The final prediction model was developed based on the ten most important features identified by the LASSO and machine learning models. On this basis, we constructed a robust risk-scoring system that is helpful for clinicians in the PICU for the early recognition of DSS patients requiring mechanical ventilation support. The total risk score ranged from minus 2 to 46 points. Notably, an elevated hematocrit > 48% is considered a protective factor, corresponding to a deduction of 2 points in the risk-scoring system. As shown in Fig 5, the risk score chart showed that DSS patients with a risk score of 14–15 had roughly 10% progression to respiratory failure requiring MV support. A higher risk score indicates an increased probability of MV support. Hence, DSS patients with a risk score of 21 had a 50% probability of receiving MV support, and those with a risk score of 38–39 had a 99% probability of receiving MV. Given the proportional relationship between MV requirement and mortality risk in hospitalized children with DSS, this risk-scoring system is a valuable tool for improving clinical management. Further external and prospective validation in other cohorts will enhance the generalizability of these findings. Dengue shock syndrome has been reported to have fatality rates ranging from 5.3% to 25.6% in all-type DSS-experiencing children admitted to the PICU [4,5]. Similarly, DSS patients who develop critical respiratory failure account for markedly elevated mortality rates in pediatric and adult cohort studies [6,8,9]. The combination of these dengue-associated complications has predominantly contributed to significantly increased fatalities among hospitalized dengue patients, as noted in our recent cohort of severely prolonged DSS patients undergoing mechanical ventilation (death rate of approximately 36%) [9]. Mechanical ventilation is a frequently observed complication among critically ill patients in the late critical stage of severe dengue infection [11]. Therefore, this study has great clinical significance. Rapid identification of patients with high-risk DSS for MV support at the time of PICU admission could help prioritize interventions that can significantly enhance clinical and survival outcomes. This study has several strengths. The predictive model for mechanical ventilation was based on a relatively large sample of hospitalized children with DSS. The most significant features were well-defined using profound clinical knowledge and robust machine learning methods. Internal validation of the final predictive model demonstrated high discrimination and calibration. However, this study had several limitations inherent to its single-center retrospective cohort design and unstandardized collection of clinical and laboratory data during hospital admission. The most significant bias was the missing information on serum lactate levels, previously reported as a prognostic indicator of death. Nevertheless, this variable was appropriately managed using the multiple imputation method. Conclusion We developed robust machine learning-based models to estimate the risk of mechanical ventilation in hospitalized children with dengue shock syndrome. The internal validation of the predictive model demonstrated high performance. Based on the most significant features identified at the time of admission, we constructed a simple risk-scoring system. This scoring system is a valuable tool for clinicians, aiding in the bedside management of DSS patients. Supporting information TRIPOD checklist. (DOCX) Indications for mechanical ventilation support in children with severe DSS. (DOCX) Data variables used in the machine learning modelling. (DOCX) The Pareto chart of missing values of the initial set of 28 candidate variables. (TIF) 9 Sep 2024 PONE-D-24-33301A machine learning-based risk score for prediction of mechanical ventilation in children with dengue shock syndrome: A retrospective cohort studyPLOS ONE Dear Dr. Nguyen, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. ============================== ACADEMIC EDITOR: 1. Please ensure that extensive language editing is performed prior to submission of the revised manuscript.2. Authors use medical terms that are not correct in many places eg: severe hepatic transaminases - this should be severe transaminitis3. It is not clear whether this risk score is to be used for patients with respiratory failure or all DSS. clarify.4. Inclusion criteria is not clear. - Was the model performed on all DSS, all DHF or DHF/DSS with respiratory failure? - Were they laboratory-confirmed dengue cases? 5. What were the exclusion criteria? Did you exclude co-infection causing pneumonia and respiratory failure? 7. Conclusion of the abstract – vague. To manage patients with respiratory failure and predict a poor outcome? Is it actually to predict MV in all severe dengue? ============================== Please submit your revised manuscript by Oct 24 2024 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript:A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols. We look forward to receiving your revised manuscript. Kind regards, Nilanka Perera, MD, PhD Academic Editor PLOS ONE Journal requirements: When submitting your revision, we need you to address these additional requirements. 1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer #1: Yes Reviewer #2: Yes ********** 2. Has the statistical analysis been performed appropriately and rigorously? Reviewer #1: No Reviewer #2: I Don't Know ********** 3. Have the authors made all data underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer #1: Yes Reviewer #2: Yes ********** 4. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer #1: Yes Reviewer #2: No ********** 5. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer #1: I have the following feedback on the write-up. The authors have applied several machine learning models to a real world dataset. Out of many variable, a selected set of variables were chosen for the analysis. -There are some grammatical errors and typos. Read the manuscript thoroughly and correct them. -In the abstract, it says SHAP model explained the significant clinical predictors. However, authors need to mention the machine learning model they used with SHAP model, eg: RF or SVM etc. -Authors used variable importance plot with RF model and SVM or RF (not clear) with SHAP to identify the importance features. Authors need to mention why they considered two methods, or why they preffered SHAP method over variable importance given by RF. -In line 292, Final predictive model, internal validation and calibration section, it's better to include the name of the final model (RF or SVM etc). -In line 233, authors mentioned that Logistic regression (LR) model has higher rate of false positives. Then they developed a risk score for mechanical ventilation among patients with DSS using LR model in line 304 under "Risk score for mechanical ventilation among patients with DSS" section, which is not justifiable. In this study, there's a significant class imbalance between MV (170) and non-MV groups (1108). Authors can try using class balancing techniques to see if they can further improve the precision and recall along with the accuracies. In addition to that, LASSO models work best for variable selection when there's multicollinearity and the usual practice is to check for multicollinearity before applying LASSO models.Out of many variable, a selected set of variables were chosen for the analysis using some ad-hoc method and it is recommended to use a proper variable selection method (from all non-missing variables) Reviewer #2: This is an interesting analysis that attempts to incorporate artificial intelligence into clinical practice. However, the flow of information presented in the text can be improved. I would like to suggest the following. Introduction section 1. The authors should verify the objective of this study. In lines 106 and 127, different aims have been mentioned. Methods section 2. Contradictory information on the study population. Please verify ( refer to line 129). 3. It is unclear how participants were recruited from the database and who was eligible for this particular study. Please explain. 4. Out of 82 variables in eligible participants, 28 variables were initially selected for analysis. Describe the rationale for selecting only some variables for analysis (lines 139 & 140). 5. Was there a uniformity of criteria used for starting mechanical ventilation at the PICU? The authors should clarify this. If different clinicians used different criteria, there is a potential for bias in this analysis. 6. Data on participant numbers should also be included in the methods section (refer lines 203 & 204). Results section 1. The authors have given general statements in the results section. For example “ patients who required MV were generally younger and more females than males…”. Please describe in terms of mean or median age and sex percentage (refer line 206). 2. Please define the terms used in the text – critical bleeding, severe transaminitis, etc. 3. How did the authors grade the severity of dengue infection? 4. Table 1 • The naming of the 2nd and 3rd columns is confusing. It sounds as if one group received a form of non-mechanical ventilation. Therefore, name them appropriately. For example, MV required and MV not required. • Mention the statistical tests used to derive p value in table 1. 5. Please describe how sub-coefficients were used to derive a risk score (refer line 307). 6. Explain how to use risk score practically. 7. Table 3 • Define severe bleeding and severe transaminitis in the main text as well. ( There are defined in supporting documents) • It is necessary to decide on a cut-off value for severe transaminitis to derive the risk score. Discussion section 1. In Table 3, it is mentioned thrombocytopenia is a risk factor but, in the text, it is said thrombocytopenia needing platelet transfusions is a risk factor. Verify if thrombocytopenia per se or thrombocytopenia requiring platelet transfusions was a risk factor for MV. 2. Have the platelet transfusions caused respiratory failure by fluid overload? 3. Investigating if blood transfusions were a risk factor for MV may be interesting. 4. Previously identified risk factors (such as obesity and lactate levels) for adverse outcomes for DSS were not significant in your analysis. Do you have any explanation for this difference? (lines 359 to 362). 5. Line 374 & 375 “Albumin solution should be administered to patients with severe DSS accompanied with unstable hemodynamic conditions” – Is this evidence-based? 6. Do you have any explanation for why increased hematocrit is a protective factor for MV? 7. Validation of this score with prospective studies is necessary before recommending this score in clinical practice (line 408). General other comments 1. Different units for platelet counts have been used. Please stick to one unit. 2. Some spelling and grammar mistakes were noted, and language editing is recommended. ********** 6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy. Reviewer #1: No Reviewer #2: No ********** [NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step. 19 Sep 2024 Response letter to reviewers Dear Editors-in-Chief of the PLOS ONE, We are very thankful for your insightful comments for our submitted manuscript. MANUSCRIPT # PONE-D-24-33301, entitled "A machine learning-based risk score for prediction of mechanical ventilation in children with dengue shock syndrome: A retrospective cohort study" We are thankful for all peer-review points. These are our responses to all the reviewers’ comments. ACADEMIC EDITOR: 1. Please ensure that extensive language editing is performed prior to submission of the revised manuscript. Response: Yes, we are thankful for your comments. We have checked the language editing in the revision manuscript and resubmitted the latest version as appropriate. 2. Authors use medical terms that are not correct in many places eg: severe hepatic transaminases - this should be severe transaminitis Response: We are thankful for your comments, and appropriate amendments have been made to the revised manuscript from your suggestions. 3. It is not clear whether this risk score is to be used for patients with respiratory failure or all DSS. clarify. Response: We would like to clarify that this risk score of mechanical ventilation (main study outcome) will be used for all patients with dengue shock syndrome (DSS) who were admitted to the PICU. Among the DSS patients, a certain proportion of children with DSS (170/1278, 13.3%) developed severe respiratory failure that required MV support. This information has been included in the revision manuscript. 4. Inclusion criteria is not clear. - Was the model performed on all DSS, all DHF or DHF/DSS with respiratory failure? - Were they laboratory-confirmed dengue cases? Response: We thank for your comment, and we have made appropriate amendments as follows: Lines 128-129 in the revision manuscript. “Eligibility criteria were defined as age below 18 years, laboratory-confirmed dengue infection, and the presence of dengue shock syndrome.” It is notable the denominator (population) are all PICU-admitted children had laboratory-confirmed dengue infection and DSS, with / without respiratory distress on PICU admission. 5. What were the exclusion criteria? Did you exclude co-infection causing pneumonia and respiratory failure? Response: We are thankful for your comments and have made clarification and amendment as appropriate. We have made amendments to clarify this point as follows: “The exclusion criteria were a lack of serologically confirmed dengue infection and missing data for variables of interest ( ≥ 50%).” Lines 129-130, in the revision manuscript. In this study, we aimed to increase the study representativeness and generalizability of the prognostic model. When we developed a baseline prognostic model for mechanical ventilation during 24h of PICU admission, we did not exclude co-infection-causing pneumonia related to mechanical ventilation. Because the ventilator-associated pneumonia (VAP) commonly occurs ≥ 48h -72h after PICU admission (relatively far beyond the baseline timepoint-24h PICU admission) Additionally, we did not exclude patients with DSS who experienced respiratory failure when they were admitted to the PICU. This is because all referred DSS patients were commonly treated previously within 12h-24h from the province hospital and then transferred to our hospital. Some patients might have respiratory failure and others did not. The most common cause of respiratory failure in the referred DSS patients was due to fluid overload and other dengue-related complications, such as severe bleeding, acute liver failure. Therefore, with these approaches, we could increase the representativeness and generalizability. 7. Conclusion of the abstract – vague. To manage patients with respiratory failure and predict a poor outcome? Is it actually to predict MV in all severe dengue? Response: We are thankful for your comments. We have made appropriate amendments for clarification, as follows: “We developed a robust machine learning-based risk score to estimate the need for MV in hospitalized children with DSS.” Lines 60-61, in the revision manuscript. Reviewer #1: I have the following feedback on the write-up. The authors have applied several machine learning models to a real world dataset. Out of many variable, a selected set of variables were chosen for the analysis. -There are some grammatical errors and typos. Read the manuscript thoroughly and correct them. Response: We are thankful for your comments, and we will make appropriate amendments. -In the abstract, it says SHAP model explained the significant clinical predictors. However, authors need to mention the machine learning model they used with SHAP model, eg: RF or SVM etc. Response: Thank for your comment. In this study, we utilized several supervised machine learning models, including Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), k-Nearest Neighbors (KNN), and Extreme Gradient Boosting (XGBoost), to predict the need for mechanical ventilation in pediatric patients with dengue shock syndrome. Upon evaluating these models, we found that SVM and RF models demonstrated very good overall performance (AUC, Sensitivity, Specificity, Precision, F1 score, and Accuracy). We based the optimal performance metrics on determining the SVM model to be additionally analyzed with SHAP. We add this information in the revised manuscript, Lines 298-300 - Authors used variable importance plot with RF model and SVM or RF (not clear) with SHAP to identify the importance features. Authors need to mention why they considered two methods, or why they preferred SHAP method over variable importance given by RF. Response: Currently, there are various supervised models, and it is well noted that under different sampling conditions that one supervised model may be better than other models and vice versa. - Balancing all performance metrics, RF and SVM showed the two best performance with AUC of 0.96 and high levels of precision and accuracy. - In addition, the SVM demonstrated slightly better performance than the RF model, in terms of higher sensitivity. Therefore, we finally determined the SVM as the best performing model. However, the RF model is also considered competitive. Additionally, SHAP analysis was employed to provide a more detailed and interpretable explanation of how each feature influenced individual predictions. SHAP was preferred because it allows for a consistent and individualized explanation of feature importance across multiple models, offering more granular insights. We applied the SHAP methodology based on the best supervised model (SVM) for studying the explanation of the fully predefined covariables (also including the 10 most important features identified). In the SHAP model, the 10 features from the SVM model outperformed the remaining risk factors for MV. Additionally, these are examples to illustrate the point that SHAP values can provide global insights into the feature importance and local explanations of individual predictions of each covariate. SHAP for Age: As can be seen from this chart, patients with DSS aged < 5 years have an increased risk of MV. The SHAP model can provide us with the distribution of age variable and an explanation for age in the overall prediction of outcome (MV). Further example for SHAP model with hematocrit (HCT) variable in this study: It can be seen from this SHAP plot that HCT > 48% corresponds to low SHAP values, indicating having a lower risk of MV. => Therefore, with all these supported points, the SHAP model was preferred over other supervised models (SVM, RF, KNN,….). -In line 292, Final predictive model, internal validation and calibration section, it's better to include the name of the final model (RF or SVM etc). Response: We are thankful for your comments. We used the ten most important predefined covariates in the final SVM model based on its highest performance, which was further analyzed with internal validation and calibration. We have clarified these points as you suggested, in the revision. Lines 273-274, in the revision manuscript “All the predictors from the SVM model demonstrated the utmost significance in the SHAP model” Lines 298-300, in the revision manuscript “The best performance model for MV, with the ten most significant features from the SVM and SHAP models, was internally validated using complete-case analysis and the bootstrap method.” -In line 233, authors mentioned that Logistic regression (LR) model has higher rate of false positives. Then they developed a risk score for mechanical ventilation among patients with DSS using LR model in line 304 under "Risk score for mechanical ventilation among patients with DSS" section, which is not justifiable. Line 233: “LR achieved an area under the curve (AUC) of 0.949, with a sensitivity of 0.995, modest specificity of 0.597, precision of 0.941, F1 score of 0.732, and accuracy of 0.942.” Line 304: “These results confirm that the final predictive model performs well in predicting severe respiratory failure requiring MV among pediatric patients with dengue shock syndrome.” Response: We thank for your comments. We will answer from point to point. - We acknowledge that the logistic regression (LR) model has modest specificity (or a higher rate of false positives). This was mainly because of non-linearity of data (skewed data) in this study. Most of the continuous covariables were not normally distributed. We also tried several conventional methods for data normalization, including categorization and log-transformation. This is a well-known limitation of the conventional logistic analysis. - However, machine learning supervised models can overcome this weakness from conventional analysis when manipulating with non-linearity data. With the similar 10 preselected covariates (used in the LR model above), the supervised machine learning models showed very good performance. In particular, the Support Vector Machine (SVM) is well-known for its capacity to analyze the non-linearity data. Furthermore, we additionally applied the SHAP methodology based on the best supervised model (SVM) for studying the explanation of the fully predefined covariables. - This is an example illustrating analyzing the peak hematocrit (HCT) with SHAP model: - It can be seen from this SHAP plot that the HCT is markedly non-linear, clearly observed HCT < 35% and HCT > 54%. Similarly, non-linearity was also significant, regarding covariables of age, INR, and vasoactive inotropic score (VIS). - Considering the predictive model performance evaluation, we used the criteria developed by Steyerberg EW et al. [1,2]. Key criteria for a well-performance model include (1) discrimination (C-statistic or AUC, discrimination slope), (2) Calibration in-the-large, calibration slope, (3) overall performance (Brier score), (4) reclassification and (5) clinical usefulness. Therefore, the specificity parameter is only one contributing indicator for assessing a predictive model, and it is not the main determinant for the evaluation of the model performance [1-5]. - In this study, despite modest specificity of LR model, the final C-statistic of LR model was still high (AUC = 94.9%), as well as precision and accuracy (of approximately 94%). Most importantly, the final predictive model (based on the 10 predefined covariables) for MV showed very good discrimination, calibration of slope, calibration plot, low Brier score as well as the clinical usefulness of the developed model [1,2]. In this study, a high consistency between predicted and observed data was prominently observed (as shown in the Figure 4). - We constructed a risk score based on logistic regression with complete-case analysis, considering its simplicity and interpretability in clinical practice. Logistic regression provides a straightforward method for computing risk scores that can be easily applied in clinical settings. Notably, we developed the risk score based on references from the Framingham risk score development, and clinical prediction model development and validation methods by Steyerberg EW and Harrell FE Jr [1-11]. References 1. Steyerberg EW, Vickers AJ, Cook NR, Gerds T, Gonen M, Obuchowski N, Pencina MJ, Kattan MW. Assessing the performance of prediction models: a framework for traditional and novel measures. Epidemiology. 2010 Jan;21(1):128-38. doi: 10.1097/EDE.0b013e3181c30fb2. PMID: 20010215; PMCID: PMC3575184. 2. Steyerberg EW, Vergouwe Y. Towards better clinical prediction models: seven steps for development and an ABCD for validation. Eur Heart J. 2014 Aug 1;35(29):1925-31. doi: 10.1093/eurheartj/ehu207. Epub 2014 Jun 4. PMID: 24898551; PMCID: PMC4155437. 3. Steyerberg EW, Eijkemans MJ, Harrell FE, Jr, Habbema JD. Prognostic modeling with logistic regression analysis: in search of a sensible strategy in small data sets. Med Decis Making. 2001;21:45–56. 4. Steyerberg EW, Harrell FE, Jr, Borsboom GJ, Eijkemans MJ, Vergouwe Y, Habbema JD. Internal validation of predictive models: efficiency of some procedures for logistic regression analysis. J Clin Epidemiol. 2001;54:774–781. 5. Steyerberg EW, Borsboom GJ, van Houwelingen HC, Eijkemans MJ, Habbema JD. Validation and updating of predictive logistic regression models: a study on sample size and shrinkage. Stat Med. 2004;23:2567–2586. 6. Harrell FE Jr, Lee KL, Califf RM, Pryor DB, Rosati RA. Regression modelling strategies for improved prognostic prediction. Stat Med. 1984 Apr-Jun;3(2):143-52. doi: 10.1002/sim.4780030207. PMID: 6463451. 7. Harrell FE. Regression Modeling Strategies: With Applications to Linear Models, Logistic Regression, and Survival Analysis. New York: Springer; 2001. 8. Harrell FE Jr. Regression Modelling Strategies. Available from: https://www.r-project.org/nosvn/conferences/useR-2007/tutorials/harrell.html 9. Steyerberg EW, Eijkemans MJ, Harrell FE, Jr, Habbema JD. Prognostic modelling with logistic regression analysis: a comparison of selection and estimation methods in small data sets. Stat Med. 2000;19:1059–1079. 10. Lloyd-Jones DM, Wilson PW, Larson MG, Beiser A, Leip EP, D'Agostino RB, Levy D. Framingham risk score and prediction of lifetime risk for coronary heart disease. Am J Cardiol. 2004 Jul 1;94(1):20-4. doi: 10.1016/j.amjcard.2004.03.023. PMID: 15219502 11. Berry JD, Lloyd-Jones DM, Garside DB, Greenland P. Framingham risk score and prediction of coronary heart disease death in young men. Am Heart J. 2007 Jul;154(1):80-6. doi: 10.1016/j.ahj.2007.03.042. PMID: 17584558; PMCID: PMC2279177. In this study, there's a significant class imbalance between MV (170) and non-MV groups (1108). Authors can try using class balancing techniques to see if they can further improve the precision and recall along with the accuracies. In addition to that, LASSO models work best for variable selection when there's multicollinearity and the usual practice is to check for multicollinearity before applying LASSO models. Out of many variable, a selected set of variables were chosen for the analysis using some ad-hoc method and it is recommended to use a proper variable selection method (from all non-missing variables). Response: The proportion of MV among the DSS population was 170 / 1278 (13.3%), which is considered to have relatively sufficient statistical power for predictive model development. Importantly, this figure (MV prevalence of 13.3% among DSS) truly reflects the real-life data of patients with dengue shock syndrome, that is consistent with the previously published data [1]. Therefore, we do not think that the MV imbalance is considerable in this study. We conservatively believe that it is not necessary to perform oversampling methods (SMOTE...etc). Also, we did not find any multicollinearity when analyzing the LASSO. Additionally, all covariates were independent. We recognized that LASSO seemed to be too stringent in variable selection, and eliminated some significant variables that are well known in disease pathogenesis and clinical importance. As stated in the manuscript, we combined and balanced different methods of LASSO, clinical knowledge, and the medical literature for preselecting the covariates. Reference [1] Submitted filename: Response to Reviewers.docx 25 Nov 2024 A machine learning-based risk score for prediction of mechanical ventilation in children with dengue shock syndrome: A retrospective cohort study PONE-D-24-33301R1 Dear Dr. Nguyen, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice will be generated when your article is formally accepted. Please note, if your institution has a publishing partnership with PLOS and your article meets the relevant criteria, all or part of your publication costs will be covered. Please make sure your user information is up-to-date by logging into Editorial Manager at Editorial Manager® and clicking the ‘Update My Information' link at the top of the page. If you have any questions relating to publication charges, please contact our Author Billing department directly at authorbilling@plos.org. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org. Kind regards, Nilanka Perera, MD, PhD Academic Editor PLOS ONE Additional Editor Comments (optional): Reviewers' comments: Reviewer's Responses to Questions Comments to the Author 1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer #1: All comments have been addressed ********** 2. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer #1: Yes ********** 3. Has the statistical analysis been performed appropriately and rigorously? Reviewer #1: Yes ********** 4. Have the authors made all data underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer #1: Yes ********** 5. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer #1: Yes ********** 6. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer #1: Thank you for addressing the comments. You have given clear explanations to all of them and amended the manuscript accordingly. ********** 7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy. Reviewer #1: No ********** 27 Nov 2024 PONE-D-24-33301R1 PLOS ONE Dear Dr. Thanh , I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now being handed over to our production team. At this stage, our production department will prepare your paper for publication. This includes ensuring the following: * All references, tables, and figures are properly cited * All relevant supporting information is included in the manuscript submission, * There are no issues that prevent the paper from being properly typeset If revisions are needed, the production department will contact you directly to resolve them. If no revisions are needed, you will receive an email when the publication date has been set. At this time, we do not offer pre-publication proofs to authors during production of the accepted work. Please keep in mind that we are working through a large volume of accepted articles, so please give us a few weeks to review your paper and let you know the next and final steps. Lastly, if your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org. If we can help with anything else, please email us at customercare@plos.org. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Dr. Nilanka Perera Academic Editor PLOS ONE