Machine learning-based models for prediction of in-hospital mortality in patients with dengue shock syndrome

Table 1 Baseline clinical and laboratory characteristics of study participants on pediatric intensive care unit admission (n = 1278), n (%)/median (interquartile range)

Characteristics	Non-survivor (n = 39)	Survivors (n = 1239)	P value
Age (years)	7 (5-10)	8.1 (5.4-10.7)	0.16
Female sex	18 (46)	601 (49)	0.77
Underlying diseases	7 (18)	73 (6)	< 0.01
Day of occurrence of dengue shock since onset of fever (days)	4 (4-5)	5 (4-5)	0.09
Grading of DSS
Compensated DSS	11 (28)	1143 (92)	< 0.001
Decompensated DSS	28 (72)	96 (8)
Severe bleeding	29 (74)	66 (5.3)	< 0.001
Respiratory rate (breaths/min)	30 (24-40)	25 (22-30)	< 0.001
Systolic shock index (bpm/mmHg)	1.47 (1.33-1.88)	1.3 (1.11-1.5)	< 0.01
White blood cell count (× 109/L)	7.15 (3.86-12.9)	4.74 (3.36-6.7)	0.93
Hemoglobin (g/dL)	13.8 (11.5-15.1)	14.9 (13.5-16.1)	0.01
Peak hematocrit (%)	46 (42-52)	48 (45-51)	0.07
Nadir hematocrit (%)	33 (29-34)	38 (35-41)	0.001
Platelet cell count (× 109/L)	25 (14-42.4)	36 (23-55)	0.21
Platelet transfusion	34 (87)	129 (10.4)	< 0.001
Aspartate aminotransferase (IU/L)	2219 (572-4458)	148 (86-314)	< 0.001
Alanine aminotransferase (IU/L)	724 (324-1670)	63 (34-149)	< 0.001
Severe transaminitis	25 (64)	105 (9)	< 0.001
International normalized ratio	2.52 (1.83-3.57)	1.21 (1.1-1.47)	< 0.001
Blood lactate (mmol/L)	7.2 (3.1-13)	2.3 (1.6-3.1)	< 0.001
Serum creatinine (mmol/L)	65 (56-111)	51 (44-59)	0.001
Cumulative fluid infused from referral hospitals and 24 h of PICU admission (mL/kg)	301 (220-393)	133 (104-176)	< 0.001
Ratio of colloid-to-crystalloid infusion	6.0 (1.7-9.1)	0.7 (0-2.0)	< 0.001
Vasoactive inotropic score during first 24 h	60 (40-165)	0 (0-15)	< 0.001
Vasoactive inotropic score > 30	30 (79)	17 (13)	< 0.001

DSS: Dengue shock syndrome; PICU: Pediatric intensive care unit.

Table 2 Multivariable models for in-hospital mortality of patients with dengue shock syndrome upon pediatric intensive care unit admission

Candidate predictors	Full model1		Reduced model 12		Reduced model 23
	OR (95%CI)	P value	OR (95%CI)	P value	OR (95%CI)	P value
Age (years)	0.85 (0.71-1.02)	0.09	0.93 (0.8-1.08)	0.32	0.94 (0.81-1.09)	0.39
Underlying diseases	5.14 (1.22-21.7)	0.03			4.1 (1.02-16.3)	0.04
Systolic shock index (mmHg/bpm)	0.46 (0.08-2.48)	0.36
Severe bleeding	4.28 (1.37-13.4)	0.01	3.38 (1.19-9.57)	0.02	3.38 (1.2-9.58)	0.02
Severe transaminitis	1.16 (0.4-3.38)	0.78
International normalized ratio	1.56 (1.12-2.17)	< 0.01	1.45 (1.05-1.99)	0.02	1.47 (1.07-2.01)	0.01
Peak hematocrit (%)	1.08 (1.0-1.16)	0.05
Platelet counts (< 20 × 10 9/L)	0.95 (0.32-2.85)	0.92
Platelet transfusion	4.48 (1.11-18.1)	0.03	4.1 (1.07-15.7)	0.03	4.89 (1.27-18.9)	0.02
Blood lactate (mmol/L)	1.2 (1.02-1.41)	0.03	1.2 (1.04-1.39)	0.01	1.19 (1.03-1.37)	0.02
Serum creatinine (µmol/L)	1.02 (1.0-1.03)	0.01	1.01 (1-1.03)	0.09	1.01 (1.0-1.03)	0.06
Log-2 Cumulative fluid infused from referral hospitals and within 24 h-PICU admission (mL/kg)4	2.76 (1.25-6.1)	0.01	3.69 (1.95-7.0)	< 0.001	3.61 (1.88-6.91)	< 0.001
Colloid to crystalloid fluid infusion ratio	1.07 (0.98-1.17)	0.15
Vasoactive inotropic score (> 30)	9.58 (2.36-39)	< 0.01	4.78 (1.49-15.4)	< 0.01	4.59 (1.42-14.8)	0.01

¹Full model based on predefined covariates with multiple imputation data analysis.

²Reduced model 1 based on Least Absolute Shrinkage and Selection Operator method.

³Reduced model 2 based on stepwise selection with the lowest value of Akaike information criterion.

⁴Standardization of covariates by Log-2 transformation.

DSS: Dengue shock syndrome; OR: Odds ratio; PICU: Pediatric intensive care unit.

Table 3 Performance of supervised models to estimate the risk of mortality in patients with dengue shock syndrome without the synthetic minority oversampling technique

Models	AUC (95%CI)	Sensitivity	Specificity	Precision	F1 score	Accuracy
Logistic regression	0.92 (0.90-0.94)	0.99	0.50	0.98	0.61	0.97
Naïve Bayes	0.86 (0.70-1)	0.91	0.82	0.83	0.87	0.86
Random forest	0.86 (0.69-1)	0.82	0.91	0.90	0.86	0.86
KNN	0.86 (0.71-1)	0.82	0.91	0.90	0.86	0.86
Decision tree	0.77 (0.59-0.94)	0.82	0.73	0.75	0.78	0.77
XGBoost	0.86 (0.71-1)	0.82	0.91	0.90	0.86	0.86

AUC: Area under the curve; KNN: K-nearest neighbors; SMOTE: Synthetic minority oversampling technique; XGboost: Extreme gradient boosting.

Table 4 Performance of supervised models to estimate the risk of mortality in patients with dengue shock syndrome from oversampling data with the synthetic minority oversampling technique

Models	AUC (95%CI)	Sensitivity	Specificity	Precision	F1 score	Accuracy
Logistic regression	0.93 (0.89-0.97)	0.94	0.93	0.77	0.85	0.93
Naïve Bayes	0.94 (0.91-0.97)	0.95	0.94	0.80	0.87	0.94
Random forest	0.97 (0.95-0.99)	0.98	0.97	0.88	0.93	0.97
KNN	0.91 (0.87-0.95)	0.92	0.89	0.69	0.78	0.90
Decision tree	0.95 (0.92-0.98)	0.95	0.96	0.85	0.90	0.96
XGBoost	0.97 (0.95-0.99)	0.98	0.96	0.86	0.92	0.96

AUC: Area under the curve; KNN: K-nearest neighbors; SMOTE: Synthetic minority oversampling technique; XGboost: Extreme gradient boosting.