Improving Mortality Risk Prediction with Routine Clinical Data: A Practical Machine Learning Model Based on eICU Patients

Figures & data

Figure 1 The process of cohort selection.

Table 1 Basic Characteristics for the Survival and Death Groups

Variable Name	All Patients	Alive	Death	P value
	n=123,929	n=114,085	n=9844
Age (years), mean (SD)	63.5+17.7	62.9+17.7	70.2+15.9	<0.01
Male	66,704(53.8%)	61,423(53.8%)	5281(54.0%)	0.72
APACHE IV scores, mean (SD)	55.2+24.9	52.4+22.1	87.6+31.4	<0.01
Ethnicity (%)				<0.01
Asian	1783(1.4%)	1626(1.4%)	157(1.6%)
Hispanic	4853(3.9%)	4475(3.9%)	378(3.8%)
Caucasian	95,102(76.7%)	87,399(76.6%)	7703(78.3%)
American	15,264(12.3%)	14,152(12.4%)	1113(11.2%)
Other	5137(5.7%)	6434(5.7%)	493(5.0%)
Length of stay in ICU (days), median(IQR)	1.9 (1.1–3.4)	1.8 (1.0–3.2)	3.1 (1.5–6.2)	<0.01
Length of stay in hospital (days), median (IQR)	5.3 (3.1–9.1)	5.3 (1.0–9.0)	5.2 (2.6–10.1)	<0.01

Abbreviations: SD, standard deviation; IQR, interquartile range (25–75%).

Table 2 Model Performance in the Independent Testing Dataset for Different Machine Learning Algorithms

Algorithms	AUROC	AUPRC	Brier Score
APACHE IV scores	0.8598	0.4341	0.0624
Logistic Regression	0.9357	0.7775	0.0306
Random Forest	0.9559	0.8172	0.0285
Artificial Neural Network	0.9620	0.8256	0.0501
Extreme Gradient Boosting	0.9702	0.8517	0.0259

Abbreviations: AUROC, area under the receiver-operator curve; AUPRC, area under the Precision and Recall curve.

Figure 2 Calibration curves of the machine learning models in the test dataset. X-axis indicates the predicted mean mortality risk, Y-axis indicates the actual mean mortality risk, the slope indicates the relation between the predicted and observed outcomes.

Table 3 The XGBoost Model Performance in Different Short-Term Mortality Risk Prediction Tasks

Prediction Tasks	AUROC	AUPRC	Brier Score
1-day mortality prediction (n=12,393)	0.9702	0.8517	0.0259
2-day mortality prediction (n=10,907)	0.9310	0.6437	0.0427
3-day mortality prediction (n=9403)	0.9148	0.5519	0.0477
4-day mortality prediction (n=7935)	0.8971	0.4835	0.0517
5-day mortality prediction (n=6697)	0.8627	0.3720	0.0592
6-day mortality prediction (n=5148)	0.8544	0.3731	0.0647
7-day mortality prediction (n=4667)	0.8406	0.3533	0.0668

Abbreviations: AUROC, area under the receiver-operator curve; AUPRC, area under the Precision and Recall curve.

Figure 3 Precision-Recall curve of the XGBoost model in different tasks. Lines with different color indicate mortality risk prediction task within different time periods. Task with bigger area under Precision-Recall curves shows better discrimination performance.