Prediction Models for AKI in ICU: A Comparative Study

Figures & data

Figure 1 Time windows for data collection.

Figure 2 Handling and splitting of imbalanced data.

Notes: (A) Imbalanced cohort with more non-AKI patients than AKI patients. (B) Balanced cohort with the same number of AKI and non-AKI patients. (C) Resorting of the sequence randomly. (D) Randomly selected 20% from c as the testing set. (E) Random selection of 20% from c as the training set.

Table 1 Demographic Characteristics of the Patient Cohort at Baseline

	Age (Years)		BMI (kg/m^2)		Male/Female
	AKI	Non-AKI	AKI	Non-AKI	AKI	Non-AKI
Mean	68.19	64.58	29.15	28.51	–	–
Standard deviation	14.90	16.92	7.71	8.09	–	–
P (95% CI)	< 2.2×10^−16		1.977×10^−8		–
Ratio					1.46	1.19

Table 2 Statistic Summary of the 10-Times AUC Result for Each Model

Group	Sum	Average	Variance
LightGBM	9.04933	0.904933	0.000678
Logistic regression	7.355297	0.73553	9.51×10^−5
Random forest	7.284963	0.728496	0.000132
Support vector machine	7.371453	0.737145	8.81×10^−5
eXtreme Gradient Boosting	7.6001	0.76001	5.96×10^−5
Convolutional neural network	7.183812	0.718381	0.000311

Table 3 ANOVA of the 10-Times AUC Result for Each Model

Source of Variation	SS	df	MS	F	P-value	F Crit
Between groups	0.247513	5	0.049503	217.7206	1.71E-34	2.38607
Within groups	0.012278	54	0.000227
Total	0.259791	59

Abbreviations: SS, sum of squares of deviation from mean; DF, degree of freedom; MS, mean square; F, F-measure.

Table 4 Predictive Models Results

	Accuracy	Precision	Recall	F1	AUC
LightGBM	0.905	0.971	0.836	0.897	0.905
eXtreme Gradient Boosting	0.76	0.745	0.8	0.771	0.76
Support vector machine	0.737	0.763	0.695	0.727	0.737
Logistic regression	0.735	0.756	0.694	0.724	0.736
Random forest	0.729	0.727	0.743	0.735	0.728
Convolutional neural network	18	0.715	0.730	0.722	18

Table 5 Studies on Predicting Acute Kidney Injury in ICU

Studies	Models Tested	Best Model
Mengxin Sun et al	LR, RF, NB, SVM, CNN	SVM
Lindsay P. Zimmerman et al^Citation21	LR, RF, MP	MP
Yikuan L et al^Citation25	LR, RF, NB, SVM, GBDT, CNN	LR and SVM
Yuan Zhan et al^Citation26	LR, RF, LightGBM	LightGBM

Abbreviations: LR, logistic regression; RF, random forest; NB, Naïve Bayes; SVM, support vector machine; CNN, convolutional neural networks; MP, multilayer perceptron; GBDT, gradient boosting decision tree; LightGBM, Light Gradient Boosting Decision.

Figure 3 The importance of each feature in each model.

Abbreviations: LGB, Light Gradient Boosting Decision; RF, random forest; XGB, eXtreme Gradient Boosting; LR, logistic regression; SVM, supported vector machine; AVG, average.

Zimmerman LP, Reyfman PA, Smith ADR, et al. Early prediction of acute kidney injury following ICU admission using a multivariate panel of physiological measurements. BMC Med Inform Decis Mak. 2019;19(Suppl 1):16. doi:10.1186/s12911-019-0733-z

 PubMed Web of Science ®Google Scholar

Li Y. Early prediction of acute kidney injury in critical care setting using clinical notes. 2018 IEEE International Conference on Bioinformatics and Biomedicine (BIBM); December; 18; Los Alamitos, CA, USA.

 Google Scholar

Yuan Z, Feng C. LightGBM model for predicting acute kidney injury risk in ICU patients. Acad J Chinese PLA Med Sch. 2019;40(4):316–320.

 Google Scholar