Dependencies

python == 3.8.10 pytorch == 1.11.0 numpy == 1.22.4 pandas == 1.4.3 scikit-learn == 1.1.2

Datasets

We made our experiments on the MIMIC-III and eICU datasets. In order to access the datasets, please refer to https://mimic.physionet.org/gettingstarted/access/ and https://eicu-crd.mit.edu/gettingstarted/access/.

MIMIC-III			eICU
Feature	Type	Missingness (%)	Feature	Type	Missingness (%)
Capillary refill rate	categorical	99.78	-	-	-
Diastolic blood pressure	continuous	30.90	Diastolic blood pressure	continuous	33.80
Fraction inspired oxygen	continuous	94.33	Fraction inspired oxygen	continuous	98.14
Glasgow coma scale eye	categorical	82.84	Glasgow coma scale eye	categorical	83.42
Glasgow coma scale motor	categorical	81.74	Glasgow coma scale motor	categorical	83.43
Glasgow coma scale total	categorical	89.16	Glasgow coma scale total	categorical	81.70
Glasgow coma scale verbal	categorical	81.72	Glasgow coma scale verbal	categorical	83.54
Glucose	continuous	83.04	Glucose	continuous	83.89
Heart Rate	continuous	27.43	Heart Rate	continuous	27.45
Height	continuous	99.77	Height	continuous	99.19
Mean blood pressure	continuous	31.38	Mean arterial pressure	continuous	96.53
Oxygen saturation	continuous	26.86	Oxygen saturation	continuous	38.12
Respiratory rate	continuous	26.80	Respiratory rate	continuous	33.11
Systolic blood pressure	continuous	30.87	Systolic blood pressure	continuous	33.80
Temperature	continuous	78.06	Temperature	continuous	76.35
Weight	continuous	97.89	Weight	continuous	98.65
pH	continuous	91.56	pH	continuous	97.91
Age	continuous	0.00	Age	continuous	0.00
Admission diagnosis	categorical	0.00	Admission diagnosis	categorical	0.00
Ethnicity	categorical	0.00	Ethnicity	categorical	0.00
Gender	categorical	0.00	Gender	categorical	0.00

Main Entrance

main.py contains both training code and evaluation code.

Ablation Studies

We present two variants of our approach as follows:
$Ours_{\alpha}$ (A variation of our approach that does not perform graph analysis-based patient stratification modeling) :
Remove the Similarity, GCN, and InfoAgg classes
$Ours_{\beta}$ (A variation of our approach in which we omit the contrastive learning component) :
Remove the CLLoss class