Comparison of machine learning algorithms to SAPS II in predicting in-hospital mortality of fractures of the pelvis and acetabulum: analyzes based on MIMIC-III database

Figures & data

Table 1. Overview of datasets.

Label	Name of data set	File types (file extension)	Data repository and identifier (DOI or accession number)
Data file 1	SAPS II score	CSV file (.csv)	Science Data Bank (http://www.dx.doi.org/10.11922/sciencedb.00787) (Cai 2021)
Data file 2	Cohort control	CSV file (.csv)	Science Data Bank (http://www.dx.doi.org/10.11922/sciencedb.00787) (Cai 2021)
Data file 3	Cohort experiment	CSV file (.csv)	Science Data Bank (http://www.dx.doi.org/10.11922/sciencedb.00787) (Cai 2021)

Note: The dataset cohort control contains data without feature expansion, while the cohort experiment file collects data after feature expansion.

Figure 1. Flowchart of patient selection and study design.

Figure 2. Distributions of the characteristics in the survivors and non-survivors. (a) Gender distribution; (b) Admission type distribution.

Figure 3. Distributions of selected variables in the survivors and non-survivors. (a) Age distribution. (b) SAPS II score distribution. SD = Standard deviation.

Table 2. First-24 hours characteristics of enrolled patients with pelvis and/or acetabulum fractures in MIMIC-III.

Characteristics	Survivors (n = 223)	Non-survivors (n = 84)
Age (years)	49.83 (19.96)	66.29 (19.69)
Sex, n (%)
Male	155 (69.5%)	55 (65.5%)
Female	68 (30.5%)	29 (34.5%)
Admission type, n (%)
Medical	198 (88.8%)	80 (95.2%)
Unscheduled surgical	24 (10.8%)	4 (0.48%)
Scheduled surgical	1 (0.04%)	0 (0)
Heart rate (max)	114.73 (19.54)	119.34 (24.29)
Heart rate (min)	74.96 (14.68)	68.48 (15.05)
Systolic BP (max, mmHg)	158.86 (21.04)	161.43 (26.05)
Systolic BP (min, mmHg)	93.61 (15.07)	85.07 (21.37)
Temperature (max, °C)	38.01 (0.65)	37.75 (1.16)
Temperature (min, °C)	36.17 (0.68)	35.64 (1.05)
GCS (min)	13.57 (2.33)	12.88 (3.02)
PaO2/FiO2 (min, mm Hg/%)	248.55 (133.16)	195.33 (126.77)
BUN (max, mg/dL) or serum urea (max, mmol/L)	20.29 (11.66)	29.94 (19.97)
BUN (min, mg/dL) or serum urea (min, mmol/L)	12.55 (8.26)	20.25 (17.58)
Urine output (mL/day)	4728.19 (2910.69)	4315.20 (3334.23)
Sodium (max, mEq/L or mmol/L)	141.91 (3.60)	143.01 (5.62)
Sodium (min, mEq/L or mmol/L)	136.07 (3.75)	135.95 (4.29)
Potassium (max, mEq/L)	4.70 (0.83)	4.75 (0.70)
Potassium (min, mEq/L)	3.53 (0.40)	3.47 (0.47)
Bicarbonate (max, mEq/L)	27.93 (3.09)	26.40 (4.38)
Bicarbonate (min, mEq/L)	22.37 (3.89)	21.39 (4.46)
Bilirubin (max, mg/dL)	1.20 (2.08)	1.31 (1.43)
Bilirubin (min, mg/dL)	0.90 (1.47)	0.84 (0.68)
WBC (max, × 10³/mm³)	16.54 (7.73)	15.06 (6.93)
WBC (min, × 10³/mm³)	8.37 (3.28)	8.33 (3.75)
Chronic disease, n (%)	1 (0.04%)	0 (0)

Notes: For the numerical attributes, the table provides mean (standard deviation); the categorical variables are listed with number (%). Scheduled surgical = surgery scheduled ≥24 hours prior; medical = no surgery within 1 week of admission; unscheduled surgical = surgery scheduled ≤24 hours prior. BP = blood pressure; GCS = Glasgow coma scale; BUN = blood urea nitrogen; WBC = white blood cells.

Table 3. Comparison of customized mortality prediction models with SAPS II.

Models		Accuracy	Precision	Recall	AUROC	H-L p Value
SAPS II		0.75	0.74	0.16	0.73	<.001
SVM	Control	0.76	0.75	0.78	0.86	NA
	Experiment	0.80	0.79	0.82	0.89	NA
LR	Control	0.77	0.77	0.77	0.76	.37
	Experiment	0.80	0.80	0.81	0.78	.27
DT	Control	0.86	0.86	0.88	0.87	NA
	Experiment	0.91	0.88	0.94	0.89	NA
RF	Control	0.87	0.87	0.88	0.87	NA
	Experiment	0.87	0.86	0.89	0.90	NA

SVM = support vector machine; LR = logistic regression; DT = decision tree; RF = random forest; NA = not applicable.

Figure 4. Receiver-operating characteristics (ROC) curves of customized machine learning models. LR = logistic regression; DT = decision tree; RF = random forest.

Figure 5. Feature importance of customized machine learning models.

Cai Y. 2021. Mortality Prediction MIMIC-III. V1. Science Data Bank. 2021. doi:10.11922/sciencedb.00787.

 Google Scholar

Data availability statement

All the data utilized in our research came from the MIMIC-III database, developed by The Laboratory of Computational Physiology at the Massachusetts Institute of Technology (MIT), which can be accessed from the official website: https://mimic.physionet.org/ (Johnson et al. 2016). The final datasets for this project are freely and openly accessed on the Science Data Bank under (http://www.dx.doi.org/10.11922/sciencedb.00787) (Cai 2021). Please see Table 1 and reference list for details and link to the data. The original codes are open to be required for other researchers.