Optimized Machine Learning Models to Predict In-Hospital Mortality for Patients with ST-Segment Elevation Myocardial Infarction

Figures & data

Table 1 Simplified Set and Full Set of Features

Simplified Set of Features	Full Set of Features
1. Sex	1. Sex
2. Age	2. Age
3. Consciousness	3. Consciousness
4. Respiration	4. Respiration
5. Pulse	5. Pulse
6. Heart rate	6. Heart rate
7. Systolic blood pressure	7. Systolic blood pressure
8. Diastolic blood pressure	8. Diastolic blood pressure
9. Killip class	9. Killip class
10. Cardiac troponin I	10. Cardiac troponin I
11. Time from symptom to first medical contact	11. Time from symptom to first medical contact
12. Prehospital mode of transport	12. Prehospital mode of transport
13. Heart failure	13. Heart failure
	14. Thrombolysis
	15. Reperfusion treatment
	16. TIMI flow after intervention

Abbreviation: TIMI, thrombolysis in myocardial infarction.

Table 2 Basic Characteristics of Continuous Numerical Variables

Features	Total	Patients Survived	Patients Died	p-value
No. of patients	5708	5554	154
Age, years	63.00 (16)	64.50 (16)	75.00 (16)	<0.001
Respiration, counts/min	19.00 (3)	19.00 (3)	19.00 (2)	0.006
Pulse, beats/min	76.00 (22)	76.00 (22)	82.00 (32)	<0.001
HR, beats/min	76.00 (23)	76.00 (22)	82.00 (32)	<0.001
SBP, mm Hg	133.00 (32)	133.00 (31)	120.50 (43)	<0.001
DBP, mm Hg	80.00 (21)	80.00 (21)	74.50 (25)	<0.001
cTnI, ng/mL	1.750 (3.420)	1.735 (3.410)	2.230 (3.863)	0.041
S to FMC, min	102.00 (179)	101.00 (178)	110.00 (291)	0.065

Abbreviations: HR, heart rate; SBP, systolic blood pressures; DBP, diastolic blood pressure; PCI, percutaneous coronary intervention; cTnI, cardiac troponin I; S to FMC, time from symptom to first medical contact.

Table 3 Basic Characteristics of Categorical Variables

Features	Total	Patients Survived	Patients Died	p-value
No. of patients	5708	5554	154
Male	4198 (73.55%)	4108 (73.96%)	90 (58.44%)	<0.001
Consciousness	5690 (99.68%)	5542 (99.78%)	148 (96.10%)	<0.001
Heart failure	401 (7.03%)	318 (5.73%)	83 (53.90%)	<0.001
Thrombolysis	314 (5.50%)	307 (5.53%)	7 (4.55%)	0.586
Prehospital mode of transport				<0.001
EMS	916 (16.05%)	876 (15.77%)	40 (25.97%)
Transferred from other hospitals	1143 (20.02%)	1102 (19.84%)	41 (26.62%)
Self-transported	3649 (63.93%)	3576 (64.39%)	73 (47.40%)
Killip classification				<0.001
I	5396 (94.53%)	5277 (95.01%)	119 (77.27%)
II	235 (4.12%)	221 (3.98%)	14 (9.09%)
III	44 (0.77%)	39 (0.70%)	5 (3.25%)
IV	33 (0.58%)	17 (0.31%)	16 (10.39%)
Reperfusion type				<0.001
Non	748 (13.10%)	669 (12.05%)	79 (51.30%)
Primary PCI	4381 (76.75%)	4317 (77.73%)	64 (41.56%)
Emergent coronary angiography	265 (4.64%)	261 (4.70%)	4 (2.60%)
Thrombolysis	245 (4.29%)	238 (4.29%)	7 (4.55%)
Thrombolysis + Primary PCI	69 (1.21%)	69 (1.24%)	0 (0.00%)
TIMI flow after intervention				0.082
0	185 (3.24%)	180 (3.88%)	5 (7.35%)
1	6 (0.11%)	6 (0.13%)	0 (0.00%)
2	214 (3.75%)	209 (4.50%)	5 (7.35%)
3	4308 (75.47%)	4250 (91.50%)	58 (85.29%)

Abbreviations: EMS, Emergency medical services; PCI, percutaneous coronary intervention; TIMI, thrombolysis in myocardial infarction.

Table 4 Performance Comparison of the Traditional and Optimized ML Models Using the Full Set

Models Trained with RUS Using the Full Set
Models	Accuracy (%)	Sensitivity (%)	Specificity (%)	G-mean (%)	AUC
LR	73.77	89.47	73.34	81.01	0.908
SVM	85.62	84.21	85.66	84.93	0.919
DT	78.33	81.58	78.24	79.89	0.819
RF	71.67	81.58	71.40	76.32	0.809
Models Trained without RUS Using the Full Set
LR	96.91	7.89	99.35	28.01	0.877
SVM	96.42	23.68	98.41	26.09	0.795
DT	97.27	5.26	99.78	22.92	0.842
RF	97.34	2.63	99.93	16.22	0.848

Abbreviations: LR, logistic regression; SVM, support vector machine; DT, decision tree; RF, random forest; AUC, area under the curve.

Figure 1 ROC analysis results of models trained without RUS using the full set.

Abbreviation: ROC curve, receiver operating characteristic curve.

Figure 2 ROC analysis results of models trained with RUS using the full set.

Abbreviation: ROC curve, receiver operating characteristic curve.

Figure 3 The confusion matrix of the SVM trained with RUS using the full set.

Abbreviations: SVM, support vector machine; RUS, random under-sampling.

Figure 4 The confusion matrix of the SVM trained without RUS using the full set.

Abbreviations: SVM, support vector machine; RUS, random under-sampling.

Table 5 Performance Comparison of the Traditional and Optimized ML Models Using the Simplified Set

Model Trained with RUS Using the Simplified Set
Models	Accuracy (%)	Sensitivity (%)	Specificity (%)	G-mean (%)	AUC
LR	85.13	68.42	85.59	76.53	0.841
SVM	78.19	89.47	77.88	83.4	0.746
DT	92.92	55.26	93.95	72.05%	0.746
RF	70.62	71.05	70.61	70.83	0.710
Model Trained without RUS using the simplified set
LR	97.27	5.26	99.78	22.92	0.877
SVM	96.35	23.68	98.34	48.26	0.656
DT	96.77	10.53	99.14	32.30	0.785
RF	96.77	15.79	98.99	39.54	0.694

Abbreviations: LR, logistic regression; SVM, support vector machine; DT, decision tree; RF, random forest; AUC, area under the curve.

Figure 5 The confusion matrix of the SVM trained with RUS using the simplified set.

Abbreviations: SVM, support vector machine; RUS, random under-sampling.

Figure 6 The feature importance scores for predicting the in-hospital mortality provided by the RF.

Abbreviation: RF, random forest.

Table 6 Best Predictive Performance Results of the Models Developed by Other Articles

Article	Accuracy	AUC
Machine learning techniques to predict in-hospital cardiovascular outcomes in elderly patients presenting with acute myocardial infarction^Citation24	-	0.74
Prediction of in-hospital mortality and length of stay in acute coronary syndrome patients using machine-learning methods^Citation25	90.0%	-
Risk prediction for in-hospital mortality in women with ST-elevation myocardial infarction: a machine learning approach^Citation26	-	0.85
An explainable machine learning-based risk prediction model for in-hospital mortality for Chinese STEMI patients: Findings from China Myocardial Infarction registry^Citation27	-	0.899
Machine learning models to predict in-hospital mortality for ST-elevation myocardial infarction: from China Acute Myocardial Infarction (CAMI) registry^Citation28	-	0.868

Bansal A, Kapadia S, Ellis S, et al. Machine learning techniques to predict in-hospital cardiovascular outcomes in elderly patients presenting with acute myocardial infarction. J Am Coll Cardiol. 2020;75(11 suppl 1):3603. doi:10.1016/S0735-1097(20)34230-3

 Web of Science ®Google Scholar

Yakovlev A, Metsker O, Kovalchuk S, Bologova E. Prediction of in-hospital mortality and length of stay in acute coronary syndrome patients using machine-learning methods. J Am Coll Cardiol. 2018;71(11 suppl S):242. doi:10.1016/S0735-1097(18)30783-6.

 Web of Science ®Google Scholar

Mansoor H, Bian J. Risk prediction for in-hospital mortality in women with ST-elevation myocardial infarction: a machine learning approach. J Am Coll Cardiol. 2017;69(11 suppl S):171. doi:10.1016/S0735-1097(17)33560-X

 Web of Science ®Google Scholar

Yang JG, Li YX, Li X, Chen TG, Xie GT, Yang YJ. An explainable machine learning-based risk prediction model for in-hospital mortality for Chinese STEMI patients: findings from china myocardial infarction registry. J Am Coll Cardiol. 2019;73(9 suppl 1):261. doi:10.1016/S0735-1097(19)30869-1.

 PubMed Web of Science ®Google Scholar

Yang JG, Li X, Chen T, Li YX, Xie GT, Yang YJ. Machine learning models to predict in-hospital mortality for ST-elevation myocardial infarction: from china acute myocardial infarction (cami) registry. J Am Coll Cardiol. 2018;71(11 suppl S):236. doi:10.1016/S0735-1097(18)30777-0.

 Web of Science ®Google Scholar