J Clin Med Res

Journal of Clinical Medicine Research, ISSN 1918-3003 print, 1918-3011 online, Open Access

Article copyright, the authors; Journal compilation copyright, J Clin Med Res and Elmer Press Inc

Journal website https://jocmr.elmerjournals.com

Review

Volume 18, Number 8, August 2026, pages 515-527

Application Progress of Machine Learning in Prognostic Prediction of Percutaneous Coronary Intervention: A Systematic Review

↓ **Table 1.** Performance Comparison of Different Models for Predicting PCI in-Hospital Mortality
Prediction model	Sample size	Core evaluation index	Sensitivity	Specificity	Key prognostic factors
PCI: percutaneous coronary intervention; AUC: area under the receiver operating characteristic curve; C-statistic: concordance statistic; GRACE: Global Registry of Acute Coronary Events; SVM: support vector machine; STEMI: ST-segment elevation myocardial infarction.
GRACE score (traditional)	21,237	C-statistic = 0.83	—	—	Conventional clinical indicators
Random forest	21,237	C-statistic = 0.92	—	—	Age, renal function, hemodynamic status
SVM	1,895 STEMI patients	AUC = 0.94	91.2%	88.7%	Creatinine level, left ventricular ejection fraction, cardiogenic shock
Deep neural network	Large-sample dataset	AUC = 0.94	—	—	Albumin level, lymphocyte count, conventional indicators

↓ Table 1. Performance Comparison of Different Models for Predicting PCI in-Hospital Mortality

Prediction model

Sample size

Core evaluation index

Sensitivity

Specificity

Key prognostic factors

PCI: percutaneous coronary intervention; AUC: area under the receiver operating characteristic curve; C-statistic: concordance statistic; GRACE: Global Registry of Acute Coronary Events; SVM: support vector machine; STEMI: ST-segment elevation myocardial infarction.

GRACE score (traditional)

21,237

C-statistic = 0.83

—

Conventional clinical indicators

Random forest

21,237

C-statistic = 0.92

—

Age, renal function, hemodynamic status

SVM

1,895 STEMI patients

AUC = 0.94

91.2%

88.7%

Creatinine level, left ventricular ejection fraction, cardiogenic shock

Deep neural network

Large-sample dataset

AUC = 0.94

—

Albumin level, lymphocyte count, conventional indicators

↓ **Table 2.** Performance Comparison of Different Models for MACE Prediction After PCI
Prediction model	Prediction cycle	Core evaluation indicators	Core advantages
PCI: percutaneous coronary intervention; MACEs: major adverse cardiovascular events; C-index: concordance index; XGBoost: extreme gradient boosting; SHAP: SHapley Additive exPlanations; AUC: area under the receiver operating characteristic curve; F1: F1-score; GRACE: Global Registry of Acute Coronary Events.
GRACE score (traditional)	1-year	C-index = 0.71	Simple operation, limited ability to capture nonlinear relationships
Random forest	1-year	C-index = 0.78	Stable performance, accurate screening of key risk factors
XGBoost	30-day	Precision = 0.89, recall = 0.85, F1 = 0.87	Excellent for imbalanced datasets, interpretable via SHAP
Multimodal deep learning	5-year	AUC = 0.83 (traditional clinical model: AUC = 0.76)	Integrate imaging and clinical data, high long-term prediction accuracy

↓ Table 2. Performance Comparison of Different Models for MACE Prediction After PCI

Prediction model

Prediction cycle

Core evaluation indicators

Core advantages

PCI: percutaneous coronary intervention; MACEs: major adverse cardiovascular events; C-index: concordance index; XGBoost: extreme gradient boosting; SHAP: SHapley Additive exPlanations; AUC: area under the receiver operating characteristic curve; F1: F1-score; GRACE: Global Registry of Acute Coronary Events.

GRACE score (traditional)

1-year

C-index = 0.71

Simple operation, limited ability to capture nonlinear relationships

Random forest

1-year

C-index = 0.78

Stable performance, accurate screening of key risk factors

XGBoost

30-day

Precision = 0.89, recall = 0.85, F1 = 0.87

Excellent for imbalanced datasets, interpretable via SHAP

Multimodal deep learning

5-year

AUC = 0.83 (traditional clinical model: AUC = 0.76)

Integrate imaging and clinical data, high long-term prediction accuracy

↓ **Table 3.** Performance Comparison of Different Models for PCI-Related Bleeding Complication Prediction
Prediction model	AUC	Sensitivity	Specificity	Model characteristics
PCI: percutaneous coronary intervention; AUC: area under the receiver operating characteristic curve; CRUSADE: Can Rapid risk stratification of Unstable angina patients Suppress ADverse outcomes with Early implementation of the ACC/AHA guidelines; LSTM: long short-term memory.
CRUSADE score (traditional)	0.64	—	—	Low accuracy, unable to process high-dimensional variables
Single random forest	0.76	—	—	Adaptable to high-dimensional clinical variables
Ensemble voting classifier	0.81	78%	76%	Integrate multiple algorithms with better robustness
LSTM time-series model	—	85% (accuracy)	—	Dynamic real-time risk assessment within 72 h after PCI

↓ Table 3. Performance Comparison of Different Models for PCI-Related Bleeding Complication Prediction

Prediction model

AUC

Sensitivity

Specificity

Model characteristics

PCI: percutaneous coronary intervention; AUC: area under the receiver operating characteristic curve; CRUSADE: Can Rapid risk stratification of Unstable angina patients Suppress ADverse outcomes with Early implementation of the ACC/AHA guidelines; LSTM: long short-term memory.

CRUSADE score (traditional)

0.64

—

Low accuracy, unable to process high-dimensional variables

Single random forest

0.76

—

Adaptable to high-dimensional clinical variables

Ensemble voting classifier

0.81

78%

76%

Integrate multiple algorithms with better robustness

LSTM time-series model

—

85% (accuracy)

—

Dynamic real-time risk assessment within 72 h after PCI

↓ **Table 4.** Comparison of Long-Term Survival Prediction Models After PCI
Prediction model	Prediction term	C-index	Core advantages	Limitations
PCI: percutaneous coronary intervention; C-index: concordance index; RSF: random survival forest.
Cox proportional hazards regression	5-year	0.71	Mature and classic, easy to implement	Poor capability to capture nonlinear variable interactions
RSF	5-year	0.76	Effectively capture nonlinear relationships between variables	Limited ability to mine ultra-high-dimensional data
DeepSurv deep learning model	10-year	0.82	Adapt to high-dimensional variables such as genetic loci with high long-term prediction accuracy	Poor model interpretability

↓ Table 4. Comparison of Long-Term Survival Prediction Models After PCI

Prediction model

Prediction term

C-index

Core advantages

Limitations

PCI: percutaneous coronary intervention; C-index: concordance index; RSF: random survival forest.

Cox proportional hazards regression

5-year

0.71

Mature and classic, easy to implement

Poor capability to capture nonlinear variable interactions

RSF

5-year

0.76

Effectively capture nonlinear relationships between variables

Limited ability to mine ultra-high-dimensional data

DeepSurv deep learning model

10-year

0.82

Adapt to high-dimensional variables such as genetic loci with high long-term prediction accuracy

Poor model interpretability

↓ **Table 5.** Characteristic Comparison of Mainstream Machine Learning Algorithms for PCI Prognostic Prediction
Algorithm type	Applicable data characteristics	Core advantages	Main application scenarios
PCI: percutaneous coronary intervention; SVM: support vector machine; XGBoost: extreme gradient boosting; MLP: multilayer perceptron; CNN: convolutional neural network; LSTM: long short-term memory; MACEs: major adverse cardiovascular events; ECG: electrocardiogram; STEMI: ST-segment elevation myocardial infarction; SHAP: SHapley Additive exPlanations.
Random forest	Large-sample and high-dimensional data with missing values	Automatic feature selection, strong robustness, partial interpretability	In-hospital mortality, bleeding risk, medium-term MACE prediction
SVM	Small-sample nonlinear datasets	Superior small-sample performance, good generalization ability	In-hospital mortality prediction of STEMI patients
XGBoost	Large-scale and imbalanced clinical data	Anti-overfitting, high training efficiency, SHAP-based interpretability	Short-term MACEs prediction, imbalanced sample prediction tasks
MLP	High-dimensional structured clinical data	Automatically learn nonlinear feature interactions	General prognostic prediction tasks
CNN	Grid-based medical image data	Automatic medical imaging feature extraction	Feature analysis of coronary angiography, ECG and echocardiography
LSTM	Time-series dynamic monitoring data	Capture long-term dynamic data dependencies	Dynamic postoperative bleeding risk assessment
DeepSurv	High-dimensional omics and survival data	Mine novel genetic prognostic biomarkers	Long-term survival prediction of PCI patients

↓ Table 5. Characteristic Comparison of Mainstream Machine Learning Algorithms for PCI Prognostic Prediction

Algorithm type

Applicable data characteristics

Core advantages

Main application scenarios

PCI: percutaneous coronary intervention; SVM: support vector machine; XGBoost: extreme gradient boosting; MLP: multilayer perceptron; CNN: convolutional neural network; LSTM: long short-term memory; MACEs: major adverse cardiovascular events; ECG: electrocardiogram; STEMI: ST-segment elevation myocardial infarction; SHAP: SHapley Additive exPlanations.

Random forest

Large-sample and high-dimensional data with missing values

Automatic feature selection, strong robustness, partial interpretability

In-hospital mortality, bleeding risk, medium-term MACE prediction

SVM

Small-sample nonlinear datasets

Superior small-sample performance, good generalization ability

In-hospital mortality prediction of STEMI patients

XGBoost

Large-scale and imbalanced clinical data

Anti-overfitting, high training efficiency, SHAP-based interpretability

Short-term MACEs prediction, imbalanced sample prediction tasks

MLP

High-dimensional structured clinical data

Automatically learn nonlinear feature interactions

General prognostic prediction tasks

CNN

Grid-based medical image data

Automatic medical imaging feature extraction

Feature analysis of coronary angiography, ECG and echocardiography

LSTM

Time-series dynamic monitoring data

Capture long-term dynamic data dependencies

Dynamic postoperative bleeding risk assessment

DeepSurv

High-dimensional omics and survival data

Mine novel genetic prognostic biomarkers

Long-term survival prediction of PCI patients

↓ **Table 6.** Advantages and Disadvantages of Machine Learning for PCI Prognostic Prediction
Advantages	Disadvantages
PCI: percutaneous coronary intervention.
Can identify nonlinear and high-dimensional relationships between variables	Limited interpretability of complex deep learning models (“black box”)
Able to fuse multimodal data (clinical indicators, imaging, time-series signals)	Performance is susceptible to data quality, missing values and inconsistent data standards
Outperforms traditional linear risk scores on most prognostic endpoints	Model performance frequently declines during external validation (domain shift)
Flexible algorithms suitable for classification, survival analysis and imbalanced data tasks	Requires computational resources and professional analytical expertise
Outputs individualized risk probabilities to support precision treatment	Lack of unified standards for modeling, verification and clinical deployment

↓ Table 6. Advantages and Disadvantages of Machine Learning for PCI Prognostic Prediction

Advantages

Disadvantages

PCI: percutaneous coronary intervention.

Can identify nonlinear and high-dimensional relationships between variables

Limited interpretability of complex deep learning models (“black box”)

Able to fuse multimodal data (clinical indicators, imaging, time-series signals)

Performance is susceptible to data quality, missing values and inconsistent data standards

Outperforms traditional linear risk scores on most prognostic endpoints

Model performance frequently declines during external validation (domain shift)

Flexible algorithms suitable for classification, survival analysis and imbalanced data tasks

Requires computational resources and professional analytical expertise

Outputs individualized risk probabilities to support precision treatment

Lack of unified standards for modeling, verification and clinical deployment