Development and validation of a machine learning model for prediction of 1-year mortality following ST-elevation myocardial infarction: a retrospective cohort study

Abstract

Objectives

To develop a machine learning (ML)-based risk prediction model for 1-year mortality in ST-elevation myocardial infarction (STEMI) patients undergoing primary or rescue percutaneous coronary intervention.

Design

Patient data, including demographic, clinical, biochemical, imaging and procedural details, were extracted from electronic medical records. Data were split into training (80%) and test (20%) sets. Eight supervised learning algorithms were evaluated: least absolute shrinkage and selection operator, ridge, Elastic Net (EN, decision tree, support vector machine, random forest, AdaBoost and gradient boosting. Feature selection was performed sequentially with subsets of the top 5/10/15/20/25/30 features. Model hyperparameters were optimised using fivefold cross-validation with area under the curve (AUC) as the scoring metric.

Setting

Single, tertiary Australian centre.

Participants

We analysed data from 1863 consecutive STEMI patients treated at a tertiary Australian centre from July 2010 to December 2019.

Outcome measures

The primary outcome was 1-year all-cause mortality.

Results

The 1-year mortality rate was 13.6% (n=254) in our cohort. The EN model with five key features (parsimonious model) demonstrated superior performance, achieving an AUC of 0.821, which was comparable to the full 30-variable model (AUC 0.821). Advanced age, pre-hospital cardiac arrest and management with balloon angioplasty alone were identified as predictors of increased mortality risk, while family history of premature coronary disease and higher left ventricular ejection fraction were associated with improved survival. To facilitate clinical implementation, we developed a user-friendly web application for individualised risk assessment.

Conclusion

Our ML model accurately predicts 1-year mortality in STEMI patients using only five clinical variables. This tool offers improved accuracy and ease of use compared with existing risk stratification methods, potentially enhancing patient stratification and guiding treatment decisions in STEMI management.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• This study presents a large, real-world dataset with comprehensive evaluation of multiple machine learning (ML) algorithms, and the final model’s high area under the curve indicates robust discriminative ability.

• The development of a web-based application represents a significant step towards translating our findings into clinical practice. By providing an easily accessible platform for individualised risk assessment, we address a key challenge in implementing ML models in healthcare settings.

• The retrospective, single-centre design may limit generalisability, and despite internal validation on a hold-out test dataset, external validation in diverse populations remains ideal.

• Missing variables, although imputed and uncommon, may influence the outcomes.

Introduction

Acute ST-elevation myocardial infarction (STEMI) remains a significant cause of morbidity and mortality worldwide, despite advances in treatment strategies.¹ Timely reperfusion through primary percutaneous coronary intervention (pPCI) has become the gold standard for managing STEMI, significantly improving outcomes.² However, accurately predicting mortality risk in STEMI patients remains challenging, with existing risk stratification tools often lacking in precision or ease of use.3,6

Multiple factors have been identified as predictors of poor outcomes in STEMI patients. Advanced age, female gender, diabetes mellitus, renal dysfunction, anterior infarct location and cardiogenic shock have consistently been associated with increased mortality.^{1 7} Additionally, lower socioeconomic status has emerged as a significant independent predictor of adverse events, with studies demonstrating that socioeconomically disadvantaged patients experience higher rates of complications, readmissions and mortality following STEMI.⁸ Longer symptom-to-balloon times, reduced adherence to guideline-directed medical therapy and limited access to cardiac rehabilitation services further contribute to the poorer prognosis observed in these vulnerable populations.^{2 8}

In recent years, machine learning (ML) techniques have shown promise in enhancing risk prediction across various medical fields.⁹ These advanced analytical methods can potentially identify complex patterns and relationships within large datasets, potentially surpassing traditional statistical approaches in accuracy and adaptability.^{9 10} However, the application of ML in STEMI risk prediction has been limited, and there is a need for models that are both highly accurate and clinically applicable.¹¹ The identification of key predictors of 1-year mortality and providing an accessible platform for risk calculation has the potential to enhance patient stratification, guide treatment decisions and ultimately improve outcomes for STEMI patients.¹²

Our study aimed to develop and validate a novel ML-based risk prediction model for 1-year mortality in STEMI patients undergoing pPCI or rescue percutaneous coronary intervention (PCI). By leveraging a comprehensive dataset from a tertiary Australian centre and employing various supervised learning algorithms, we sought to create a tool that could provide individualised risk assessments with high accuracy. Furthermore, we aimed to translate our findings into a user-friendly web application, facilitating the implementation of this predictive model in clinical practice.

Materials and methods

We conducted a retrospective analysis of all consecutive patients who underwent pPCI or rescue PCI for STEMI at our tertiary Australian centre from July 2010 to December 2019. All patients received standard-of-care treatment, including aspirin administration prior to PCI (unless contraindicated) and intra-arterial therapeutic heparin at the procedure’s onset. Glycoprotein IIb/IIIa inhibitor use was at the discretion of the interventional cardiologist. The study was approved by the Northern Sydney Local Health District Human Research Ethics Committee (reference number: HREC/RESP/18/155). All methods were performed in accordance with the relevant guidelines and regulations. All research was performed in accordance with the Declaration of Helsinki.

We analysed patient electronic medical records to extract key demographic, clinical, biochemical, imaging and procedural details. Invasive haemodynamic parameters, including heart rate and aortic systolic blood pressure, were recorded at the start of coronary angiography. Coronary angiograms were analysed to assess collateral circulation using the Rentrop classification. Left ventricular function was evaluated using post-STEMI transthoracic echocardiography or, if unavailable, ventriculography during the index procedure. Ischaemic time was defined as the duration from the onset of continuous chest pain to the acquisition of the first angiographic image during PCI. The primary outcome was 1-year all-cause mortality. Mortality data were obtained from electronic medical records, which capture deaths occurring within our health network and deaths reported to our facility.

The study included 1863 consecutive STEMI patients treated between July 2010 and December 2019, with 254 mortality events (13.6%) at 1 year. Data were split into training (80%, approximately 1490 patients with 203 events) and test sets (20%, approximately 373 patients with 51 events). This allocation provided an events-per-variable (EPV) ratio of approximately 40:1 for our final 5-predictor model, substantially exceeding the commonly recommended minimum of 10–20 events per predictor for stable model development. The 80/20 split ratio was chosen to optimise the balance between robust model training and sufficient test data for performance evaluation, while maintaining adequate precision in our estimated metrics.

We employed eight supervised learning classification models to create risk prediction algorithms for 1-year mortality: three linear models (least absolute shrinkage and selection operator (LASSO or L1), ridge (L2) and Elastic Net (EN)) and five non-linear models (decision tree, support vector machine, random forest, AdaBoost and gradient boosting).

Missing values were imputed using median values. Feature selection for logistic regression was performed sequentially, with subsets of the top 5/10/15/20/25/30 (all) features, based on coefficient magnitude and then used for other models. Model hyperparameters were optimised using fivefold cross-validation with area under the curve (AUC) as the scoring metric. Final model selection was based on AUC scores on the test set. The chosen models were then calibrated using sigmoid and isotonic regressors, with calibration assessed via Brier scores and calibration plots. All statistical tests used a 5% significance threshold. We performed a retrospective power calculation from our final model.

We performed our analyses with Python (V.3.7), and the methodology code can be found on https://github.com/harisritharan/stemi_risk_prediction/blob/master/stemimodelbuilding.ipynb.

To facilitate real-world implementation, we developed a user-friendly web application using Python Dash. This application incorporates our top-performing model for predicting 1-year mortality following STEMI. Users can input relevant clinical variables to generate a risk assessment score, presented as a percentage probability of 1-year mortality.

Patient and public involvement

There was no patient or public involvement in the design, conduct, reporting or dissemination of this study.

Results

Our study encompassed 1863 STEMI patients, with a 1-year mortality rate of 13.6% (n=254). The cohort’s demographic profile revealed a mean age of 64.9±13.7 years, with a predominance of male patients (77.1%). Cardiovascular risk factors were prevalent, with hypertension (46.4%) and hypercholesterolaemia (39.1%) being the most common. Notably, 29.1% reported a family history of premature coronary disease, while 16.4% had diabetes. Pre-hospital cardiac arrest occurred in 11.7% of cases. The left anterior descending coronary artery was identified as the primary culprit vessel in 47.5% of patients. Comprehensive baseline characteristics are presented in table 1.

Table 1. Baseline characteristics^*.

	All patients	Proportion of missing values
Mean age (SD), year	64.9 (13.7)	0%
Male, no. (%)	1436 (77.1%)	0%
Body mass index (SD), kg/m2	27.3 (4.8)	5.4%
Hypertension, no. (%)	807 (46.4%)	6.7%
Hypercholesterolaemia, no. (%)	677 (39.1%)	7.0%
Diabetes mellitus, no. (%)	278 (16.4%)	9.2%
Family history of coronary disease before age 50, no. (%)	458 (29.1%)	15.5%
Smoking history		19.2%
Never smoker, no. (%)	604 (40.2%)
Ex-smoker, no. (%)	431 (28.7%)
Current smoker, no. (%)	470 (31.3%)
Pre-hospital cardiac arrest, no. (%)	218 (11.7%)	0%
Ischaemic time (SD), min	484.5 (657.0)	11.1%
Starting heart rate (SD), beats/min	79.8 (19.9)	0.6%
Starting systolic blood pressure (SD), mm Hg	122.7 (28.2)	0.9%
Previous stent, no. (%)	186 (10.1%)	7.5%
Culprit coronary artery		0%
Left anterior descending, no. (%)	885 (47.5%)
Left circumflex, no. (%)	269 (14.4%)
Right coronary, no. (%)	709 (38.1%)
Robust collateral recruitment		0%
Yes (Rentrop grade 2 or 3), no. (%)	399 (21.4%)
No (Rentrop grade 0 or 1), no. (%)	1464 (78.6%)
Thrombolysis in myocardial infarction flow pre-PCI		0%
0, no. (%)	1096 (58.8%)
1, no. (%)	156 (8.4%)
2, no. (%)	511 (27.4%)
3, no. (%)	100 (5.4%)
Thrombolysis in myocardial infarction flow post-PCI		0%
0, no. (%)	21 (1.1%)
1, no. (%)	29 (1.6%)
2, no. (%)	138 (7.4%)
3, no. (%)	1675 (89.9%)
Presence of chronic total occlusion in a remote vessel, no. (%)	116 (6.2%)	0%
Percutaneous coronary intervention performed, no. (%)	1714 (92.0%)	0%
Number of stents (SD), no.	1.2 (0.6)	0%
Length of stented segment (SD), mm	30.1 (16.1)	8.0%
Glycoprotein IIb/IIIa inhibitor use, no. (%)	1002 (54.8%)	1.8%
Inotrope use during procedure, no. (%)	269 (14.5%)	0.2%
Intra-aortic balloon pump (IABP) or extracorporeal membrane oxygenation (ECMO) use during procedure, no. (%)	44 (2.7%)	11.8%
Ventricular arrhythmia during procedure, no. (%)	124 (6.7%)	0.2%
Left ventricular ejection fraction less than 40%, no. (%)	373 (20.8%)	3.8%
1-year mortality, no. (%)	254 (13.6%)	0%

^*Note: This table has been adapted from our previous publication.11

We evaluated eight ML algorithms for predicting 1-year mortality. The EN model demonstrated superior performance across various feature subsets, maintaining an AUC of 0.821 for 30, 25, 20, 15 and 5 features. The L1 model slightly outperformed others when incorporating 10 variables (AUC 0.822). Figure 1 illustrates the performance variations of these algorithms relative to feature inclusion.

Figure 1. Predictive performance of models by number of variables for 1-year mortality. Predictive performance of models by number of variables for 1-year mortality. AB, AdaBoost; AUC, area under the curve; DT, decision tree; EN, Elastic Net; GB, gradient boosting; L1, least absolute shrinkage and selection operator (LASSO); L2, ridge regression; RF, random forest; SVM, support vector machine.

Considering both predictive accuracy and clinical utility, we selected the EN model with five features as our definitive algorithm. This model achieved an AUC of 0.821, accuracy of 0.855, precision of 0.375, recall of 0.06 and an F1 score of 0.10. A comparative analysis of this EN model against other models and other five-feature models is provided in online supplemental figure S1 and table S1, respectively. The calibration of our final model was performed using a sigmoid regressor; this demonstrated superior performance compared with an isotonic regressor in calibrating the predicted probabilities to observed outcomes with a lower Brier score and a calibration plot showing closer alignment to the ideal calibration line (online supplemental figure S2).

Our final model identified key predictors of 1-year mortality. Advanced age, pre-hospital cardiac arrest and management solely with balloon angioplasty were associated with increased mortality risk. Conversely, a family history of premature coronary disease and higher left ventricular ejection fraction correlated with improved survival outcomes. The EN model coefficients formed the basis of our final risk stratification score, as depicted in figure 2.

Figure 2. Included coefficients and weighting in the final model for 1-year mortality. LVEF, left ventricular ejection fraction; FHx, Family history; CAD, Coronary artery disease.

To facilitate clinical implementation, we developed an interactive web-based application for individualised mortality risk assessment in STEMI patients. This tool, accessible at https://stemi-ml-score.onrender.com/mortality, is illustrated in figure 3 and provides a user-friendly interface for rapid risk calculation based on our validated model.

Figure 3. Risk score web application. LVEF, left ventricular ejection fraction.

Power and sample size considerations

With 254 mortality events and a final model of 5 predictors, our study achieved an EPV ratio of approximately 51:1, substantially exceeding the commonly recommended minimum of 10–20 events per predictor for stable model development. For our test set validation (51 events, 373 patients), we performed a retrospective power calculation for AUC estimation. With an observed AUC of 0.821, our sample size provided >90% power to detect a true AUC ≥0.75 with a two-sided alpha of 0.05. While fivefold cross-validation was employed for hyperparameter tuning, our final model was evaluated on an independent test set comprising 20% of the data, providing robust validation.

Discussion

Our study demonstrates the successful application of ML techniques in developing a highly accurate predictive model for 1-year mortality following STEMI. The EN model with five key features achieved an impressive AUC of 0.821, outperforming other algorithms and feature combinations for accuracy and practicality. This model’s performance underscores the potential of ML approaches in risk stratification for STEMI patients.

The identified predictors of 1-year mortality align with established clinical knowledge while offering new insights. Advanced age and pre-hospital cardiac arrest are well-recognised risk factors, and our model reinforces their importance.^{7 13} The association of balloon angioplasty alone with increased mortality likely reflects its use in more complex cases or as a salvage strategy when stenting is not feasible.^{14 15} Conversely, the protective effect of a family history of premature coronary disease is intriguing and may indicate earlier medical attention or different pathophysiological mechanisms in these patients. This is supported by results from a large-scale retrospective analysis examining 2 123 492 STEMI admissions, which revealed a notable protective effect associated with a family history of coronary artery disease.¹⁶ Patients with such a history exhibited markedly reduced odds of in-hospital mortality compared with those without (1.4% vs 8.1%; adjusted OR 0.42, 95% CI 0.41 to 0.44; p<0.001), with putative protective effect stemming from increased cardiovascular health awareness and more proactive engagement in both pharmacological and lifestyle-based cardiovascular risk reduction strategies among individuals with a family history of heart disease.¹⁶ Left ventricular ejection fraction (LVEF) emerged as a crucial protective factor, consistent with its known prognostic value in post-STEMI outcomes.^{17 18} The strong predictive power of LVEF in our model highlights the importance of early assessment of left ventricular function in risk stratification.

Our composite endpoint of 1-year mortality includes both in-hospital and post-discharge deaths, which may be driven by different mechanisms. In-hospital mortality is typically related to acute complications such as cardiogenic shock, ventricular arrhythmias, mechanical complications or acute heart failure, while post-discharge mortality may be more influenced by chronic factors including left ventricular remodelling, medication adherence, comorbidities and secondary prevention measures. The protective effect of higher LVEF and the adverse effect of pre-hospital cardiac arrest in our model likely reflect their importance in both phases, whereas factors such as family history of premature CAD (associated with earlier medical attention and better adherence to preventive strategies) may predominantly influence post-discharge outcomes. While our model demonstrates strong overall predictive performance for 1-year mortality, future studies could benefit from developing separate models for in-hospital and post-discharge mortality to better capture the distinct risk profiles and potentially guide phase-specific interventions.

The observation that model performance remained stable or improved when reducing from 30 to 5 variables, while seemingly counterintuitive, is a well-recognised phenomenon in ML known as the ‘curse of dimensionality’ or overfitting. Several factors explain this finding: (1) Many of the 30 variables may be correlated or redundant, providing overlapping information rather than independent predictive value, (2) including weakly predictive or noisy variables can introduce variance that degrades model performance, particularly in relatively modest sample sizes, (3) regularisation techniques like LASSO and EN are specifically designed to identify the most informative features while penalising less important ones and (4) parsimonious models with fewer features often generalise better to new data by avoiding overfitting to training set noise. The consistency of performance across feature subsets demonstrates that the key predictive information is captured by a small number of highly informative variables, while the remaining variables contribute marginal additional information at the cost of increased complexity.

Our study has several strengths, including a large, real-world dataset and the comprehensive evaluation of multiple ML algorithms. The model’s high AUC indicates robust discriminative ability. The development of a web-based application represents a significant step towards translating our findings into clinical practice. By providing an easily accessible platform for individualised risk assessment, we address a key challenge in implementing ML models in healthcare settings.⁶ This tool could aid in clinical decision-making, patient counselling and resource allocation in the critical post-STEMI period.

However, we acknowledge limitations. The retrospective, single-centre design may limit generalisability, and despite internal validation on a hold-out test dataset, external validation in diverse populations remains ideal. Our reliance on electronic medical records for mortality ascertainment, as opposed to linkage with national registries, may have missed some deaths occurring outside our health network, potentially leading to slight underestimation of the true mortality rate. Finally, missing variables, although imputed and uncommon, may influence the outcomes. We handled missing data through imputation using median values for continuous variables and mode for categorical variables. This simple imputation approach was chosen because the proportion of missing data was low (<5% for most variables) and median imputation avoids introducing artificial relationships between variables that can occur with multiple imputation methods when data are missing at random. More sophisticated approaches such as multiple imputation by chained equations might better preserve variable relationships and should be explored in future studies, particularly when higher proportions of missing data are present. Furthermore, important variables that may influence long-term mortality were not systematically available in our dataset, including comprehensive echocardiographic parameters (left ventricular volumes, diastolic function indices), pulmonary haemodynamics and detailed discharge medication data. The absence of these variables may limit the model’s predictive performance, and future prospective studies incorporating comprehensive echocardiographic data, renal function metrics and medication adherence patterns may further improve risk prediction. A recent study by Prasad et al demonstrated that incorporating comprehensive echocardiographic data improves survival prediction following Myocardial infarction (MI), suggesting that integration of such parameters could enhance our model’s performance in future iterations.¹⁹ Future research should focus on prospective validation and assessment of the model’s impact on clinical outcomes when used in practice. Integration with electronic health records could further enhance its utility in real-time decision support.

Conclusion

Our ML-based risk prediction model for 1-year mortality in STEMI patients demonstrates high accuracy and clinical applicability. The development of a user-friendly web application facilitates the implementation of this model in clinical practice, potentially improving patient care through more precise risk assessment and tailored treatment strategies.

Data availability statement

All data relevant to the study are included in the article or uploaded as supplementary information.