Machine learning to predict long-term cardiovascular death following myocardial infarction: incremental value of echocardiographic data

Abstract

Aims

Machine learning (ML) for prediction of cardiovascular (CV) death following myocardial infarction (MI) has not been well studied. This study sought to define the incremental value of (i) integrating comprehensive echocardiographic data in ML models and (ii) ML approaches over Cox Regression (CPH), for predicting CV death following MI.

Methods and results

Retrospective cohort study of consecutive patients with MI admitted at a tertiary referral hospital, with echocardiography performed within 24 h of admission. Models were trained on a cohort admitted between 2013 and 2017 (n & 1568) and validated on a separate temporal holdout cohort from 2018 to 2021 (n & 1634). Two ML models Gradient Boosted Cox and a DeepSurv Neural Network were developed and compared with conventional multivariable Cox regression. The SHapley Additive exPlanations (SHAP) method was used for ML model interpretation. In the final study population of 3202 patients (mean age 63.2 ± 12.5 years; 29.2% females), 28.8% had ST-elevation MI and the mean left ventricular ejection fraction (LVEF) was 52.5 ± 11.2%. At a median follow-up of 4.5 years, there were 139 (4.3%) CV deaths. In the validation set, Gradient Boosted Cox achieved the highest performance (C-index 0.861), compared with conventional Cox regression (C-index 0.813, P & 0.037) and the DeepSurv Neural Network (C-index 0.847, P & 0.38) for the prediction of CV death. Within the GB Cox model, 14 out of the top 20 features for predicting CV death were echocardiographic variables, including LV size, LVEF, and diastolic parameters. Further, in nested ML models, the addition of echocardiographic parameters provided incremental value beyond clinical variables + LVEF alone (C-index 0.861 vs. 0.792, P & 0.017).

Conclusion

ML integration of comprehensive echocardiographic data leads to improved prediction of CV death following MI, with key measures of LV size and systolic and diastolic function contributing substantially to prognostic models.

Graphical Abstract

Graphical Abstract.

Introduction

Patients with myocardial infarction (MI) remain at significantly increased risk of recurrent cardiac events and death following their first event.^1–3 Identifying individuals at high risk of cardiovascular death is critical so that more intensive preventive cardiovascular measures can be applied. Machine learning (ML) can potentially improve risk prediction following MI, but previous studies have produced mixed results for stratifying prognosis following MI.^4,5 Previous studies of ML for risk stratification following MI have used all-cause mortality as the most common endpoint.^4–6 Whilst focusing on all-cause mortality is a useful surrogate for estimating cardiac survival, studies have shown that around two thirds of all deaths following MI are due to non-cardiac causes.⁷ Therefore, focusing on cardiovascular death may be more relevant as this provides an estimate of mortality directly related to the myocardial insult as a result of MI. Furthermore, many previous studies of ML for risk stratification following MI did not incorporate echocardiographic data beyond left ventricular ejection fraction (LVEF).^5,6 Whilst LVEF is the most commonly used echocardiographic parameter for risk stratification clinically, a large body of literature shows that several measures of LV size, geometry and diastolic function, may be superior to LVEF for predicting outcomes following MI.⁸ Therefore, combining the prognostic utility of multiple prognostically validated parameters alongside LVEF using ML algorithms may produce better risk stratification compared to using LVEF alone.

Accordingly, the current study was designed to test the prognostic utility of ML prognostic models combining comprehensive clinical and echocardiographic data for predicting CV death following MI. In a previous study, our group demonstrated the utility of combining comprehensive echocardiographic data alongside clinical and angiographic data for predicting all-cause mortality following MI.⁹ We re-analysed our data to examine the impact of combining comprehensive echocardiographic data with clinical and angiographic data for the purposes of predicting CV death. Specifically, we sought to define (i) the incremental prognostic value of integrating comprehensive echocardiographic data in ML models and (ii) the incremental value of ML survival approaches over standard Cox Regression, for predicting CV death following MI. We hypothesized that ML survival models would outperform conventional Cox Regression models, enable precise identification of the predictors contributing to mortality risk, and that incorporating comprehensive echocardiographic data would enhance risk prediction with ML following MI. While our group has previously explored the use of ML for predicting all-cause mortality, this study provides a distinct and clinically crucial contribution by focusing specifically on CV death.⁹ This endpoint is more directly related to the pathophysiology of the index MI and is of greater relevance when considering targeted cardiovascular interventions. By isolating cardiovascular events, we aim to build a more specific and actionable risk prediction tool, addressing a key gap in literature.

Methods

Study overview

A total of 3464 consecutive patients with MI (ST-elevation MI [STEMI] and non-ST-elevation MI [NSTEMI]) during the study period between January 2013 and December 2021 at a single tertiary level referral centre was considered for inclusion in this study. Exclusion criteria included significant haemodynamic instability (shock, acute pulmonary oedema, requirement for mechanical ventilation, inotropes, intra-aortic balloon pump, and those with ventricular tachyarrhythmia), in-hospital death, limited or point of care echocardiograms, and incomplete clinical, echocardiographic, or follow-up data. All clinical, angiographic, and echocardiographic data were retrospectively collected from prospectively maintained institutional cardiac catheterization and echocardiographic databases. Outcome data were obtained from the State Births, Deaths and Marriages registry, where each patient is tracked through a unique identification number, and which is linked to the National Death Registry. Cause of death was adjudicated by study investigators (EX and SP) on the basis of the Medical Certificate of Cause of Death issued by a medical practitioner upon death of the patient and individually cross-checked with the medical records, including access to postmortem data where available, at the study institution by the investigators. In cases of disagreement, a final consensus adjudication was achieved through joint review of medical records.

Outcome measures and definitions

The primary outcome measures were cardiovascular death and non-cardiovascular death, defined according to currently accepted conventional definitions.^10,11 CV death was further sub-categorized as ischaemic, heart failure related or arrhythmic (including sudden cardiac death). The underlying root cause of death was classified on the basis of the World Health Organization definition which defines cause of death as the disease or injury which initiates the train of morbid events leading directly to death, or the circumstances of the accident or violence which produced the fatal injury.^10,11 CV death was defined as death primarily due to cardiovascular causes in patients who had symptoms and signs of cardiac dysfunction leading up to death, including arrhythmia, heart failure, or ischaemia. Patients who could not be classified according to these definitions on the basis of the available information were classified as indeterminate. The primary endpoint was time to cardiovascular death. Patients were followed from the date of admission until the date of death or the end of the study period. Patients who were alive at the end of follow-up or died from non-cardiovascular causes were censored at the date of last contact or death, respectively.

Institutional protocol for MI

All patients with MI at the study institution were considered for an invasive approach unless significant contraindications existed. The default strategy for management of STEMI was primary PCI with 24-h catheterization laboratory activation (mean door to balloon time 73 ± 12 min) and an early invasive approach for NSTEMI. The completeness of revascularization was at the discretion of the operator depending on the patient’s condition, with an emphasis on culprit lesion PCI in primary PCI. All patients were started on evidence based medical therapy for MI including aspirin, dual antiplatelet therapy, statins, angiotensin converting enzyme inhibitors, and beta-blockers on admission unless contraindications existed.^2,3

Echocardiogram protocol and measures

A comprehensive transthoracic echocardiogram was performed within 24 h of admission for all patients. All echocardiograms were performed on either General Electric (GE) Vivid E9 or E95 machines (Horten, Norway) or Phillips IE33 or EPIQ machines (Andover, MA, USA) with tissue Doppler imaging (TDI) software and a 2.5–5 MHz variable frequency, phased array transthoracic transducer. The echocardiography protocol, performed by experienced clinical sonographers, followed a standard format with image acquisition from the parasternal, apical and subcostal acoustic windows, and included 2D, colour flow mapping, continuous and pulse-wave Doppler, and TDI.¹² All measurements obtained were in accordance with current ASE recommendations.¹² Left ventricular systolic function was assessed by LV ejection fraction (LVEF) obtained using Simpson’s biplane method of discs from the apical 4 and 2 chamber views.¹² Left ventricular volumes were obtained from Simpson’s biplane method of discs, and indexed to body surface area. Left ventricular mass was obtained using Devereaux’s formula.¹² Diastolic variables were obtained as recommended in current guidelines.¹³ Mitral inflow Doppler was obtained using pulsed wave Doppler. TDI was obtained at the septal and lateral mitral annulus. E/e′ ratio was calculated using the early mitral inflow E wave velocity and septal e′ (E/e′ septal), lateral e′ (E/e′ lateral), and the average of septal and lateral e′ (E/e′ average). Left atrial volume index was assessed using a Simpson’s biplane method with an inbuilt disk summation algorithm on the echo machines used in this study.¹² TRV was obtained from continuous wave Doppler signal of tricuspid regurgitation jets. We acknowledge that the timing of the echocardiogram relative to coronary angiography differed between MI subtypes (typically following angiography in STEMI and prior to angiography in NSTEMI); this was unavoidable due to the distinct clinical management pathways for these presentations.

Statistical analysis

Continuous variables are presented as mean ± standard deviation and compared using an unpaired t-test. Categorical variables are presented as percentages and compared the Chi-squared test. A P < 0.05 was considered significant. A conventional multivariable Cox Regression model was fitted as a baseline comparator. To address collinearity among the 59 candidate predictors, backward elimination feature selection was performed using five-fold cross-validation on the training cohort. This process identified an optimal set of 18 predictors (cross-validated C-index & 0.844). The temporal validation cohort was held out entirely during feature selection and was used only for final model evaluation. The proportional hazards assumption was checked using Schoenfeld residuals. Results are presented as hazard ratios (HRs) with 95% confidence intervals (CIs). The performance of all models was evaluated using the Harrell’s Concordance Index (C-index).¹⁴ Pairwise comparisons between C-indices were performed using a bootstrap test (20 000 resamples) with a two-sided z-test for the mean difference.

ML model construction

A total of 59 clinical and echocardiographic variables (Tables 1 and 2) were included in creation of the models for prediction of CV death. For ML, a gradient boosting survival model (Gradient Boosted Cox) was chosen as the primary algorithm, with a DeepSurv neural network and a standard Cox Regression model serving as comparators. Gradient Boosted Cox fits an ensemble of regression trees using the negative log partial likelihood of the Cox model as the loss function, enabling it to capture non-linear relationships and higher-order interactions while retaining the semi-parametric survival framework.¹⁵ DeepSurv extends the Cox model by replacing the linear predictor with a feed-forward neural network, optimized via the same Cox partial likelihood loss.¹⁶ Variables with >25% missing data were excluded from analyses. The remainder of missing data was imputed using the median (for continuous variables) or mode (for categorical variables). To prevent data leakage, imputation values were derived from the training cohort and applied to the validation cohort. Data splitting (temporal split) was performed prior to imputation of missing values, model training, and hyperparameter tuning.

Table 1.

Baseline characteristics of training and validation cohorts

Baseline characteristics	All	Pre-2018	Post-2018	P-value
Count	n = 3202	n = 1568	n = 1634
Cardiovascular death (Outcome)	139 (4.3%)	103 (6.57%)	36 (2.20%)	<0.001
Demographics
Age	63.1 (12.5)	63.0 (12.6)	63.3 (12.5)	0.413
Sex (Male)	2268 (70.8%)	1116 (71.2%)	1152 (70.5%)	0.705
BMI	29.4 (8.9)	29.2 (6.1)	29.6 (10.9)	0.204
Height (cm)	171.5 (9.3)	171.4 (9.3)	171.7 (9.3)	0.381
Weight (kg)	86.6 (19.9)	86.2 (19.0)	87.0 (20.7)	0.255
Body surface area	2.0 (0.3)	2.0 (0.3)	2.0 (0.3)	0.307
Laboratory
Creatinine	101.0 (90.7)	100.0 (89.2)	102.0 (92.2)	0.532
eGFR	73.2 (20.3)	73.4 (19.9)	73.0 (20.7)	0.583
CKD (eGFR <90)	2110 (65.9%)	1029 (65.6%)	1081 (66.2%)	0.780
History
Diabetes	902 (28.2%)	447 (28.5%)	455 (27.8%)	0.706
Hypertension	1709 (53.4%)	834 (53.2%)	875 (53.5%)	0.866
Hypercholesterolemia	1493 (46.6%)	768 (49.0%)	725 (44.4%)	0.010
Smoking	1934 (60.4%)	923 (58.9%)	1011 (61.9%)	0.088
End stage renal disease	39 (1.2%)	33 (2.1%)	6 (0.4%)	<0.001
Previous MI	76 (2.4%)	56 (3.6%)	20 (1.2%)	<0.001
Previous PCI	278 (8.7%)	154 (9.8%)	124 (7.6%)	0.029
Previous CABG	165 (5.2%)	92 (5.9%)	73 (4.5%)	0.087
Family history	1176 (36.7%)	542 (34.6%)	634 (38.8%)	0.014
Previous CVA	30 (0.9%)	24 (1.5%)	6 (0.4%)	0.001
Prosthetic MV	5 (0.2%)	4 (0.3%)	1 (0.1%)	0.209
Prosthetic AV	22 (0.7%)	10 (0.6%)	12 (0.7%)	0.907
Presentation
STEMI	921 (28.8%)	409 (26.1%)	512 (31.3%)	0.001
Rhythm
Atrial fibrillation	337 (10.5%)	160 (10.2%)	177 (10.8%)	0.602
Angiographic
Three VD	483 (15.1%)	222 (14.2%)	261 (16.0%)	0.166
PCI	1769 (55.2%)	853 (54.4%)	916 (56.1%)	0.364
Medical management	1024 (32.0%)	492 (31.4%)	532 (32.6%)	0.498
LV size, geometry, and systolic function
IVS diastolic thickness	1.1 (0.2)	1.1 (0.2)	1.1 (0.2)	<0.001
LV PW diastolic thickness	1.0 (0.2)	1.0 (0.2)	1.0 (0.2)	<0.001
LV diastolic diameter	47.1 (6.3)	47.0 (6.2)	47.2 (6.5)	0.382
LV systolic diameter	33.2 (6.1)	33.0 (6.2)	33.3 (5.9)	0.129
LV diastolic volume MOD BP	104.6 (29.3)	103.2 (30.1)	106.1 (28.5)	0.005
LV systolic volume MOD BP	50.3 (22.5)	49.5 (23.0)	51.1 (22.0)	0.042
LV end-diastolic volume index	51.8 (12.8)	51.3 (13.4)	52.4 (12.2)	0.011
LV end-systolic volume index	24.8 (10.5)	24.4 (10.8)	25.1 (10.2)	0.066
LV EF	52.5 (11.2)	52.8 (11.1)	52.3 (11.3)	0.251
LV mass index	85.6 (22.4)	87.4 (22.6)	83.9 (22.0)	<0.001
LV OT peak velocity	97.5 (19.2)	97.4 (18.6)	97.6 (19.7)	0.765
LV OT velocity time integral	19.3 (4.4)	19.4 (4.3)	19.1 (4.5)	0.098
LV diastolic parameters
LA area	20.4 (4.2)	20.5 (4.1)	20.4 (4.4)	0.459
LA volume index	32.3 (10.0)	32.6 (9.9)	32.0 (10.1)	0.093
Mitral A point velocity	71.4 (22.3)	71.7 (23.1)	71.2 (21.6)	0.469
Mitral E point velocity	70.3 (22.4)	71.8 (23.6)	68.8 (21.0)	<0.001
Mitral E to A ratio	1.0 (0.5)	1.1 (0.5)	1.0 (0.4)	0.034
MV deceleration time	217.2 (62.2)	219.0 (62.4)	215.5 (61.9)	0.116
LV Ep lateral velocity	7.9 (2.3)	8.0 (2.4)	7.8 (2.2)	0.005
LV Ep septal Velocity	6.1 (1.7)	6.2 (1.8)	6.0 (1.7)	0.002
Mitral E to LV Ep average ratio	9.2 (1.8)	9.0 (0.0)	9.3 (2.5)	<0.001
Valve Parameters
Aortic stenosis (Mod/Sev)	75 (2.3%)	38 (2.4%)	37 (2.3%)	0.857
Aortic regurgitation (Mod/Sev)	22 (0.7%)	13 (0.8%)	9 (0.6%)	0.460
Mitral regurgitation (Mod/Sev)	212 (6.6%)	106 (6.8%)	106 (6.5%)	0.811
Mitral stenosis (Mod/Sev)	1 (0.0%)	0 (0.0%)	1 (0.1%)	1.000
Tricuspid regurgitation (Mod/Sev)	62 (1.9%)	36 (2.3%)	26 (1.6%)	0.187
Right heart parameters
RA area	15.5 (3.6)	15.6 (3.5)	15.3 (3.7)	0.026
Right atrial pressure	4.5 (3.0)	4.7 (3.1)	4.3 (2.9)	<0.001
Right ventricular systolic pressure	30.8 (7.3)	30.9 (7.2)	30.8 (7.5)	0.781
RV diastolic basal diameter	3.6 (0.6)	3.6 (0.0)	3.6 (0.8)	0.436
RV OT velocity time integral	14.2 (2.5)	14.3 (2.9)	14.0 (2.1)	0.001
RV systolic lateral velocity	15.5 (199.0)	11.5 (2.1)	19.4 (278.5)	0.252
Tricuspid annular plane systolic excursion	2.0 (0.1)	2.0 (0.1)	2.0 (0.1)	0.549
TR peak velocity	246.9 (32.0)	247.1 (31.1)	246.8 (32.8)	0.773

Split by admission date of 01 January 2018.

Table 2.

Performance metrics of all models on the temporal holdout set at 4 years

Model	C-index (95% CI)	AP	Sens	Spec	PPV	NPV	F1	Brier
GB Cox (Full Echo)	0.861 (0.821–0.900)	0.140	0.944	0.728	0.254	0.993	0.400	0.020
DeepSurv NN	0.847 (0.803–0.887)	0.133	0.694	0.875	0.352	0.967	0.467	0.019
Cox Regression Baseline	0.813 (0.763–0.857)	0.106	0.722	0.785	0.248	0.967	0.369	0.018
GB Cox (Clinical + LVEF)	0.792 (0.724–0.851)	0.129	0.667	0.826	0.273	0.962	0.387	0.019
GB Cox (Clinical)	0.793 (0.740–0.841)	0.103	0.861	0.652	0.195	0.98	0.318	0.02

Performance metrics for prediction of 4-year cardiovascular death on the temporal validation cohort (patients from 2018 onwards). C-index represents time-dependent concordance index across all time points with 95% confidence intervals derived from 2,000 bootstrap samples. Classification metrics (Sensitivity, Specificity, PPV, NPV, F1-score) are evaluated at the 4-year horizon using Youden’s optimal threshold applied to predicted 4-year event probabilities; patients censored before 4 years without an event were excluded from classification metrics to ensure clinically evaluable outcomes. AP (Average Precision) reflects area under the precision-recall curve using 4-year predicted probabilities. Brier score represents inverse probability of censoring weighted (IPCW) calibration metric at 4 years, with lower values indicating better agreement between predicted probabilities and observed outcomes (range 0–1).

Abbreviations: NN, Neural Network; GB, Gradient Boosting; LVEF, left ventricular ejection fraction; AP, average precision; Sens, sensitivity; Spec, specificity; PPV, positive predictive value; NPV, negative predictive value; F1, F1-score (harmonic mean of precision and recall); Brier, Brier score.

Model training

A temporal validation strategy was employed, with the entire cohort split based on admission date. Patients admitted between January 2013 and December 2017 were assigned to the training cohort (n = 1568). Model development and comparison followed a rigorous, standardized framework. All feature scaling and missing-value imputation were performed after the temporal split, with parameters derived exclusively from the training cohort and applied to the validation cohort, to prevent data leakage. Continuous features were standardized to zero mean and unit variance using the training-set distribution, with HRs representing the change in risk per 1 standard deviation increase. Hyperparameters for the Gradient Boosted Cox model were selected through systematic evaluation on the training cohort using internal cross-validation and are reported in Supplementary material online, Table S1. Following hyperparameter selection, all three models were trained on the entire training cohort. Detailed specifications of the DeepSurv architecture can be found in the Supplementary material online, Methods.

Model testing

Patients admitted between January 2018 and December 2021 constituted the holdout temporal test cohort (n & 1634), which was used for final, unbiased model evaluation. For each patient, each model generated a continuous risk score representing the predicted log HR relative to the baseline hazard. The primary measure of discriminative performance was Harrell’s C-index. Given the class imbalance, the area under the precision-recall curve was also calculated at 4 years to provide a more robust assessment of model performance on the positive class. Secondary measures included sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1 score. For the best performing model, the Youden’s J statistic (J & Sensitivity + Specificity − 1) was calculated and used to derive the optimal prognostic cutoff for the ML score. This statistic identifies the threshold that maximizes the distance from the no-discrimination line, thereby providing an optimal balance between sensitivity and specificity. This threshold was then used to classify patients into high-risk and low-risk groups. To account for the competing risk of non-cardiovascular death, cumulative incidence functions were generated to visualize the probability of cardiovascular death in the high- and low-probability groups. Differences between the cumulative incidence curves were assessed using Gray’s test. Calibration was assessed visually using a calibration plot, which compares the mean predicted probability against the observed fraction of events within quintiles of predicted probability, at 1-, 2-, 3- and 4-year time horizons. Agreement between the predictions and observed outcomes was quantified using the inverse probability of censoring weighted Brier score (range 0–1), with smaller values indicating better calibration.^17,18 Predicted event probabilities were derived from the Breslow baseline survival function estimated on the training cohort, ensuring no data leakage.¹⁹ Finally, decision curve analysis was performed at a 4-year horizon to evaluate the clinical net benefit of each model across a range of threshold probabilities.

ML interpretability

To assist with clinical interpretability, the SHAP method was used to generate a summary plot for the entire cohort and to explain individualized ML prediction for representative patients with high vs. low ML scores. This is a game theory-based statistical method that is used to address the ‘black box’ nature often associated with complex algorithms such as Gradient Boosted Cox, by calculating the contributions of each feature to the overall prediction.

Nested model training

To assess the incremental value of echocardiographic parameters over and above clinical data, three nested Gradient Boosted Cox models were trained and compared on the temporal test set: (i) a baseline model using only clinical variables (Clinical-Only), (ii) a model using clinical variables plus LVEF (Clinical + LVEF), and (iii) the full model incorporating all clinical and comprehensive echocardiographic variables (Full Model). These nested models were developed and evaluated using the same training and testing set as the comprehensive model. Pairwise differences in C-index were calculated to quantify incremental prognostic value. Additionally, to evaluate the feasibility of a streamlined clinical tool, we developed a parsimonious model using only the top eight features ranked by importance and assessed its non-inferiority compared to the full 59-variable model.

Sensitivity analysis for competing risks

To address the competing risk of non-cardiovascular death, a sensitivity analysis was performed by excluding all patients who died from non-cardiovascular causes from both the training (n = 214 excluded) and validation (n = 94 excluded) cohorts. The Gradient Boosted Cox models (full and clinical-only) were then re-trained and re-evaluated on these modified datasets. This approach isolates the model’s ability to discriminate cardiovascular death from survival, without potential confounding from non-cardiovascular mortality treated as censored observations in the primary analysis.

Ethical considerations

The study was approved by the institutional Human Research Ethics committee (Metro North Hospital and Health Service Human Research Ethics Committee, Study number: HREC/2022/QPCH/86459). A separate Public Health Act mandated permission was obtained to access data stored on institutional databases.

Results

Study population and baseline characteristics

The study included 3202 patients with MI (29.2% female) after exclusions, with a mean age of 63.2 ± 12.5 years. Detailed patient flow data and exclusions are shown in Figure 1. Of these, 465 patients experienced all-cause mortality at a median follow-up period of 4.5 years, with 139 classified as CV death. Of the 139 adjudicated CV death in the cohort, the primary causes were classified as heart failure-related in 78 cases (56.1%), arrhythmic or sudden cardiac death in 34 cases (24.5%), and ischaemic (including recurrent MI) in 27 cases (19.4%). Baseline clinical characteristics comparing those with and without CV death are summarized in Supplementary material online, Table S2. Importantly, age, diabetes, hypertension, chronic renal failure, NSTEMI presentation, atrial fibrillation on presentation, three-vessel disease, and medical management were all associated with CV death on univariate comparisons. Baseline echocardiographic characteristics comparing those with and without CV death are presented in Supplementary material online, Table S3, encompassing 35 different variables. Importantly, LV size, wall thickness, LVEF, LV mass, multiple diastolic variables (left atrial size, diastolic tissue velocities, E/e′ ratios, tricuspid regurgitation velocities), and severity of aortic stenosis and mitral regurgitation were all associated with CV death on univariate comparison.

Figure 1.

Flow of patients, exclusions and included patients, along with their splits for prediction algorithms.

Training and validation cohorts

The training (2013–17) and validation (2018–21) cohorts were well-matched for key demographic variables, including age, sex, and BMI (Table 1). However, several clinically relevant differences reflective of evolving patient populations and treatment patterns were observed. Notably, the absolute rate of cardiovascular death was lower in the more recent validation cohort (2.20%) compared to the training cohort (6.57%, P < 0.01). Patients in the validation cohort also presented more frequently with STEMI (31.3% vs. 26.1%, P & 0.001). The robust performance of the model despite these temporal shifts in baseline characteristics and event rates underscores its potential generalizability.

Results of cox regression analysis

The results of multivariable Cox regression analysis are presented in Supplementary material online, Table S4 and listed in order of HR. The strongest independent clinical predictors of shorter time to cardiovascular death included medical management (HR 1.60, 95% CI 1.32–2.03, P < 0.001), diabetes (HR 1.42, 95% CI 1.19–1.78, P < 0.001), male sex (HR 1.38, 95% CI 1.10–1.87, P & 0.003), and three-vessel disease (HR 1.36, 95% CI 1.14–1.62, P & 0.001). Among echocardiographic variables, markers of diastolic function were strong predictors of mortality, including higher mitral E-point velocity (HR 1.43, 95% CI 1.19–1.72, P < 0.001) and mitral A-point velocity (HR 1.41, 95% CI 1.13–1.81, P & 0.004). Conversely, several factors were identified as protective, associated with a reduced risk of cardiovascular death. These included higher LVEF (HR 0.68, 95% CI 0.55–0.83, P < 0.001), larger body surface area (HR 0.61, 95% CI 0.44–0.82, P & 0.003), and a family history of coronary artery disease (HR 0.77, 95% CI 0.59–0.95, P & 0.020).

ML model performance

A comparison of the test performance of the Cox Regression model with the two ML models (Gradient Boosted Cox and Neural Network) is presented in Table 2. The Gradient Boosted Cox model demonstrated the highest discriminative performance (C-index 0.861, 95% CI 0.821–0.900), outperforming the standard Cox Regression model (C-index 0.813, 95% CI 0.763–0.857; P = 0.037), but not the DeepSurv Neural Network (C-index 0.847, P = 0.38). Figure 2 illustrates the C-index comparison. In terms of calibration, the Gradient Boosted Cox model demonstrated reasonable agreement between predicted and observed risks across 1- to 4-year time horizons (see Supplementary material online, Figure S1), though with a tendency to slightly overestimate risk in the highest quintiles at later time points, likely reflecting the lower event rate in the validation cohort. Cumulative incidence curves (Figure 3) demonstrated a clear separation between high- and low-probability groups (Gray’s test P < 0.001). The SHAP beeswarm plot (Figure 4) highlights the most influential predictors in the Gradient Boosted Cox model, with age, mitral E wave velocity, LVEF, diabetes, LA volume, LV mass, eGFR, mitral A wave velocity, and LV wall thickness comprising the top 10 predictors with the largest impacts on CV death prediction.

Figure 2.

C-Index model comparison. Comparison of Harrell’s Concordance Index (C-index) for the Gradient Boosted Cox, DeepSurv, and standard Cox regression models, as well as the nested Gradient Boosted Cox models (Clinical + LVEF and Clinical Only) on the validation cohort. Error bars represent 95% CIs derived from bootstrapping. The Gradient Boosted Cox model significantly outperformed the standard Cox model (P & 0.023).

Figure 3.

Cumulative incidence curves for cardiovascular death by gradient boosted cox model. Cumulative incidence of cardiovascular death stratified by the Gradient Boosted Cox model risk score. Patients were stratified into high-risk and low-risk groups based on the optimal threshold. Gray’s test indicates a highly significant difference between groups (P < 0.001).

Figure 4.

Gradient boosted cox SHAP beeswarm plot. Visual representation of feature importance for the survival model. Features are ranked by the sum of SHAP value magnitudes. Each dot represents a patient. Red indicates a high feature value, while blue indicates a low feature value. A positive SHAP value (right side of axis) indicates an increased hazard (higher risk of death), while a negative value indicates a decreased hazard.

Incremental value of echocardiography

Nested Gradient Boosted Cox models demonstrated the incremental value of comprehensive echocardiographic data (Figure 2). The Clinical Only achieved a C-index of 0.793. The addition of LVEF (Clinical + LVEF) did not yield an improvement in the C-index (0.792, P & 0.949 vs. Clinical Only). However, the Full Model incorporating comprehensive echocardiographic parameters outperformed the Clinical + LVEF model (C-index 0.861; Δ0.070, P & 0.017).

Decision curve analysis

Decision curve analysis at a 4-year time horizon (see Supplementary material online, Figure S4) illustrated that the Gradient Boosted Cox model offered the highest net benefit across the majority of clinically relevant threshold probabilities, indicating superior clinical utility for decision-making compared to the standard Cox model and default strategies.

Individualized explanations

Figure 5 shows waterfall plots for two individual patients, illustrating how the Gradient Boosted Cox model arrived at its prediction for each case. The baseline value (0.016) represents the average log hazard for the dataset. Each bar represents a feature’s contribution to moving the individual’s risk score away from this baseline. Patient 3 (Top) illustrates a low-risk profile (Risk Score −0.261; 4-year predicted risk 3.3%) where younger age (45 years) and normal diastolic function (Mitral E-wave 47.9 cm/s) acted as strong protective factors. In contrast, Patient 9 (Bottom) illustrates a high-risk profile (Risk Score 0.378; 4-year predicted risk 6.1%) who subsequently died of cardiovascular causes. Crucially, this patient had a preserved LVEF (57.5%), which the model correctly identified as a protective feature (blue bar). However, the model assigned a net high risk due to the overwhelming contribution of non-systolic echocardiographic markers: elevated Mitral E-wave velocity (136 cm/s), severe left atrial enlargement (LA volume index 126 mL/m²), and pulmonary hypertension (TR peak velocity 345 cm/s). This case exemplifies the incremental value of comprehensive echocardiography, as a model relying solely on LVEF might have underestimated this patient’s risk.

Figure 5.

Gradient boosted cox SHAP waterfall plot. Visual representation of two selected patients and how the model used the attributes to calculate their final score. Beginning from the mean prediction at the bottom (ML score of 1.265), the waterfall plots display SHAP(SHapley Additive exPlanations) values in increasing order of magnitude, with red for increasing risk and blue for decreasing risk of death. The positive or negative SHAP value of each feature is progressively added to reach the individual prediction. Case 1 shows a 75 year old patient with a high Mitral E point velocity and LVEF of 18.3%, without diabetes that is given a high score of 1.839, and subsequently experienced cardiovascular death. Case 2 shows a patient with LVEF of 69.3%, LVOT of 27.1 and LVESVI of 12.8, with diabetes who was given a low score of −1.468 by the model, and did not have subsequent cardiac death recorded. In this example based on the ML score a different approach to management could be taken.

Sensitivity analysis for competing risks

After excluding 308 patients who died from non-cardiovascular causes (training: 214; validation: 94), the full Gradient Boosted Cox model maintained excellent discrimination (C-index 0.864, 95% CI 0.809–0.915), comparable to the primary analysis (C-index 0.861). The incremental value of comprehensive echocardiographic data over clinical variables alone persisted (C-index 0.864 vs. 0.806, P & 0.011), confirming the robustness of the main findings to the competing risk of non-cardiovascular death.

Parsimonious model for potential clinical application

To evaluate the feasibility of a streamlined clinical tool, we developed and validated a parsimonious Gradient Boosted Cox model using only the top eight most influential features identified by SHAP analysis. This simplified model, representing an 86% reduction in feature complexity (from 59 to 8 variables), achieved a C-index of 0.857 (95% CI 0.819–0.891). This performance was statistically non-inferior to the full 59-variable model (C-index 0.861, P = 0.66), as shown in Supplementary material online, Figure S2. This finding has significant implications for clinical translation, suggesting that a highly accurate risk prediction tool can be implemented using a small, manageable set of key clinical and echocardiographic variables, thereby reducing the data collection burden and facilitating integration into electronic health record workflows.

Discussion

Summary

The main findings of this study are that ML integration of comprehensive echocardiographic data leads to improved prediction of CV death following MI, with key measures of LV systolic and diastolic function contributing substantially to prognostic models. In the final Gradient Boosted Cox model, 14 out of the top 20 predictors (70%) were from the baseline echocardiogram, highlighting the powerful role of echocardiographic data when ranked alongside important clinical data. Of the top 10 highest ranked variables, seven were echocardiographic measurements, including LVEF, four diastolic parameters (mitral E and A wave velocities, left atrial volume, septal e′ velocity), LV mass and LV wall thickness. These findings are broadly consistent with previous studies demonstrating that diastolic parameters and LVMI as well as LVEF are powerful independent predictors for prognosis following MI.^20–23 Crucially, the high ranking of variables like mitral E-wave velocity and LA volume index in our model validates the inclusion of comprehensive diastolic and structural assessment, demonstrating their essential role in accurate risk stratification. The incremental value of adding comprehensive echocardiographic data to ML models was further illustrated in the sequential, nested ML models, where the addition of comprehensive echocardiographic data over and above clinical data resulted in significant improvement of model discriminative performance (C-index 0.86 vs. 0.79). In the broader context, ML models built with Gradient Boosted Cox incorporating comprehensive echocardiographic data have the potential to improve risk stratification for CV death following MI.

Previous studies of ML for risk prediction following MI and echocardiographic inclusions

Whilst there have been a number of previous studies investigating the value of ML for risk prediction following MI, most have focused on all-cause mortality and none have included comprehensive echocardiographic data beyond LVEF (most common measurement included in a number of studies), MR severity (1 study), and regional wall motion index (two studies) in predictive models.^4,5,24–26 In contrast, our study focused on CV death and included a comprehensive range of prognostically validated echocardiographic measures, including measures of LV size, geometry, mass, systolic and diastolic function, as well as comprehensive valvular assessment.²⁷ While a comprehensive review of all published studies is beyond the scope of this discussion, it is notable that the majority of prior ML models for risk prediction following MI have utilized all-cause mortality as the primary outcome. Furthermore, echocardiographic data in these studies, when included, have often been limited to basic parameters such as LVEF. Of the few studies that studied CV death, echocardiographic inclusions were rudimentary: Zoni-Berrisso included LVEF and segmental LV dyskinesis, Myers et al. included LVEF only, and Xiao et al. (primary endpoint composite MACE) included LVEF and LVEDD.^28–30 As highlighted previously, a number of studies have confirmed the independent prognostic value of a range of echocardiographic parameters beyond LVEF, including diastolic parameters, LV mass, diastolic function, pulmonary haemodynamics, and severity of valvular lesions, all of which were available for inclusion in our present study.²⁷ Mechanistically, the high predictive ranking of diastolic indices in our model, specifically mitral E-wave velocity and LA volume index, likely reflects their ability to capture different temporal aspects of haemodynamic stress. A high mitral E-wave velocity is a marker of elevated acute LV filling pressures and increased stiffness, whereas LA enlargement serves as a barometer of chronic diastolic dysfunction and cumulative haemodynamic burden. Both physiological states are potent drivers of post-MI heart failure and mortality, explaining their retained prognostic value even in the absence of advanced strain imaging.

Advantages of ML for risk prediction following MI

ML has a significant advantage over traditional statistical models for risk prediction in that it can integrate a large number of variables with in-built techniques to avoid overfitting and can also be used to explore the complex inter-relationships between variables. ML may be particularly useful for integrating a diverse range of echocardiographic data: given that there are a number of measures of LV size, mass, and diastolic function that have been shown to be superior to LVEF for predicting survival following MI, integrating all these prognostically validated measures synergistically using ML models has the potential to improve risk prediction following MI.²⁷ A key advantage of our approach is its clinical interpretability, which addresses the common ‘black box’ criticism of ML through the use of SHAP. SHAP provides both global insights into the most influential predictors across the cohort (Figure 4) and, crucially, local explanations for individual patient predictions (Figure 5). The ability to explain why a specific patient is flagged as high-risk is essential for building clinician trust and transforming the model from a score into a true decision support tool. This granular explainability is a distinct advantage of tree-based models like Gradient Boosted Cox.

Gradient boosting cox ML for predicting outcomes following MI

Previous studies including Khera et al. and Shetty et al. have highlighted the potential superiority of gradient boosted techniques for risk stratification following MI. Gradient Boosted Cox offers several theoretical advantages in a dataset of this size and type. Firstly, Gradient Boosted Cox handles mixed data types (numerical and categorical) better than neural networks, which require pre-processing into numerical data. Secondly, its tree-based structure captures non-linear relationships better without specific feature engineering on smaller tabular datasets (n & 3202), whereas neural networks require extensive tuning on much larger datasets to learn these patterns effectively. Lastly, one of the key strengths of Gradient Boosted Cox ML is its clinical interpretability through SHAP values, allowing for clear visualization of individual risk drivers (log HRs). Neural networks lack this interpretability and are more of a ‘black box’ technique.

Clinical translation

ML models could potentially be implemented in clinical practice as a clinical decision support tool, guiding risk stratification and outpatient management of high vs. low-risk patients, with intensification of therapy or closer surveillance as appropriate. A potential way to achieve this would be integrate these algorithms into digitally integrated electronic medical record systems as a risk calculator with risk displayed on a dashboard accessible to clinicians. In the present study, all baseline data were derived from three digital data repositories: the institutional cardiac catheterization, laboratory, and echocardiography databases. Whilst these were stand-alone repositories and required manual integration of the data from the three sources following extraction for the purposes of this study, using current technology, it is now feasible to integrate these databases using either a single repository for all data or using vendor neutral archiving to a common data lake to achieve automation and streamlining of accessing key data points. Additionally, using online database management tools such as Research Electronic Data Capture (REDCap) makes this type of integration across separate clinical databases accessible for clinicians. The potential clinical impact of using the ML based prognostic score includes improved patient risk stratification, better targeted therapies, and potentially more efficient use of device-based therapies. A limited number of previous studies have attempted to use innovative approaches to combine multiple echocardiographic variables, including using ML/artificial intelligence methods to image analysis, and the emerging data suggests that using combination approaches may be superior to LVEF alone, but further validation and proof of concept are required, including randomized clinical trials to test any novel ML based risk score.

Limitations

This study has a number of limitations. Whilst the data presented in this study included patients from 16 hospitals within a single state (1 central referral hospital and 15 peripheral referring hospitals), true external validation in an independent dataset was not performed, and generalizability could be improved with inclusion of multi-state and multi-national patient cohorts. The retrospective nature of the study imposes inherent limitations. Measurement of systolic function in the early phase following MI may underestimate LVEF due to myocardial stunning. Only standard echocardiographic parameters were available in this clinical data-set. Pre-existing echocardiographic abnormalities prior to the index MI represent an important confounder. Inclusion of more comprehensive laboratory data, which were not available for this study, including haemoglobin, peak Troponin levels, electrolytes, baseline glucose levels, and serum lipids, could potentially enhance risk prediction models. The cause of death was adjudicated retrospectively on the basis of the medical certificate of cause of death and available case notes. Whilst the accuracy of medical certificates of causes of death is recognized as imperfect, the auditors had access to postmortem findings (where available) and medical records to adjudicate where necessary, and thus the final adjudicated cause of death was based on the best available data. Moreover, a recent study documented that cause of death certification prior to post mortem was accurate in 83% of causes.³¹ Furthermore, key clinical variables such as Killip class, inpatient medications, and biomarkers like NT-proBNP were not available in this retrospective dataset. The absence of these powerful predictors is a significant limitation, as they may have provided prognostic information not captured by the included variables. While our study utilized standard echocardiographic parameters available in historical datasets, we acknowledge the growing evidence supporting the superior prognostic value of myocardial strain imaging. Recent studies have demonstrated that both global longitudinal strain and left atrial strain provide incremental prognostic information beyond LVEF and standard diastolic indices following acute MI.^32,33 Although strain data were not uniformly available for this retrospective cohort, future ML models incorporating these sensitive markers of subclinical dysfunction would likely achieve even greater predictive accuracy and should be a priority for prospective data collection. Finally, we acknowledge that biological risk is only one component of outcomes. Further enhancements to the prognostic utility of ML predictive models can be achieved by inclusion of genomic data, as well as psychosocial, cultural, and socio-economic indices which are all important determinants of mortality in ACS.

Conclusions

The main finding of this study is that a ML model incorporating comprehensive echocardiographic data using Gradient Boosted Cox was superior to Cox regression in predicting CV death after MI, and the inclusion of comprehensive echocardiographic data significantly enhanced ML-enabled prediction of survival. In the final ML models, the majority (75%) of the top 20 performing predictors were echocardiographic measurements. Furthermore, the addition of comprehensive echocardiographic data over and above clinical data and LVEF resulted in significant improvement of model performance. Together, our findings suggest that ML models incorporating comprehensive echocardiographic and clinical data have the potential to improve prediction of CV death following MI.

Supplementary material

Supplementary material is available at European Heart Journal – Digital Health.

Author contributions

Liam Scanlon (Conceptualization, Data curation, Formal analysis, Writing—review & editing [equal]), Eddy Xiong (Data curation, Investigation, Writing—review & editing [equal]), Nicole Chan (Data curation, Investigation, Writing—review & editing [equal]), Michael B Mallouhi (Data curation, Investigation, Methodology [equal]), William Vollbon (Data curation, Investigation, Methodology [equal]), John J. Atherton (Conceptualization, Writing—review & editing [equal]), Andrew K Lin (Conceptualization, Formal analysis, Writing—review & editing [equal]), and Sandhir B Prasad (Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration [lead], Writing—original draft [equal])

Funding

1. Heart Foundation of Australia – Vanguard Grant 2. Metro North Hospital and Health Service – Clinician Research Fellowship (SP).

Data availability

Availability of data to external parties is restricted by institutional HREC and research governance policies.

Lead author biography

A/Prof Sandhir B Prasad is a senior staff cardiologist and clinical lead in echocardiography at the Royal Brisbane and Women’s Hospital. He has been awarded research fellowships from the National Heart Foundation (NZ), Centres of Health Research (Qld) and Metro North (Clinician Research Fellowship). He is the recipient of Awards of Excellence from ASUM (Australasian Sonologist of the Year) and Metro North Hospital and Health Service (Clinical Research Award). His PhD focussed on the pathophysiology of left ventricular diastolic dysfunction. He is an Associate Professor at the University of Queensland and Griffith University.