About the CV Risk Post MI model

Most previous machine learning work after myocardial infarction has predicted all-cause mortality. That is a blunt surrogate, because roughly two thirds of deaths after MI are from non-cardiac causes. Predicting cardiovascular death isolates the outcome that is most directly related to the myocardial insult, and therefore the one most likely to respond to targeted cardiovascular intervention.

In this cohort, 139 of 465 deaths were adjudicated as cardiovascular (4.3% of the whole cohort) over a median 4.5 years of follow-up. Heart failure accounted for 56.1% of these, arrhythmic or sudden cardiac death for 24.5%, and ischaemic causes for 19.4%.

The full model considered 59 clinical and echocardiographic variables. Ranking them by model feature importance and keeping the top eight produced a parsimonious model that is statistically indistinguishable from the full one, and that a clinician can complete from a standard echocardiogram report.

Six of the eight are echocardiographic, and they span LV systolic function, LV geometry, diastolic function, atrial size and right ventricular function. That breadth is the substantive finding of the study: no single echocardiographic index carries the prognostic signal on its own.

Of 3,464 consecutive patients presenting with STEMI or NSTEMI between January 2013 and December 2021, 262 were excluded for a limited or delayed echocardiogram, inadequate image quality, in-hospital death or haemodynamic instability, or incomplete follow-up. That left 3,202 patients, of whom 29.2% were female and 28.8% presented with STEMI. Mean LVEF was 52.5%.

Every echocardiogram was performed within 24 hours of admission on GE Vivid E9 or E95 or Philips iE33 or EPIQ machines, with measurements following American Society of Echocardiography recommendations. Cause of death was adjudicated by two study investigators against the Medical Certificate of Cause of Death, cross-checked against medical records and post-mortem data where available, with disagreements resolved by consensus.

The cohort was split by admission date rather than at random: a training cohort of 1,568 patients admitted 2013 to 2017 (103 CV deaths), and a temporal holdout of 1,634 patients admitted from 2018 onwards (36 CV deaths). This is deliberately the harder test. It approximates prospective deployment and forces the model to survive genuine shifts in practice, which a random split would hide. The CV death rate did shift, from 6.57% before 2018 to 2.20% after.

A Gradient Boosted Cox model was fitted, using an ensemble of regression trees with the negative log partial likelihood of the Cox model as the loss function. This captures non-linear relationships and higher-order interactions while retaining the semi-parametric survival framework. Feature scaling and missing value imputation were derived from the training cohort only. Variables with more than 25% missing data were excluded.

All figures below are from the independent temporal validation cohort: patients the model never saw during training, admitted in a later era.

The Gradient Boosted Cox model discriminated significantly better than conventional multivariable Cox regression (0.861 vs 0.813, P = 0.037). A DeepSurv neural network reached 0.847, not significantly different from the boosted model (P = 0.38). The eight-variable parsimonious model retained essentially all of the discrimination of the full model.

The incremental value of echocardiography was tested with nested models. Clinical variables alone gave a C-index of 0.793. Adding LVEF gave 0.792, no improvement at all (P = 0.949). Adding comprehensive echocardiographic data gave 0.861, an increase of 0.070 (P = 0.017). Decision curve analysis at four years showed the boosted model offered the highest net benefit across most clinically relevant threshold probabilities.

The calculator reports predicted probability of cardiovascular death at 1 and 4 years, and classifies the patient as higher or lower risk using the Youden optimal cutpoint on 4-year predicted risk. That cutpoint falls at 4.5% predicted 4-year risk.

• Lower risk: below 4.5% at 4 years

  Specificity 85% in temporal validation. Most patients fall here, consistent with an overall 4.3% cardiovascular death rate.

• Higher risk: 4.5% or above at 4 years

  Sensitivity 81% in temporal validation. Identifies patients in whom more intensive secondary prevention, closer follow-up or heart failure surveillance may be worth considering.

Two cautions on this threshold. It is a statistically optimal cutpoint, not a clinically validated decision boundary, and no trial has shown that acting on it improves outcomes. It was also derived on the same temporal holdout used to report discrimination, and only patients with at least four years of follow-up or an event could be included in that calculation, so the operating characteristics are likely optimistic.

This model has not been externally validated. It was developed at a single Australian tertiary centre, and while the temporal holdout is a stringent internal test, performance in other health systems and demographic compositions is unknown.

Other constraints worth weighing:

• The retrospective design admits selection bias and unmeasured confounding.
• Non-cardiovascular death was treated as a censoring event in the primary analysis rather than as a competing risk, which can inflate absolute risk estimates. A sensitivity analysis excluding non-cardiovascular deaths was reported in the paper.
• Calibration in the paper tended to slightly overestimate risk in the highest predicted quintiles at later time points, reflecting the lower event rate in the validation era.
• The number of events is modest (103 in training, 36 in validation), which limits the precision of both the model and its calibration.
• Two inputs require tissue Doppler imaging, which is not reported in every echocardiogram.
• eGFR was recorded to a ceiling of 90 in the source dataset, so values above 90 cannot be distinguished.
• All eight inputs are required; the tool does not handle missing or low quality data.
• Early post-infarct ventricular function may be underestimated because of myocardial stunning.

What does the CV Risk Post MI score predict?

It estimates an individual patient's probability of cardiovascular death at 1 year and 4 years after an acute myocardial infarction. It is a time-to-event (survival) prediction rather than a fixed-horizon classifier, so both horizons come from the same underlying model.

Which patients was it developed for?

Adults admitted with STEMI or NSTEMI who had a comprehensive transthoracic echocardiogram within 24 hours of admission. Patients with significant haemodynamic instability, in-hospital death, inadequate image quality, or incomplete follow-up were excluded during development.

Which variables does it need?

Eight: mitral E point velocity, left ventricular ejection fraction, age, eGFR, left atrial volume index, LV posterior wall diastolic thickness, right ventricular systolic lateral velocity, and septal e' velocity. Six of the eight are echocardiographic, and two require tissue Doppler.

How accurate is it?

In an independent temporal validation cohort of 1,634 patients admitted from 2018 onwards, the eight-variable model achieved a Harrell C-index of 0.858. This was not significantly different from the full 59-variable model (C-index 0.861), which itself outperformed conventional multivariable Cox regression (0.813, P = 0.037).

Why is echocardiography needed, rather than ejection fraction alone?

Because in this cohort LVEF added nothing on its own. A model using clinical variables alone reached a C-index of 0.793, and adding LVEF to it changed nothing (0.792, P = 0.949). Only when comprehensive echocardiographic data were added did discrimination improve materially, to 0.861 (P = 0.017). Measures of LV size, diastolic function and right heart function carried the incremental signal.

Where do the absolute probabilities come from?

The model outputs a log hazard ratio. That is converted to an absolute probability using the Breslow baseline cumulative hazard function estimated on the training cohort only, so no information from the validation cohort enters the prediction.

Can I use it to make treatment decisions?

No. This is a research tool. It has not been externally validated outside the development centre, and it is not cleared by the FDA, EMA or TGA for standalone clinical decision-making. It is intended to support, not replace, clinical judgement.

Machine learning to predict long-term cardiovascular death following myocardial infarction: incremental value of echocardiographic data. Scanlon L, Xiong E, Chan NI, Mallouhi M, Vollbon W, Atherton JJ, Lin A, Prasad SB. European Heart Journal - Digital Health 2026;7:ztag048. doi:10.1093/ehjdh/ztag048. Published open access under CC BY-NC 4.0.

Affiliations: Monash Victorian Heart Institute and Monash Health Heart, Victorian Heart Hospital, Monash University; Department of Cardiology, Royal Brisbane and Women's Hospital, Herston, Queensland; Queensland Statewide Cardiac Network (Queensland Cardiac Outcomes Registry); Faculty of Medicine, University of Queensland; and School of Medicine and Dentistry, Griffith University.

The study was approved by the institutional Human Research Ethics Committee (Metro North Hospital and Health Service, HREC/2022/QPCH/86459), with separate Public Health Act mandated permission to access institutional databases. Patient-level data cannot be shared publicly for privacy reasons, but may be available on reasonable request to the corresponding author subject to ethics committee approval.