This study is part of the TOTAL-AMI (Tailoring Of Treatment in All comers with Acute Myocardial Infarction) project, previously described by Eggers et al. [9].

The original population of 1385 patients was a subset from the larger TOTAL-AMI cohort with subjects hospitalized due to ACS between March 2008 and September 2014 at the departments of cardiology in Uppsala (Uppsala University Hospital, site 1), Lund (Skåne University Hospital, site 2), and Stockholm (Danderyd Hospital, site 3). This subset had been randomly singled out for multimarker panel sampling at admission and was deemed an appropriate population for this study with testing of the echocardiographic measurements given its sample size.

The subjects received guideline-directed therapy and underwent transthoracic echocardiography according to clinical routine. Medical history and patient characteristics upon presentation were retrieved from the SWEDEHEART registry. The outcome measures, time to all-cause death and time to HF re-hospitalization, were collected from the Swedish Patient Registry (PAR) in July 2018. In PAR, discharge diagnoses were based on International Classification of Diseases, 10th revision, Clinical Modification (ICD-10-CM) codes.

After exclusion of patients with missing or unsatisfactory echocardiographic images, 941 patients remained. Data on previous myocardial infarction and smoking status in the SWEDEHEART associated registry RIKS-HIA was missing in 124 patients. Missing variables were imputed by multiple imputation in the regression analyses. The inclusion process is described in Fig. 1.

The study was approved by the Regional Committee for Medical Research Ethics (DNR 2017/759–31).

Transthoracic 2D echocardiography was performed as part of clinical routine within 72 h from admission. For the purpose of this study, the echocardiographic raw data was collected from the imaging databases at the participating hospitals and re-analyzed at the core lab in Uppsala. Echocardiographic analyses were performed using TomTec-Arena version 2.30 (TomTec, Unterschleißheim, Germany). Volumetric measurements, including LVEF, were obtained by manual delineation of the endocardium and calculated according to the modified Simpson’s method. Strain measurements based on speckle tracking were obtained by software-automated delineation of the endocardium with manual corrections performed when deemed necessary. The same experienced reviewer performed all measurements.

Peak GLS was assessed in monoplane from the four-chamber-, two-chamber, and three-chamber view, respectively. Hence, triplane GLS was calculated as an average of the three views. In subjects that lacked a feasible three-chamber view, GLS was calculated as an average from the four- and two-chamber views (i.e., biplane). Patients recruited from the site in Lund were often examined according to a truncated protocol focused on biplane LVEF assessments and therefore did not always have an available apical three-chamber view (biplane GLS: n = 256).

In accordance with recommendations from the American Society of Echocardiography and the European Association of Cardiovascular Imaging, images with suboptimal tracking of the endocardium in more than two segments in one single view were excluded [10]. This principle was applied both in GLS and volumetric tracings. Thirteen patients (1.4%) were in atrial fibrillation during examination. Their heart rate was below 90 beats per minute and care was taken to measure GLS at somewhat regular RR intervals.

The predictive value of GLS and LVEF was investigated with a univariable receiver operating characteristic (ROC) analysis against the combined outcome. Their optimal cut-off values were obtained according the Youden index. Spearman’s rho was assessed to test correlation between the two echocardiographic parameters and cubic spline analyses were performed to test for a non-linear predictive relation between the echo measurements and the combined endpoint.

Clinical variables with assumed prognostic importance were selected (Table 2) for univariable analysis by Cox proportional hazards regression (congestive heart failure, hypertension, previous MI, chronic kidney disease, diabetes, smoking status). A following multivariable analysis was adjusted for clinical variables with a p value < 0.1 from the univariable analysis. Age and sex were included in the multivariable analysis as pre-specified covariables. Reported results from the multivariable analyses with Cox proportional hazards regression are based on imputed datasets.

Harrell’s C-index was assessed in a step-wise manner with the initial addition of LVEF to clinical data in a first model followed by GLS on top of LVEF and clinical data in a second model to evaluate improvement in model prediction. A sub-analysis of Harrell’s C was performed in the population with ejection fraction above 40% by the same step-wise approach as in the full population. The cut-off, 40%, was selected due to its clinical relevance in prognostication and treatment guidance. Change in C-index between the models within Figs. 4 and 5 was tested with DeLong’s test [11].

Differences between included and excluded patients is reported in Supplemental Table 1. An in-depth multivariable subgroup analysis with stratification for infarction type (STEMI vs NSTEMI), sex, age (cut-off at 65 years), and LVEF (cut-off at 40% and 50%) was also performed and is reported in Supplemental Table 2.

In order to explore the impact of additional baseline clinical parameters, yet with care taken as to avoid model over-fit, another stepwise analysis was performed with baseline clinical parameters of significance in ACS that were not selected for the main analysis (NTproBNP, pathologic Q-wave, bundle branch block, and infarction type). These results are presented in Supplemental Table 3. Sex stratified baseline data is reported in Supplemental Table 4.

Statistical testing was performed in SPSS version 26.0 (IBM Corp. Released 2019. IBM SPSS Statistics for Windows, version 26.0. Armonk, NY: IBM Corp) and in R (4.0.2).

Multiple imputation of missing values was performed using the SAS function PROC MI and Arbitrary Missing Patterns. Twenty imputed datasets were created to obtain smaller standard errors and to ensure that effect estimates were accurate. The results for each imputation were combined using SAS procedure PROC MIANALYZE. Level of significance was set to p < 0.05.