Abstract
Background
Risk stratification in suspected coronary artery disease (CAD) relies on clinical history and imaging, but longitudinal medical history may provide additional prognostic information.
Objectives
This study evaluated whether machine learning models using 10-year historical registry data improves prediction of major adverse cardiovascular events (MACE) beyond demographics and stress perfusion cardiovascular magnetic resonance (CMR) in patients with suspected CAD.
Methods
A total of 2,159 patients (65 ± 12 years, 44% women) undergoing stress perfusion CMR were included. Models used demographics, CMR findings, and registry diagnoses and procedures from the prior 10 years. MACE was defined as cardiovascular death, myocardial infarction, or coronary intervention. Random Forest models were used for binary prediction, and Random Survival Forests (RSF) and penalized Ridge Cox regression for time-to-event analysis.
Results
During 2.7 ± 0.9 years of follow-up, 212 patients (9.8%) experienced MACE. For binary prediction, the area under the receiver-operating characteristic curve (AUROC) was 0.66 (95% CI: 0.62-0.69) using demographics and 0.72 (95% CI: 0.68-0.75; P = 0.001) after adding historical registry data. The full model achieved an AUROC of 0.82 (95% CI: 0.79-0.85) and outperformed the demographics-plus-CMR model (AUROC: 0.81; 95% CI: 0.78-0.84; ΔAUROC: 0.014; 95% CI: 0.002-0.027; P = 0.029). In time-to-event analysis, RSF and Cox regression performed similarly using demographics and CMR (concordance index [C-index] 0.80 for both; P = 0.546). After including registry data RSF achieved a C-index of 0.81 (P = 0.144), and Cox performance decreased to C-index 0.77 (P < 0.001).
Conclusions
In patients with suspected CAD, machine learning models using 10-year historical registry data have the potential to improve prognostic risk prediction beyond demographics and stress perfusion CMR.
Key words: cardiovascular magnetic resonance, major adverse cardiovascular events, registry data, risk prediction
Central Illustration

Coronary artery disease (CAD) is the leading cause of morbidity and mortality globally, despite major advances in medical and interventional treatments.1 The clinical presentation and prognosis of CAD vary considerably among individuals, making patient management and risk stratification challenging.2 Over the last decades, cardiovascular magnetic resonance (CMR) with stress perfusion has established itself as a reliable, noninvasive test for CAD. Stress perfusion CMR provides detailed information on myocardial ischemia, infarction, and cardiac function, all of which are strong predictors of major adverse cardiovascular events (MACE).3, 4, 5, 6 Current guidelines for management of patients with suspected CAD recommend calculation of a pretest probability for CAD before deciding on diagnostic strategy.2 Pretest probability is primarily based on symptoms and demographics. Stress perfusion CMR is recommended for those with intermediate to high pretest probability. To what extent data on the patient’s medical history can improve risk prediction is not completely known.
Swedish national health registries provide longitudinal data with high validity and excellent nationwide coverage.7 These registries contain data on diagnoses and procedures spanning decades, capturing the cumulative burden of disease over time. The volume and complexity of these historical parameters often present a significant analytical challenge. Many statistical methods may struggle with this complex data, as they typically assume that risk factors affect outcomes in a linear, additive way.8 Machine learning (ML) algorithms may be suitable for handling high-dimensional and nonlinear data.9 Random Survival Forests (RSF) extends this approach to time-to-event data10 that may detect complex patterns hidden within years of medical records. As longitudinal medical history captures the cumulative burden of disease and prior health care utilization over time, incorporating historical registry data into risk stratification may provide additional prognostic information.
Therefore, the aim of this study was to evaluate whether ML-methods applied to 10-year historical registry data adds prognostic value to demographics and CMR findings in patients with suspected CAD. It was further hypothesized that an RSF model would be well suited to incorporating high-dimensional historical registry information in time-to-event analysis.
Methods
Data availability statement
Individual patient data will not be shared due to patient integrity. Deidentified registry data can be shared on reasonable request to the investigator. This exploratory analysis was conducted within a documented study protocol but was not prospectively registered. The analytical code is available at https://github.com/Cardiac-MR-Group-Lund/cmr-cad.
Study population and design
Patients eligible for participation were identified from a hospital database of stress perfusion CMR examinations performed at the Department of Clinical Physiology, Skåne University Hospital, Lund, Sweden. The initial cohort consisted of 2,239 patients with suspected CAD who were referred for stress perfusion CMR between January 1, 2019, and December 31, 2022. Eighty patients from the initial cohort were excluded due to missing perfusion data. This resulted in a final study population of 2,159 patients. Baseline demographic data, collected at the time of the stress perfusion CMR examination, included age, sex, and body mass index. Time zero was defined as the date of the index stress perfusion CMR examination, and patients were followed through linkage to the Swedish National Patient Register and Cause of Death Register (maintained by the National Board of Health and Welfare) until an endpoint or administrative censoring on December 31, 2023. The study was approved by the Swedish Ethical Committee (Dnr 2023-02246-01), which waived the requirement for written informed consent due to the registry-based design of the study.
Stress perfusion CMR
All examinations were performed on either of two 1.5 T MR systems (Siemens Magnetom Aera or Siemens Magnetom Sola, Siemens Healthineers). All patients were examined using a standardized clinical stress perfusion protocol including steady-state free precession cine images for myocardial function, single-bolus dual-sequence perfusion mapping for quantification of myocardial perfusion11 and late gadolinium enhancement (LGE) imaging for detection of myocardial infarction. The clinical image analysis was performed using Segment CMR (Medviso, versions 2.2-4.1).12
All stress perfusion CMR results were retrieved from clinical reports written at the time of the examination. This included left ventricular ejection fraction (LVEF), left ventricular end-diastolic volume, and left ventricular end-systolic volume, both volumes indexed to body surface area, and the presence or absence of ischemia or infarction. Stress-induced ischemia was defined as decreased perfusion in noninfarcted myocardium by visual assessment of quantitative perfusion maps. Quantitative perfusion maps were routinely used as visual clinical decision support, whereas absolute myocardial blood flow or myocardial perfusion reserve values were not included in the clinical report. Therefore, numerical values for myocardial blood flow and myocardial perfusion reserve were not extracted or included as model predictors. Myocardial infarction was defined based on the presence of an LGE pattern consistent with CAD.
Historical registry data
Longitudinal clinical history was captured by extracting International Classification of Diseases (ICD)-10 diagnosis codes and Swedish procedure codes from the Swedish National Patient Register. In total, 343 ICD-10 diagnosis codes and 142 procedure codes were selected by 2 experienced clinicians. These represented cardiovascular risk factors, cardiovascular diseases and relevant comorbidities, cardiac investigations, and cardiovascular procedures. This extraction covered a 10-year look-back period from the date of each patient's stress perfusion CMR examination. Within this window, a history of ischemic heart disease was defined by the presence of at least 1 relevant ICD-10 code (I21-I25) or procedure code for coronary intervention, including percutaneous coronary intervention (PCI) or coronary artery bypass grafting.
To transform these raw registry records into a structured feature set for ML, we applied a standardized preprocessing pipeline. First, ICD codes were truncated to their first 3 characters to capture broader disease categories and reduce sparsity. Each unique truncated code was then treated as a binary feature, representing the presence or absence of that specific diagnosis or procedure within the 10-year window. To ensure statistical robustness and reduce noise, only features present in at least 10 patients within the total cohort were included. This process resulted in a binary patient-feature matrix with 2,159 patients and 62 registry-derived features, representing each patient’s cumulative comorbidity and procedural burden before the index stress perfusion CMR examination (Supplemental Table 1).
Outcome data were obtained from the Swedish National Cause of Death Register and Swedish Patient Register. The primary endpoint of the study was MACE defined as a composite of cardiovascular death, myocardial infarction, or coronary intervention (PCI or coronary artery bypass grafting). A 30-day exclusion window was applied to the time-to-event analysis to ensure that early revascularizations or events directly triggered by the stress perfusion CMR findings did not confound the long-term prognostic assessment.
Prediction modeling and statistics
To quantify the incremental prognostic value of historical registry data, 2 complementary prediction tasks were evaluated: binary prediction of incident MACE and time-to-event prediction of MACE. Event-free survival was estimated using the Kaplan-Meier method and compared between patients with and without stress-induced ischemia using the log-rank test.
To train the binary and time-to-event models, the data were grouped into 3 data domains: 1) demographics, containing age, sex, and body mass index; 2) CMR findings, containing stress-induced ischemia, infarction, LVEF, left ventricular end-diastolic volume, and left ventricular end-systolic volume; and 3) historical registry data, containing diagnoses and procedure codes. From these data, features used in the ML models were extracted. The variables included in each data domain and model configuration are listed in Table 1.
Table 1.
Baseline Characteristics and Clinical Outcomes
| Total Population (N = 2,159) | No Stress-Induced Ischemia (n = 1,645) | Stress-Induced Ischemia (n = 425) | |
|---|---|---|---|
| Demographics | |||
| Age (y) | 65 ± 12 | 64 ± 13 | 70 ± 10 |
| Women, n (%) | 947 (44) | 796 (48) | 117 (28) |
| BMI (kg/m2) | 28 ± 5 | 28 ± 5 | 28 ± 5 |
| CMR metrics | |||
| Stress-induced ischemia, n (%) | 425 (20) | 0 (0) | 425 (100) |
| Myocardial Infarction by LGE, n (%) | 432 (20) | 204 (12) | 198 (47) |
| LVEF (%) | 56 ± 12 | 57 ± 11 | 54 ± 12 |
| LVEDV (mL/m2) | 86 ± 26 | 85 ± 25 | 89 ± 30 |
| LVESV (mL/m2) | 40 ± 25 | 39 ± 23 | 43 ± 28 |
| Historical registry data | |||
| Prior ischemic heart disease, n (%) | 497 (23) | 294 (18) | 172 (41) |
| Hypertension, n (%) | 693 (32) | 460 (28) | 195 (46) |
| Diabetes mellitus, n (%) | 260 (12) | 150 (9) | 94 (22) |
| Follow-up and outcomes | |||
| Mean follow-up (y) | 2.7 ± 0.9 | 2.7 ± 0.9 | 2.7 ± 0.9 |
| Total MACE, n (%) | 212 (9.8) | 67 (4.1) | 129 (30.4) |
| Cardiovascular death | 34 (1.6) | 15 (0.9) | 13 (3.1) |
| Myocardial infarction | 44 (2.0) | 26 (1.6) | 17 (4.0) |
| Coronary intervention (PCI/CABG) | 134 (6.2) | 26 (1.6) | 99 (23.3) |
| MACE after 30-day exclusion window | 196 (9.1) | 65 (4.0) | 116 (27.3) |
| MACE without coronary revascularization | 83 (3.8) | 43 (2.6) | 33 (7.8) |
Values are mean ± SD or n (%). MACE (shown in bold) is defined as a composite of cardiovascular death, myocardial infarction, or coronary intervention.
BMI = body mass index; CMR = cardiovascular magnetic resonance; LGE = late gadolinium enhancement; LVEDV = left ventricular end-diastolic volume; LVEF = left ventricular ejection fraction; LVESV = left ventricular end-systolic volume; MACE = major adverse cardiovascular events; PCI/CABG = percutaneous coronary intervention/coronary artery bypass grafting.
Missing continuous and binary demographic and CMR features were imputed within each training set using the median and most frequent category. Repeating the principal model comparisons using 5 imputations by chained equations within the training folds did not substantially change the interpretation of the results. Missing data for predictors before imputation are presented in Supplemental Table 2.
Model configurations were defined to reflect sequential information available during clinical evaluation. Prognostic performance was therefore assessed using demographics alone, then after addition of CMR findings, and finally after addition of historical registry data. To examine the potential contribution of historical registry data before imaging, models combining demographics and historical registry data without CMR findings were also evaluated. Class weighting was used in the binary Random Forest models to account for the imbalance between patients with and without MACE.
For time-to-event modeling, the analyses focused on clinically relevant features. Hence, models were trained using demographics and CMR findings with the subsequent addition of 10 years of historical registry data. For each feature configuration, a linear penalized Cox proportional hazards model with Ridge regularization was compared with a nonlinear Random Survival Forest (RSF) model (800 trees, log-rank splitting).10 This design allowed assessment of both the incremental value of registry data and the ability of different modeling approaches to capture this information.
Ridge Cox regression was included as a conventional linear benchmark, as it represents a widely used conventional approach for time-to-event modeling. Ridge regularization was applied to reduce overfitting and improve coefficient stability in the presence of correlated and high-dimensional predictors, while retaining all variables in the model. The regularization parameter was optimized within the cross-validation framework. Because the performance of penalized models may depend on the regularization and tuning approach, supplementary analyses were performed using Least Absolute Shrinkage and Selection Operator (LASSO)13 and elastic-net Cox regression.14 These models used the same predictors, outer folds, preprocessing, and nested tuning procedure as the primary Ridge Cox analysis. The proportional-hazards assumption was assessed for the demographics-plus-CMR Cox model using Grambsch-Therneau tests15 and visual inspection of scaled Schoenfeld residuals16 (Supplemental Table 3, Supplemental Figure 1).
Binary and time-to-event models were evaluated using nested 5-fold cross-validation. In each outer iteration, 4 folds were used for training and 1 as the held-out test set. Imputation and hyperparameter tuning were performed within the training data only, and performance was estimated from pooled out-of-fold predictions. Registry predictors were defined and frozen before modeling without reference to the outcome. Performance was compared using area under the receiver-operating characteristic curve (AUROC). Probability accuracy was assessed using the Brier score, and calibration using calibration plots, intercepts, and slopes based on out-of-fold predictions. Model discrimination for time-to-event analyses was evaluated using Harrell concordance index (C-index), with 2-sided paired bootstrap sampling (10,000 replicates) used to determine 95% CIs and assess differences between models. Differences in AUROC for binary prediction models were assessed using DeLong test for correlated receiver-operating characteristic curves. A 2-sided P value <0.05 was considered statistically significant. All analyses utilized Python (v3.12) with the scikit-learn and scikit-survival libraries.
The binary analysis classified MACE occurring after the index CMR and before administrative censoring. The earliest event occurred 2 days and the latest 3.80 years after CMR. Follow-up among patients without an event ranged from 1.00 to 4.47 years.
Patient and public involvement
Patients and members of the public were not involved in the design, conduct, reporting, or interpretation of this study.
Results
Baseline characteristics
Baseline characteristics for the study population are presented in Table 1. The final study population consisted of 2,159 patients with a mean age of 65 ± 12 years (44% women). The prevalence of traditional risk factors included hypertension in 32%, diabetes in 12%, and a history of prior ischemic heart disease in 23%. Stress-induced ischemia was detected in 425 (20%) patients; ischemia status was unavailable in 89 patients as the assessment was missing in the report data. Myocardial infarction was present in 432 (20%) patients.
Clinical outcomes
During a mean follow-up of 2.7 ± 0.9 years, a total of 212 MACE events (9.8%) occurred in the full cohort. Registry outcome follow-up was available for all 2,159 patients. Clinical characteristics and outcomes stratified by stress-induced ischemia are presented in Table 1. For the binary risk prediction analysis, all 212 events observed during follow-up were utilized. The time-to-event analysis was based on all patients with 196 events (9.1%) after censoring events during the first 30 days after the stress perfusion CMR examination. Event-free survival differed between patients with and without stress-induced ischemia (P < 0.001) (Supplemental Figure 2).
Binary risk prediction during available follow-up
The AUROCs for binary risk prediction analyses are shown in Figure 1. The models based on demographics alone (yellow line) and historical registry data alone (dashed dark gray line) both achieved an AUROC of 0.66. By combining data from these 2 models (green line), performance improved (AUROC: 0.72, ΔAUROC: 0.059; 95% CI: 0.024-0.095; P = 0.001). By including CMR findings to the model, AUROC increased to 0.82 (ΔAUROC: 0.106; 95% CI: 0.078-0.136; P < 0.001). This model, containing demographics, historical registry data and CMR findings (blue line), outperformed a model including the data available in the clinical situation, containing demographics and CMR findings (red line), showing an AUROC of 0.81 (ΔAUROC: 0.014; 95% CI: 0.002-0.027; P = 0.029). The improved risk prediction persisted after excluding coronary interventions from the composite endpoint and focusing on cardiovascular death and acute myocardial infarction alone (ΔAUROC: 0.037; 95% CI: 0.006-0.070) (Supplemental Table 4). Including registry data improved discrimination but increased the Brier score from 0.128 to 0.139, and the raw predicted probabilities showed imperfect calibration (Supplemental Table 5, Supplemental Figure 3). In a supplementary sex-stratified analysis, the full binary model showed higher discrimination in women than in men (AUROC: 0.85; 95% CI: 0.80-0.89 vs 0.79; 95% CI: 0.75-0.82; P = 0.033). In the full model, containing demographics, historical registry data, and CMR findings, held-out permutation-importance analysis identified stress-induced ischemia as the strongest individual predictor (Figure 2). Infarction on LGE and age were the next most important predictors, followed by LVEF and prior PCI with stent, although their individual contributions were substantially smaller.
Figure 1.

Binary Prediction
Area under the receiver-operating characteristic curve (AUROC) for binary prediction of major adverse cardiovascular events. Note the significant increase in AUROC when adding historical registry data to demographics alone (yellow → green; P = 0.001), and when added to the cardiovascular magnetic resonance findings and demographics model (red → blue; P = 0.029). CMR = cardiovascular magnetic resonance.
Figure 2.

Permutation Importance Analysis for the Comprehensive Random Forest Model
Stress-induced ischemia was identified as the strongest predictor for major adverse cardiovascular events (MACE), followed by clinical and imaging features. AUROC = area under the receiver-operating characteristic curve; BMI = body mass index; LGE = late gadolinium enhancement; LVEF = left ventricular ejection fraction; LVESV = left ventricular end-systolic volume; PCI = percutaneous coronary intervention.
Time-to-event modeling analysis
When trained using demographics and CMR findings, Ridge Cox regression and RSF showed comparable performance, with C-indices of 0.80 (95% CI: 0.77-0.83) and 0.80 (95% CI: 0.77-0.83), respectively (P = 0.546) (Figure 3). When the 10-year historical registry data were added, RSF achieved the highest C-index of 0.81 (95% CI: 0.78-0.83), although the increase from 0.80 was not statistically significant (ΔC-index 0.008; 95% CI: −0.002 to 0.020; P = 0.144). With the full model, RSF outperformed the primary Ridge Cox model (ΔC-index 0.037; 95% CI: 0.019-0.058; P < 0.001). In supplementary analyses, LASSO and elastic-net Cox remained stable after addition of registry data. With the full model, RSF outperformed LASSO Cox (ΔC-index, 0.015; 95% CI: 0.003-0.029; P = 0.024), but not elastic-net Cox (ΔC-index, 0.010; 95% CI: −0.003 to 0.024; P = 0.129) (Supplemental Table 6). In the analysis of the composite endpoint without coronary revascularization, including registry data produced little change in RSF performance, whereas RSF outperformed elastic-net Cox with the full model (Supplemental Table 7).
Figure 3.

Time-to-Event Model Performance Comparison
C-index and 95% CIs are shown in brackets, for Ridge Cox regression (Cox) and Random Survival Forests (RSF) models using 2 model configurations. Left: models containing demographics and cardiovascular magnetic resonance (CMR) findings. Right: after addition of registry data, RSF achieved the highest C-index, although the within-model increase was not statistically significant. With the full model, RSF outperformed Ridge Cox regression. ⁎P < 0.05, P < 0.01, ⁎⁎⁎P < 0.001. C-index = concordance index; ns = not significant; other abbreviation as in Figure 1.
Discussion
ML models utilizing 10-year historical registry data provided modest incremental discrimination beyond demographics and stress perfusion CMR alone in a cohort of 2,159 patients with suspected CAD (Central Illustration). The incremental improvement in binary discrimination persisted for the composite endpoint without coronary revascularization, defined as cardiovascular death or acute myocardial infarction. In the primary time-to-event analysis, RSF achieved the highest C-index with the full model and outperformed Ridge Cox regression. Ridge Cox regression was included as a conventional linear benchmark for time-to-event modeling. The supplementary analyses showed that RSF outperformed LASSO Cox but not elastic-net Cox, indicating that comparative performance depended on the regularization approach. The binary analysis provides proof of concept for the incremental value of registry data, whereas the survival analysis provides the more appropriate longitudinal assessment by accounting for event timing and censoring.
Central Illustration.

Historical Registry Data May Improve Risk Prediction
This study evaluated whether 10-year historical registry data improve prediction of major adverse cardiovascular events in patients with suspected coronary artery disease undergoing stress perfusion cardiovascular magnetic resonance. Registry data improved binary risk prediction beyond demographics combined with cardiovascular magnetic resonance findings. AUROC = area under the receiver-operating characteristic curve; CMR = cardiovascular magnetic resonance; MACE = major adverse cardiovascular events.
Binary risk prediction
The binary risk prediction analyses in the present study showed that the inclusion of historical registry data increased the pretest risk assessment compared to demographics alone. Furthermore, the addition of historical registry data provided incremental prognostic value in relation to data available in the clinical setting, that is, demographics and CMR findings. This improvement in discrimination persisted in the composite endpoint without coronary revascularization. Together, these results support the use of historical registry data both in early risk assessment and for postimaging risk prediction. Furthermore, the use of historical registry data could potentially enable improved phenotype matching between different treatment arms in prospective clinical trials.
In this model, stress-induced ischemia was the most influential predictor of MACE, which is consistent with previous studies showing that stress-induced ischemia by stress perfusion CMR is a strong predictor of future events. A systematic review and meta-analysis including more than 74,000 patients showed that stress-inducible ischemia and positive LGE were associated with higher mortality and cardiovascular event rates, whereas a normal stress perfusion CMR was associated with a lower likelihood of cardiovascular events for at least 3.5 years.6 Similar findings have been reported from the multicenter SPINS registry.4 Earlier work has also shown that reversible perfusion defects and infarct-related markers provide complementary prognostic information on stress perfusion CMR.17 However, these studies primarily focused on imaging findings and conventional clinical variables and did not evaluate high-dimensional historical registry data over several years. This study therefore extends this literature by examining whether historical registry data add prognostic information beyond stress perfusion CMR, and whether the ability to use such information depends on different modeling approaches.
Time-to-event survival analysis
In the primary time-to-event analysis, RSF achieved the highest C-index with the full model and outperformed Ridge Cox regression. Including historical registry data resulted in a small, nonsignificant increase in RSF performance (P = 0.144), whereas Ridge Cox performance declined. In supplementary analyses, LASSO and elastic-net Cox remained stable after addition of registry data, and elastic-net Cox achieved performance comparable to RSF. This is consistent with previous evidence that machine-learning methods do not necessarily outperform regression-based prediction models when model development and validation procedures are appropriately implemented.18 RSF remains well suited to right-censored data and can accommodate nonlinear effects and interactions without prespecification,10 whereas elastic-net regularization may also effectively incorporate sparse and correlated registry features. Historical registry data may complement CMR by reflecting comorbidity and procedural history not captured by the index examination.
Study limitations
The single-center design may limit generalizability, and external validation in other regions and health care systems would improve the generalizability of the findings. The number of events was limited relative to the number of features, particularly in the time-to-event analyses. Although nested cross-validation reduced the risk of overfitting, it cannot eliminate this risk, and the findings require external validation. The primary endpoint included coronary revascularization, which may have been influenced by ischemia on the index CMR examination. However, a sensitivity analysis excluding revascularization showed consistent results.
Registry data provide broad longitudinal coverage, but the validity of individual diagnostic codes is not uniform. Diagnostic accuracy could vary across diseases, and coding may differ between providers and vary over time. Furthermore, diagnoses from primary care facilities were not included; hence, some comorbidities may be under-represented. The study period included the COVID-19 pandemic, during which changes in health care use and registry coding may have affected the recording of predictors and clinical events. In the present study, model discrimination was primarily evaluated, which is appropriate for comparing performance between modeling approaches. Although registry data modestly improved discrimination, the higher Brier score indicates that the predicted probabilities did not translate into improved overall probabilistic performance. The registry-enhanced model may better rank patients according to risk but should not be interpreted as providing improved absolute risk estimates without external validation and recalibration.19
Conclusions
In patients with suspected CAD, ML models using 10-year historical registry data have the potential to improve risk prediction beyond demographics and perfusion CMR findings.
Perspectives.
COMPETENCY IN MEDICAL KNOWLEDGE: In patients with suspected CAD undergoing stress perfusion CMR, ML models, incorporating 10-year historical registry data, provided modest incremental prognostic information beyond demographics and imaging findings alone. By utilizing registry data on comorbidity and clinical procedures, these models may enhance risk stratification both before and after imaging.
TRANSLATIONAL OUTLOOK: External validation in other cohorts and health care systems is required before clinical implementation. Future studies should assess calibration, interpretability, and clinical utility, and determine whether integration of ML-based risk estimates could facilitate clinical decision-making or improve patient outcomes.
Funding support and author disclosures
Dr Engblom has received project grants from the Swedish Heart-Lung Foundation (HLF20240659), Swedish Research Council (VR2025-02865), Region of Scania and the Swedish Governmental Funding for Clinical Research (ALF). All other authors have reported that they have no relationships relevant to the contents of this paper to disclose.
Acknowledgments
The authors would like to thank the grant providers who made this project possible: Swedish Heart-Lung Foundation (HLF20240659), Swedish Research Council (VR2025-02865), and the Region of Scania and the Swedish Governmental Funding for Clinical Research (ALF) project grant.
Footnotes
The authors attest they are in compliance with human studies committees and animal welfare regulations of the authors’ institutions and Food and Drug Administration guidelines, including patient consent where appropriate. For more information, visit the Author Center.
Appendix
For supplemental tables and figures, please see the online version of this paper.
Supplementary material
References
- 1.Mensah G.A., Fuster V., Murray C.J.L., Roth G.A. Global burden of cardiovascular diseases and risks, 1990-2022. J Am Coll Cardiol. 2023;82:2350–2473. doi: 10.1016/j.jacc.2023.11.007. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Vrints C., Andreotti F., Koskinas K.C., et al. 2024 ESC guidelines for the management of chronic coronary syndromes: developed by the task force for the management of chronic coronary syndromes of the European Society of Cardiology (ESC) endorsed by the European Association for Cardio-Thoracic Surgery (EACTS) Eur H J. 2024;45:3415–3537. doi: 10.1093/eurheartj/ehae177. [DOI] [PubMed] [Google Scholar]
- 3.Engblom H., Xue H., Akil S., et al. Fully quantitative cardiovascular magnetic resonance myocardial perfusion ready for clinical use: a comparison between cardiovascular magnetic resonance imaging and positron emission tomography. J Cardiovasc Magn Reson. 2017;19:78. doi: 10.1186/s12968-017-0388-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Kwong R.Y., Ge Y., Steel K., et al. Cardiac magnetic resonance stress perfusion imaging for evaluation of patients with chest pain. J Am Coll Cardiol. 2019;74:1741–1755. doi: 10.1016/j.jacc.2019.07.074. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Nagel E., Greenwood J.P., McCann G.P., et al. Magnetic resonance perfusion or fractional flow reserve in coronary disease. N Engl J Med. 2019;380:2418–2428. doi: 10.1056/NEJMoa1716734. [DOI] [PubMed] [Google Scholar]
- 6.Ricci F., Khanji M.Y., Bisaccia G., et al. Diagnostic and prognostic value of stress cardiovascular magnetic resonance imaging in patients with known or suspected coronary artery disease: a systematic review and meta-analysis. JAMA Cardiol. 2023;8:662–673. doi: 10.1001/jamacardio.2023.1290. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Everhov Å.H., Frisell T., Osooli M., et al. Diagnostic accuracy in the Swedish national patient register: a review including diagnoses in the outpatient register. Eur J Epidemiol. 2025;40:359–369. doi: 10.1007/s10654-025-01221-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Deo R.C. Machine learning in medicine. Circulation. 2015;132:1920–1930. doi: 10.1161/CIRCULATIONAHA.115.001593. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Liu T., Krentz A., Lu L., Curcin V. Machine learning based prediction models for cardiovascular disease risk using electronic health records data: systematic review and meta-analysis. Eur Heart J Digit Health. 2025;6:7–22. doi: 10.1093/ehjdh/ztae080. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Ishwaran H., Kogalur U.B., Blackstone E.H., Lauer M.S. Random survival forests. Ann Appl Stat. 2008;2:841–860. [Google Scholar]
- 11.Kellman P., Hansen M.S., Nielles-Vallespin S., et al. Myocardial perfusion cardiovascular magnetic resonance: optimized dual sequence and reconstruction for quantification. J Cardiovasc Magn Reson. 2017;19:43. doi: 10.1186/s12968-017-0355-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Heiberg E., Sjögren J., Ugander M., Carlsson M., Engblom H., Arheden H. Design and validation of Segment--freely available software for cardiovascular image analysis. BMC Med Imaging. 2010;10:1. doi: 10.1186/1471-2342-10-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Tibshirani R. The lasso method for variable selection in the cox model. Stat Med. 1997;16:385–395. doi: 10.1002/(sici)1097-0258(19970228)16:4<385::aid-sim380>3.0.co;2-3. [DOI] [PubMed] [Google Scholar]
- 14.Simon N., Friedman J., Hastie T., Tibshirani R. Regularization paths for cox's proportional hazards model via coordinate descent. J Stat Softw. 2011;39:1–13. doi: 10.18637/jss.v039.i05. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Grambsch P.M., Therneau T.M. Proportional hazards tests and diagnostics based on weighted residuals. Biometrika. 1994;81:515–526. [Google Scholar]
- 16.Schoenfeld D. Partial residuals for the proportional hazards regression model. Biometrika. 1982;69:239–241. [Google Scholar]
- 17.Steel K., Broderick R., Gandla V., et al. Complementary prognostic values of stress myocardial perfusion and late gadolinium enhancement imaging by cardiac magnetic resonance in patients with known or suspected coronary artery disease. Circulation. 2009;120:1390–1400. doi: 10.1161/CIRCULATIONAHA.108.812503. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Christodoulou E., Ma J., Collins G.S., Steyerberg E.W., Verbakel J.Y., Van Calster B. A systematic review shows no performance benefit of machine learning over logistic regression for clinical prediction models. J Clin Epidemiol. 2019;110:12–22. doi: 10.1016/j.jclinepi.2019.02.004. [DOI] [PubMed] [Google Scholar]
- 19.Van Calster B., McLernon D.J., van Smeden M., et al. Calibration: the Achilles heel of predictive analytics. BMC Med. 2019;17:230. doi: 10.1186/s12916-019-1466-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
Individual patient data will not be shared due to patient integrity. Deidentified registry data can be shared on reasonable request to the investigator. This exploratory analysis was conducted within a documented study protocol but was not prospectively registered. The analytical code is available at https://github.com/Cardiac-MR-Group-Lund/cmr-cad.
