Electro-mechanical risk score for coronary artery disease risk estimation using seismocardiography Authors: Parastoo Dehkordi, Kouhyar Tavakolian, Thomas Rocco, Khurshid Fozilov, Alisher Yuldashov, Wojciech Zareba, Baxtiyorjon Yuldashov, Baxtiyor Atamuratov, Hamid R Marateb, Shukhrat Doniyorov, Bernara Kurbanova, Farzad Khosrow-Khavar Published: 2026 ABSTRACT AbstractAimsCoronary artery disease (CAD) remains a leading cause of morbidity. Existing clinical likelihood models often lack specificity, contributing to unnecessary diagnostic testing. To develop, train, and validate the electro-mechanical risk (EMR) Score. This machine learning model uses cardiac mechanical information from resting seismocardiography (SCG) recordings and patient-level clinical risk factors to estimate obstructive CAD likelihood.Methods and resultsThis multi-centre clinical study included 2110 adults. Resting SCG was recorded using a sternum accelerometer. Obstructive CAD was defined as 50% or greater stenosis on coronary computed tomography angiography or invasive coronary angiography. A one-dimensional convolutional neural network was trained to compute the EMR Score. Performance was evaluated using repeated cross-validation and external-centre validation and compared with the 2024 ESC Risk-Factor–weighted Clinical Likelihood (RF-CL) model. Among 2110 participants (mean [SD] age, 57.8 [10.3] years; 801 women [38%]), 760 had obstructive CAD. In symptomatic individuals, the EMR Score achieved an AUC of 0.88, outperforming RF-CL (AUC, 0.85; P = 0.023), with higher specificity (53% vs. 35%) and comparable sensitivity (94% vs. 97%). In external-centre validation, the EMR Score achieved an AUC of 0.91, sensitivity of 97.8%, specificity of 52.0%, PPV of 78.0%, and NPV of 93.0%. In asymptomatic participants, the AUC was 0.89. The EMR Score classified 23% of the cohort as very low likelihood, with 2% CAD prevalence.ConclusionThe EMR Score non-invasively estimated obstructive CAD likelihood, with external-centre validation supporting generalizability beyond internal cross-validation.Clinical trial registration ClinicalTrials.gov (NCT06880120, NCT06880133)Keywords: Seismocardiography, Neural networks, Non-invasive diagnostics, Coronary artery disease, Cardiovascular risk factors Introduction Coronary artery disease (CAD) remains a leading global cause of morbidity and mortality, affecting both symptomatic and asymptomatic individuals.1 This persistent burden underscores the need for accessible, non-invasive, and cost-effective diagnostic tools to identify individuals at risk. Current diagnostic pathways often begin with guideline-directed pre-test probability (PTP) estimates, based on age, sex, symptoms, and conventional cardiovascular risk factors such as smoking and hypertension.2–4 While PTP models provide an initial framework for clinical decision-making, their limited specificity contributes to high false-positive rates and unnecessary downstream testing. The 2019 European Society of Cardiology (ESC) guidelines introduced the ESC2019 PTP model, which addressed the overestimation inherent in earlier methods such as the Diamond-Forrester algorithm.2 Despite improved calibration, its clinical utility is hindered by low specificity. In a cohort of 15 411 patients, the ESC2019 model achieved a sensitivity of 97.6% for detecting obstructive CAD but a specificity of only 12.1%, resulting in a substantial number of false-positive classifications.5 To improve discrimination, the 2024 ESC guidelines proposed the risk-factor–weighted clinical likelihood (RF-CL) model, which integrates clinical risk factors alongside demographic and symptom data.3,5 This approach enhances calibration and increases the proportion of individuals reclassified into the very low-likelihood (≤5%) category, a group that may safely forgo further testing. However, the model still leaves diagnostic uncertainty in the low-likelihood range (>5–15%), often requiring adjunctive assessments such as coronary artery calcium (CAC) scoring or stress testing. Addressing these challenges, seismocardiography (SCG) offers a novel, non-invasive approach to assessing cardiac function. SCG records subtle vibrations on the chest wall generated by cardiac motion during the cardiac cycle, providing detailed insights into cardiac mechanical activity. These vibrations, typically captured using a highly sensitive accelerometer, reflect the dorsoventral, head-to-foot, and lateral movements of the heart (see Supplementary material online, Figure S1 in the Supplementary material online, Appendix).6 Its portability and low cost make it especially suitable for diverse clinical environments. SCG has been studied since the 1960s, initially to examine cardiac timing and variability.6,7 Later research linked SCG waveforms to myocardial ischaemia and coronary stenosis, including during balloon angioplasty and exercise stress testing.8–10 Despite early promise, adoption was limited by inadequate signal processing capabilities and lack of integration into diagnostic workflows. Recent advances in sensor design and machine learning have revived interest in SCG. Our prior studies demonstrated that applying machine learning to SCG signals yielded high diagnostic accuracy for detecting CAD, including models trained on rest and exercise recordings.11,12 These findings support SCG’s potential to enable non-invasive, data-driven CAD assessment. To further expand its clinical utility, we developed the electro-mechanical risk (EMR) Score, a predictive model that integrates SCG-derived features with clinical variables.13 The EMR Score is designed to improve initial risk stratification by identifying patients who may safely avoid unnecessary diagnostic testing while still capturing those at high risk for obstructive CAD. Our preliminary study showed that the EMR Score outperformed the ESC2019 PTP model in specificity. In this study, we report the development and evaluation of the EMR Score in a large, prospectively enrolled multi-centre cohort. We refined SCG feature extraction, optimized the model architecture, and evaluated performance in both symptomatic and asymptomatic individuals. Following model locking, the EMR Score’s risk discrimination and stratification performance were compared with the updated RF-CL model. These refinements were designed to enhance specificity and improve PTP estimation for obstructive CAD. Methods Study design and participants This multi-centre diagnostic study included 2600 participants prospectively enrolled between 2019 and 2023 at regional hospitals in Qarshi and Tashkent, Uzbekistan. The study population comprised patients with suspected CAD and asymptomatic healthy volunteers. Ethical approval was obtained from the hospital ethics board (record no. 1-16-02-388-19), and all participants provided written informed consent. The trial protocol is shown in Supplement 1. The study adhered to the Declaration of Helsinki and Good Clinical Practice guidelines. This study followed the transparent reporting of a multivariable prediction model for individual prognosis or diagnosis reporting guideline. Patients with suspected CAD were referred for coronary computed tomography angiography (CCTA) or invasive coronary angiography (ICA). Coronary anatomy was documented using a standardized 15-segment model. Healthy controls underwent a structured screening protocol, including laboratory testing, electrocardiography (ECG), echocardiography, and CAC scoring. For all participants, SCG and ECG signals were recorded simultaneously for 5 min in the supine position using a triaxial accelerometer–based device with a single-lead ECG configuration, placed on the sternum near the xiphoid process (CardioClin, HeartForce Medical Inc, Canada). SCG signals were captured in three orthogonal directions: Dorsoventral, head-to-foot, and lateral (x, y, z). Signals were sampled at 250 Hz with 16-bit resolution. Standardized acquisition protocols were applied across all study sites. Details of inclusion and exclusion criteria are provided in Supplementary material online, Figure S2 in the Supplementary material online, Appendix. EMR score development and validation The EMR Score was developed through a multi-step process involving cardiac-cycle segmentation, representation learning, and model training as illustrated in the Central Illustration. SCG signals were bandpass filtered (0.5–40 Hz), normalized, and transformed into time-frequency representations using a continuous wavelet transform. Cardiac cycles were segmented using Q-wave detection from simultaneously recorded ECG signals. Information from all three SCG axes was used to generate fixed-length cardiac mechanical representations. Additional pre-processing and segmentation details are provided in the Supplement. The EMR Score was trained using cardiac mechanical information from resting SCG recordings together with patient-level clinical variables, including age, sex, chest pain symptoms, smoking, hypertension, hyperlipidaemia, diabetes, and family history of CAD. A one-dimensional convolutional neural network was used to estimate the likelihood of obstructive CAD. Network architecture and training parameters are detailed in Figure 1. Architecture of the EMR model combining SCG features and clinical data. Diagram of the 1D convolutional neural network (CNN) integrating SCG-derived time-frequency features with clinical risk factors to estimate the EMR Score. Abbreviations: EMR, electro-mechanical risk; SCG, seismocardiography; CNN, convolutional neural network. Model development was first assessed using five-fold cross-validation. To address generalizability beyond internal validation, an additional external held-out centre validation was performed. For this analysis, the model was trained using the Tashkent datasets and tested on the geographically distinct Qarshi cohort. The held-out centre was excluded from model training, pre-processing parameter estimation, model selection, and threshold selection. Comparator models The EMR Score was benchmarked against the 2019 ESC PTP model2 and the 2024 risk-factor–weighted clinical likelihood (RF-CL) model. The ESC2019 PTP model uses age, sex, and symptom type, while the RF-CL model incorporates additional cardiovascular risk factors via logistic regression. Model scores were computed using tables published in the original guideline manuscripts.2,3 Assessment of CAD and CAD severity In this study, obstructive CAD refers specifically to 50% or greater epicardial diameter stenosis on either CCTA or ICA. Individuals without such lesions are defined as having non-CAD, meaning no coronary stenosis 50% or greater. To avoid implying the absence of any atherosclerosis, we use the term non-CAD only to denote the absence of obstructive disease. The Gensini Score was used to quantify angiographic CAD burden by assigning weighted values to coronary lesions based on the degree of luminal narrowing and lesion location.13 Statistical analysis Continuous variables were summarized as mean (SD) or median (IQR) and compared using the Mann–Whitney U test. Categorical variables were compared using χ2 or Fisher exact test. Effect sizes were reported using Cramér’s V and rank-biserial correlation, with thresholds of 0.1 (small), 0.3 (moderate), and 0.5 (large). Model discrimination was assessed by the area under the receiver operating characteristic curve (AUC), and model comparisons were tested using DeLong’s test. Calibration was evaluated graphically and with calibration slope and intercept estimates. Net reclassification improvement (NRI) was calculated to assess reclassification accuracy, and decision curve analysis was used to estimate net clinical benefit. A pre-specified subgroup analysis was conducted in asymptomatic individuals using the EMR score only, as guideline-based tools (ESC2019 and RF-CL) are not validated for this population. Metrics included AUC, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), with 95% confidence intervals obtained via bootstrapping. For the external held-out centre validation, model performance was evaluated only once in the held-out cohort after model training and threshold selection had been completed using the training cohorts. The held-out cohort was not used for model fitting, hyperparameter selection, pre-processing parameter estimation, or threshold selection. All statistical analyses were conducted using Python version 3.10.13 (Python Software Foundation). A two-sided P-value <0.05 was considered statistically significant. Results The final dataset included 2110 participants: 1350 without obstructive CAD and 760 with CAD confirmed by ICA or CCTA. Of these, 467 participants were asymptomatic at enrolment, and CAC-CT data were available for 1097 individuals (see Supplementary material online, Figure S2 in the Supplementary material online, Appendix). Baseline characteristics of participants were summarized in Table 1. Table 1Clinical and demographic characteristics by CAD Statusnon-CADCAD P-valueEffective sizeCoronary artery diseaseNumber of patients, No. (%)1350 (64%)760 (36%)NANACharacteristicsSex, No. (%) Male592 (52%)551 (48%)<0.0010.275 Female758 (78%)209 (22%)Age, mean (SD), year55.17(11.09)61.78(44)<0.0010.368Risk factors and symptomsFamily History of early CAD, No. (%) Yes370 (56%)285 (44%)<0.0010.103 No980 (67%)475 (33%)Smoking, No. (%) Yes166 (34%)326 (66%)<0.0010.346 No1184 (73%)434 (27%)Dyslipidaemia, No. (%) Yes669 (66%)347 (34%)0.09090.037 No681 (62%)413 (35%)Hypertension, No. (%) Yes944 (57%)706 (43%)<0.0010.265 No406 (88%)54 (12%)Diabetes, No. (%) Yes109 (45%)135 (55%)<0.0010.144 No1241 (66%)625 (32%)Ejection Fraction, mean (SD), %59.51 ± 5.7953.78 ± 9.48<0.001-0.420Cardiac symptoms at referral Typical chest pain, No. (%)350 (35%)662 (65%)<0.0010.587 Atypical chest pain, No. (%)225 (84%)42 (16%)<0.0010.16 Non-angina chest pain, No. (%)323 (90%)37 (10%)<0.0010.247 Asymptomatic, No. (%)452 (96%)19 (4%)<0.0010.356CACs (n = 1096) 0 (n = 868), No. (%)826 (95%)42 (5%)<0.0010.542 1–10 (n = 41), No. (%)34 (83%)7 (17%)0.01690.052 11–100 (n = 90), No. (%)62 (69%)28 (31%)0.37760.019 101–400 (n = 62), No. (%)27 (44%)35 (56%)0.00110.071 > 400 (n = 35), No. (%)13 (37%)22 (63%)0.00160.069Comparison of clinical variables, cardiac function, symptoms, and CAC categories between participants with and without obstructive CADAbbreviations: CAD, coronary artery disease; CAC, coronary artery calcium. Tests include χ2 for categorical and Mann–Whitney U test for continuous variables. Effect sizes are shown as Cramér’s V or rank-biserial r. Clinical and demographic characteristics by CAD Status Comparison of clinical variables, cardiac function, symptoms, and CAC categories between participants with and without obstructive CAD Abbreviations: CAD, coronary artery disease; CAC, coronary artery calcium. Tests include χ2 for categorical and Mann–Whitney U test for continuous variables. Effect sizes are shown as Cramér’s V or rank-biserial r. Group differences by CAD Status Significant clinical and imaging differences were observed between patients with and without CAD (Table 1). Typical chest pain and CAC = 0 were the most discriminative features (Cramér’s V = 0.59 and 0.54, respectively; both P < 0.001). Moderate associations were found for smoking (V = 0.35), hypertension (V = 0.27), and male sex (V = 0.28). In contrast, dyslipidaemia showed no significant group differences. Age, which was included in the model, and ejection fraction, which was not, also differed significantly between groups, with rank-biserial correlation values of 0.368 and –0.42, respectively, indicating moderate associations in opposite directions—age tended to be higher in CAD patients, while ejection fraction was lower. The EMR Score was significantly higher among CAD patients, with a mean of 55.2 (SD = 25.0) compared with 23.0 (SD = 20.9) in non-CAD participants (P < 0.01), further supporting its ability to differentiate between disease states (see Supplementary material online, Table S1 in the Supplementary material online, Appendix). Model performance Across the entire dataset, the EMR Score achieved an AUC of 0.90 (95% CI: 0.88–0.91) (Table 2). In symptomatic patients, the EMR Score outperformed the RF-CL model (AUC, 0.88 vs. 0.85; P = 0.023) and the ESC2019 PTP model (AUC, 0.83; P < 0.001) (see Supplementary material online, Figure S3A in the Supplementary material online, Appendix). At the pre-specified cut-off of 15%, the EMR Score achieved a sensitivity of 94% and specificity of 53% in symptomatic individuals, compared with 97% sensitivity and 35% specificity for RF-CL. A cut-off of 5% was used for the RF-CL and ESC2019 PTP models. Table 2Discrimination and performance of ESC2019 PTP, RF-CL, and EMR score modelsNo.AUC (95% CI)Sensitivity (95% CI)Specificity (95% CI)PPV (95% CI)NPV (95% CI)ESC2019 PTP (symptomatic patients)16400.83 (0.81–0.85)97% (96–98%)19% (17–22%)50% (49–51%)89% (84–93%)RF-CL (symptomatic patients)16400.85 (0.83–0.86)97% (96–98%)35% (32–39%)41% (55–57%)96% (93–97%)EMR score (symptomatic patients)16400.88 (0.85–0.89)94% (92–95%)53% (50–57%)63% (61–64%)91% (89–93%)EMR score (symptomatic + asymptomatic patients)21100.90 (0.88- 0.91)95% (92–96%)61% (58–63%)58% (57–60%)95% (93–96%)EMR score (asymptomatic patients)4700.89 (0.86–0.92)80% (65–94%)80% (77–84%)15% (12–19%)99% (97–100%)EMR score—external-centre validation (Qarshi cohort)4320.91 (0.88–0.93)98% (96–99%)52% (43–60%)78% (75–81%)95.0% (88–98%)Performance metrics include AUC, sensitivity, specificity, PPV, and NPV. Internal model performance was assessed using cross-validation. External-centre validation was performed by training the EMR Score using the Tashkent datasets and evaluating performance in the held-out Qarshi cohort. Cut-off values were 5% for ESC2019 PTP and RF-CL and 15% for the EMR Score.Abbreviations: AUC, area under the curve; CAD, coronary artery disease; EMR, electro-mechanical risk; ESC2019 PTP, European Society of Cardiology 2019 pre-test probability; RF-CL, risk factor–weighted clinical likelihood; PPV, positive predictive value; NPV, negative predictive value. Discrimination and performance of ESC2019 PTP, RF-CL, and EMR score models Performance metrics include AUC, sensitivity, specificity, PPV, and NPV. Internal model performance was assessed using cross-validation. External-centre validation was performed by training the EMR Score using the Tashkent datasets and evaluating performance in the held-out Qarshi cohort. Cut-off values were 5% for ESC2019 PTP and RF-CL and 15% for the EMR Score. Abbreviations: AUC, area under the curve; CAD, coronary artery disease; EMR, electro-mechanical risk; ESC2019 PTP, European Society of Cardiology 2019 pre-test probability; RF-CL, risk factor–weighted clinical likelihood; PPV, positive predictive value; NPV, negative predictive value. Among asymptomatic participants, for whom RF-CL and ESC2019 PTP are not directly applicable, the EMR Score achieved an AUC of 0.89 (95% CI: 0.86–0.92). In this subgroup, the EMR Score correctly identified 15 of 19 CAD cases, yielding a sensitivity of 80%, specificity of 80%, and NPV of 99% (see Supplementary material online, Table S2 in the Supplementary material online, Appendix and Table S2). To assess generalizability beyond internal cross-validation, an external held-out centre validation analysis was performed. The model was trained using the Tashkent datasets and evaluated in a geographically distinct held-out cohort, which was excluded from model training, pre-processing parameter estimation, model selection, and threshold selection. In this held-out cohort, the EMR Score achieved an AUC of 0.91 (95% CI: 0.88–0.93), sensitivity of 97.8% (95% CI: 95.2–99.2), specificity of 52.0% (95% CI: 43.7–60.1), PPV of 78.0% (95% CI: 75.0–81.0), and NPV of 93.0% (95% CI: 85.0–97.0). Risk stratification by likelihood category Table 3 details the distribution of patients and CAD prevalence across likelihood categories. The RF-CL model assigned most symptomatic patients to the low (35%) and moderate (44%) categories, with no classifications in the high or very high categories—both of which require additional testing under the 2024 ESC guidelines. Similarly, the ESC2019 PTP model concentrated patients in the moderate (55%) and low (27%) groups, with smaller proportions in the very low (12%) and high (6%) categories. Table 3Distribution of risk categories and CAD prevalence across prediction modelsVery low (≤5%)Low (>5–15%)Moderate (>15–50%)High (>50–85%)Very high (>85%)ESC2019 PTP (symptomatic patients)198 (12%)450 (27%)898 (55%)94 (6%)0 (0%)10% CAD17% CAD62% CAD91% CAD0% CADRF-CL (symptomatic patients)337 (20%)579 (35%)724 (44%)0 (0%)0 (0%)4% CAD32% CAD75% CAD0% CAD0% CADEMR Score (symptomatic patients)193 (11%)283 (17%)367 (22%)505 (31%)292 (18%)4% CAD11% CAD27% CAD68% CAD90% CADEMR Score (symptomatic + asymptomatic patients)492 (23%)391 (18%)426 (20%)509 (24%)292 (14%)2% CAD38 CAD26% CAD67% CAD90% CADEMR Score (asymptomatic patients)299 (64%)108 (23%)59 (12%)4 (1%)0 (0%)0% CAD8 CAD17% CAD25% CAD0% CADRisk category assignments and corresponding CAD prevalence for ESC2019 PTP, RF-CL, and EMR models in symptomatic and asymptomatic participants. Values are the number of patients in each likelihood category (percentage of total), followed by CAD prevalence in that group.Abbreviations: CAD, coronary artery disease; EMR, electro-mechanical risk; ESC2019 PTP, 2019 ESC pre-test probability model; RF-CL, risk factor–weighted clinical likelihood. Distribution of risk categories and CAD prevalence across prediction models Risk category assignments and corresponding CAD prevalence for ESC2019 PTP, RF-CL, and EMR models in symptomatic and asymptomatic participants. Values are the number of patients in each likelihood category (percentage of total), followed by CAD prevalence in that group. Abbreviations: CAD, coronary artery disease; EMR, electro-mechanical risk; ESC2019 PTP, 2019 ESC pre-test probability model; RF-CL, risk factor–weighted clinical likelihood. In contrast, the EMR Score demonstrated a more balanced distribution across likelihood categories. It classified 11% of patients into the very low-risk group (EMR score ≤5%), with a CAD prevalence of only 4%, aligning with ESC criteria for deferring further testing. An additional 17% were categorized as low-risk, with an observed CAD prevalence of 11%. These distributions support the EMR Score’s ability to streamline diagnostic pathways and reduce unnecessary testing (Figure 2; Supplementary material online, Figure S5). Notably, none of the patients classified in the very low-risk group had severe CAD (Gensini ≥40), reinforcing the model’s safety in ruling out clinically significant disease (see Supplementary material online, Table S3 in Supplement S2). In contrast, the RF-CL model assigned 3 patients with severe CAD to the very low-risk category (see Supplementary material online, Table S4 in Supplement S2). Distribution of patients and observed CAD prevalence by risk category. Bar plots show the number of patients classified into each predicted risk category by the ESC2019 PTP model, the RF-CL model, and the EMR Score. Risk categories are defined according to ESC guidelines as very low, low, moderate, high, and very high. Dashed lines represent observed CAD prevalence in each group, indicating the alignment (or mismatch) between estimated risk and actual disease burden. Abbreviations: CAD, coronary artery disease; EMR, electro-mechanical risk; ESC, European Society of Cardiology; PTP, pre-test probability; RF-CL, risk factor-weighted clinical likelihood. Reclassification performance Compared with the RF-CL model, the EMR Score improved reclassification performance, with a NRI of +0.098. This improvement was driven by down-classification of non-CAD patients (NRInon-CAD = +0.129), while a modest net loss in reclassification occurred among CAD patients (NRICAD = −0.031) (see Supplementary material online, Figure S3.B in the Supplementary material online, Appendix). Association with CAD severity and coronary calcification The EMR Score showed a strong positive correlation with the Gensini Score (Spearman ρ = 0.677, P < 0.001), indicating its ability to reflect both CAD presence and burden. Median EMR scores increased progressively with higher Gensini categories (see Supplementary material online, Figure S5–S6 in the Supplementary material online, Appendix). Similarly, EMR scores rose across increasing CAC score categories (see Supplementary material online, Figure S7 in the Supplementary material online, Appendix). A moderate correlation was observed between EMR and CAC scores (Spearman ρ = 0.393, P < 0.001). A significant difference in EMR values was observed between those with CAC = 0 and those with higher scores (P = 0.001), supporting its relevance as a non-invasive indicator of subclinical atherosclerosis. Discussion The EMR Score offers a non-invasive, scalable method for estimating the likelihood of obstructive CAD by using cardiac mechanical information from resting SCG recordings and patient-level clinical risk factors. In this large, prospectively enrolled cohort, the model demonstrated strong discriminative performance in symptomatic individuals and maintained performance in a pre-specified asymptomatic subgroup. Its ability to generate individualized likelihood estimates directly from cardiac mechanical activity represents an important step towards expanding PTP assessment beyond traditional symptom-based frameworks. A key strength of the EMR Score is its ability to distribute patients across the full spectrum of clinical likelihood categories. This balanced stratification enabled clear differentiation between very low-, low-, moderate-, high-, and very high-likelihood groups, with each showing consistent alignment with observed disease prevalence. Individuals classified as very low-likelihood exhibited a CAD prevalence of only 2%, and none had severe disease, supporting the clinical relevance of this low-likelihood classification. Conversely, patients assigned to high and very-high likelihood categories demonstrated high observed disease burden, indicating that EMR effectively concentrates risk at the upper end of the spectrum. Compared with contemporary clinical models, EMR provides complementary insights derived from physiologic cardiac motion. The RF-CL model remains effective at identifying patients with very low clinical likelihood based on demographic and risk-factor information, whereas the EMR Score uses mechanical information from resting SCG recordings to improve separation across likelihood categories. These differences highlight the distinct foundations of clinical and physiologic prediction approaches and underscore the value of integrating mechanical cardiac information into PTP estimation. SCG captures mechanical vibrations generated during the cardiac cycle, and obstructive CAD can alter contractility, electro-mechanical coupling, and mechanical synchrony, producing measurable changes in SCG waveform timing, amplitude, and spectral characteristics, even at rest. Within the EMR Score model, continuous wavelet transform representations preserve the temporally localized spectral energy of cardiac mechanical activity. Although the convolutional neural network component does not provide direct feature-level interpretability, it learns discriminative patterns from physiologically meaningful signals. Consistent with this framework, the EMR Score correlated with angiographic severity and showed a moderate association with coronary calcium burden. The observed performance in asymptomatic individuals further suggests that alterations in cardiac mechanical function may complement symptom-based risk assessment. The EMR Score is designed as a pre-imaging risk stratification instrument rather than a substitute for confirmatory anatomical testing. SCG is rapid, portable, and non-invasive, without exposure to ionizing radiation or contrast agents, allowing integration into outpatient and primary care settings where decisions regarding referral for CCTA or invasive angiography must be made. Within this context, an electro-mechanical assessment may help identify individuals with a very low likelihood of obstructive CAD who could reasonably defer advanced imaging, while prioritizing higher-risk patients for further evaluation. To address generalizability beyond internal cross-validation, we performed an external-centre validation analysis using a geographically distinct cohort excluded from all stages of model development and threshold selection. The EMR Score maintained strong performance in this external-centre validation, supporting generalizability across recruitment centres. Nevertheless, this analysis does not replace prospective validation in fully independent healthcare systems and intended-use populations. Several methodological considerations merit acknowledgment. Although data were prospectively collected across multiple centres and an external-centre validation analysis was performed, additional prospective validation in independent clinical settings remains required. The study population consisted primarily of patients referred for CCTA or ICA, resulting in a higher prevalence of obstructive CAD than would be expected in community settings and potentially introducing referral bias. Obstructive CAD was defined as 50% or greater stenosis on CCTA or invasive angiography; while invasive angiography represents the anatomical reference standard, CCTA may introduce false-positive classifications and potential label noise. Finally, patients with arrhythmias were excluded due to the requirement for stable cardiac cycles, and future studies should evaluate model performance in broader rhythm populations. Overall, these findings support the EMR Score as a physiologically grounded, non-invasive tool for estimating the likelihood of obstructive CAD. By using cardiac mechanical information from resting SCG recordings together with patient-level clinical risk factors, the model provides individualized likelihood estimates that may inform referral decisions. Further prospective validation is required to confirm generalizability and determine clinical impact in independent clinical settings. Conclusion The EMR Score uses cardiac mechanical information from resting SCG recordings and patient-level clinical risk factors to estimate the likelihood of obstructive CAD. In this prospectively enrolled multi-centre cohort, the model improved discrimination and provided more balanced classification across likelihood categories compared with established clinical likelihood models. External-centre validation in a geographically distinct cohort supported generalizability beyond internal cross-validation. Further prospective validation in independent populations is required to confirm clinical impact and guide implementation. Supplementary Material Acknowledgements Parastoo Dehkordi and Farzad Khosrow-Khavar are employees of HeartForce Medical Inc., which develops the CardioClin device used in this study. No other disclosures were reported. The authors used ChatGPT (OpenAI) to assist in revising and editing the manuscript for clarity and grammar. All content was reviewed and verified by the authors to ensure accuracy and integrity. Contributor Information Parastoo Dehkordi, Research and Development Department, HeartForce Medical Inc., 1201-570 Granville Street, Vancouver, BC V6C 1W6, Canada. Kouhyar Tavakolian, Biomedical Engineering Department, University of North Dakota, ND, USA. Thomas Rocco, Department of Medicine, University of Rochester, New York, USA. Khurshid Fozilov, Interventional Cardiology Department, Republican Specialized Scientific and Practical Medical Center of Cardiology, Tashkent, Uzbekistan. Alisher Yuldashov, Interventional Cardiology Department, Republican Specialized Scientific and Practical Medical Center of Cardiology, Tashkent, Uzbekistan. Wojciech Zareba, Department of Medicine, University of Rochester, New York, USA. Baxtiyorjon Yuldashov, Interventional Cardiology Department, Republican Specialized Scientific and Practical Medical Center of Cardiology, Tashkent, Uzbekistan. Baxtiyor Atamuratov, Interventional Cardiology Department, Republican Specialized Scientific and Practical Medical Center of Cardiology, Tashkent, Uzbekistan. Hamid R Marateb, Research Center for Biomedical Engineering, Polytechnic University of Catalonia, Barcelona, Spain. Shukhrat Doniyorov, Interventional Cardiology Department, Republican Specialized Scientific and Practical Medical Center of Cardiology, Tashkent, Uzbekistan. Bernara Kurbanova, Interventional Cardiology Department, Republican Specialized Scientific and Practical Medical Center of Cardiology, Tashkent, Uzbekistan. Farzad Khosrow-Khavar, Research and Development Department, HeartForce Medical Inc., 1201-570 Granville Street, Vancouver, BC V6C 1W6, Canada. Supplementary material Supplementary material is available at European Heart Journal – Digital Health. Author contributions Parastoo Dehkordi (Data curation [equal], Formal analysis, Methodology, Writing—original draft [lead]), Kouhyar Tavakolian (Conceptualization, Methodology, Writing—review & editing [supporting]), Thomas Rocco (Methodology, Writing—review & editing [supporting]), Khurshid Fozilov (Data curation [lead], Writing—review & editing [supporting]), Alisher Yuldashov (Data curation [equal], Writing—review & editing [supporting]), Wojciech Zareba (Writing—review & editing [supporting]), Baxtiyorjon Yuldashov (Data curation, Writing—review & editing [supporting]), Baxtiyor Atamuratov (Data curation, Writing—review & editing [supporting]), Hamid R. Marateb (Formal analysis, Writing—review & editing [supporting]), Shukhrat Doniyorov (Data curation, Writing—review & editing [supporting]), Bernara Kurbanova (Data curation, Writing—review & editing [supporting]), and Farzad Khosrow-Khavar (Conceptualization, Data curation, Formal analysis, Methodology, Writing—review & editing [equal], Supervision [lead]) Funding This study was supported by HeartForce Medical Inc. Data availability The data underlying this article will be shared on reasonable request to the corresponding author. References Associated Data Supplementary Materials Data Availability Statement The data underlying this article will be shared on reasonable request to the corresponding author.