This single-center, retrospective cohort study included adult patients suspected of having coronary artery disease (CAD) who underwent both CCTA and MCG examinations within a two-week interval from December 2023 to June 2025. CT-FFR data were available for all patients, and a subset underwent invasive FFR. Functionally significant ischemia was defined as CT-FFR $\le$0.80. In cases where both CT-FFR and invasive FFR were available, discrepancies were resolved by FFR. This study was conducted in accordance with the revised 2013 Helsinki Declaration and approved by the Ethics Committee of the Affiliated Hangzhou First People’s Hospital, Westlake University School of Medicine (Ethic Approval Number: IIT-20231214-0298-02). Written informed consent was obtained from all participants prior to enrollment.

Initially, a total of 450 consecutive patients were screened. The inclusion criteria were as follows: (1) aged over 18 years; (2) presence of angina or angina-like symptoms, such as chest pain or dyspnea, suspected of CAD; (3) completion of both CCTA and MCG within a two-week interval; (4) availability of FFR or CT-FFR measurements. Exclusion criteria included: (1) severe cardiac valvular disease; (2) abnormal heart rhythm; (3) prior coronary artery revascularization; (4) poor image quality; (5) inability to complete MCG acquisition due to intolerance of the magnetic shielding chamber or interference from metallic implants/devices; (6) severe renal dysfunction (estimated glomerular filtration rate (eGFR) 30 mL/min/1.73 m²) or dialysis; (7) history of severe hypersensitivity to iodinated contrast media; (8) pregnancy or breastfeeding. After applying these criteria, 275 patients were included in the final analysis. The patient enrollment process is detailed in Fig. 1. All 275 patients had evaluable CT-FFR values, and 12 patients (4.4%) also underwent invasive FFR.

Fig. 1.

Flowchart of study population. CCTA, coronary computed tomography angiography; MCG, magnetocardiography; FFR, fractional flow reserve.

This study utilized an ultra-high-sensitivity MCG system based on the SERF principle, comprising a 64-channel atomic magnetometer array, a high-performance magnetic shielding chamber, and a multifunctional non-magnetic examination bed. This system can detect cardiac magnetic fields as weak as 10^-7 times the strength of the Earth’s magnetic field. Data acquisition followed a standardized protocol: subjects were positioned in a supine position on the examination bed, with the sensor array center placed 2 cm below the xiphoid process. Resting MCG signals were continuously recorded for 3 minutes. ECG signals were simultaneously acquired within 5 minutes before and after the MCG scan for heartbeat segmentation and signal comparison. Patients with metallic implants or other strong magnetic interference sources were excluded prior to scanning. During data acquisition, real-time monitoring identified potential issues like body motion, respiratory drift, channel instability, or transient environmental magnetic fluctuations, ensuring only high-quality signal segments were recorded. The system’s performance was regularly validated to ensure stable signal acquisition, with no significant variance observed during repeated scans in the same setting. Additionally, a Signal Quality Score was calculated for all MCG scans, with all patients achieving a score $\ge$90, ensuring the reliability of the acquired signals. MCG data were collected in a controlled, magnetically shielded environment to minimize external magnetic interference, ensuring high data quality and consistency. This technology offers significant advantages, including non-invasiveness, contactless measurement, and the lack of cryogenic cooling requirements, while enabling high-precision detection of magnetic field variations related to myocardial electrophysiological activity.

Coronary CT angiography was performed using a second-generation dual-source CT scanner (SOMATOM Definition Flash, Siemens Healthineers, Forchheim, Germany). Five minutes before scanning, patients received sublingual nitroglycerin (0.5 mg) to dilate the coronary arteries. A non-ionic contrast agent (Iodomane 370 mg/mL, Bayer, Germany, 60 mL) was injected intravenously at 4.5–5.0 mL/s, followed by a 30 mL saline flush at the same rate using a dual-head power injector. Scanning was initiated by automated bolus tracking with a 7-second delay after the attenuation in the ascending aorta reached 100 Hounsfield units. CT scanning was performed during 30%–80% of the R-R interval with prospective ECG gating, and the scanner’s automatic phase selection feature was used to obtain the optimal systolic (33%–46% of the R-R interval) and diastolic (66%–75% of the R-R interval) images. Reconstructions were generated with a slice thickness of 0.75 mm and a field of view of 200–250 mm. Additional acquisition parameters included a tube voltage of 120 kV, a reference tube current of 320 mAs with CARE Dose 4D modulation, a gantry rotation time of 0.28 s per rotation, and a collimation of 64 $\times$ 2 $\times$ 0.6 mm. This standardized protocol reduced radiation exposure through prospective gating and adaptive phase selection, while dual-phase reconstruction minimized coronary motion artifacts and improved the accuracy of stenosis assessment.

The multichannel cardiac magnetic signals collected by the MCG system were automatically processed to extract 23 quantitative parameters, including temporal intervals, magnetic field strength indices, as well as angular and spatial dipole metrics (Supplementary Table 1). CCTA images were analyzed using an AI-assisted platform (CoronaryDoc, Shukun Technology, Beijing, China), which generated 21 anatomical and functional parameters (Supplementary Table 2). For coronary plaque assessment, vessel-level plaque composition was recorded for the left anterior descending artery (LAD), left circumflex artery (LCX), right coronary artery (RCA), and as no plaque, non-calcified, calcified, or mixed plaque. CT-FFR was used solely to define the reference-standard outcome and was not used as an input feature for model development. Parameters with more than 30% missing data were excluded. Missing values, which occurred only in a few continuous variables, were imputed using median substitution. For feature screening, univariate analyses were performed to identify variables potentially related to functionally significant ischemia. Variables demonstrating at least borderline association (p 0.1) were considered candidates for inclusion. Clinically important parameters supported by prior physiological and imaging evidence were also retained to avoid omitting relevant predictors. Feature selection thus followed a combined strategy integrating both statistical signals and clinical relevance, rather than relying solely on p-values.

All 275 patients were included in model development, with five-fold stratified cross-validation used to generate training and validation folds. This framework was consistently applied across three model variants: MCG-only, CCTA-only, and combined MCG-CCTA models. For each model variant, five ML algorithms were implemented: logistic regression (LR), support vector machine (SVM), random forest (RF), naive bayes (NB), and extreme gradient boosting (XGBoost). Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1 score. Comparative analyses across the three model variants and five algorithms were performed to identify the best-performing configuration for diagnosing functional myocardial ischemia. To enhance interpretability, Shapley additive explanations (SHAP) was conducted to quantify the contribution of individual features to model predictions.

Statistical analyses were performed using SPSS software version 25.0 (SPSS Inc, Chicago, IL, USA) and R software version 4.5.1 (R Foundation for Statistical Computing, Vienna, Austria). The normality of continuous variables was assessed using standard tests for Gaussian distribution. As none of the continuous variables followed a normal distribution, continuous variables are presented as median (interquartile range, IQR), and group comparisons for continuous variables were performed using the Mann–Whitney U test. Categorical variables were compared utilizing the $\chi$² test or Fisher’s exact test, as appropriate. The diagnostic performance of the MCG model, CCTA model, and the combined MCG-CCTA model was evaluated using receiver operating characteristic (ROC) curve analysis, with optimal cutoff values determined by the Youden index. AUC was calculated and compared across models using the DeLong test to evaluate discriminatory ability. Additionally, sensitivity, specificity, PPV, NPV, and accuracy were reported with 95% confidence interval (CI). Two-tailed p-values 0.05 were considered significant.

A total of 275 eligible patients were included in this study, comprising 98 (35.6%) patients in the ischemia group and 177 (64.4%) patients in the non-ischemia group. Baseline characteristics are presented in Table 1. The ischemia group had a notably higher proportion of males (78.6% vs. 47.5%, p 0.01), along with a higher prevalence of smoking (26.5% vs. 11.9%, p 0.01), diabetes mellitus (35.7% vs. 18.6%, p 0.01), and hypertension (71.4% vs. 59.3%, p = 0.046). In addition, the ischemia group showed lower total cholesterol (4.04 mmol/L vs. 4.41 mmol/L, p = 0.012) and high-density lipoprotein cholesterol levels (1.04 mmol/L vs. 1.14 mmol/L, p 0.01). There were no other significant differences between the two groups.

Univariate analysis revealed that several variables had at least a borderline association (p 0.1) with functionally significant ischemia, including PR interval, ST-segment duration, the R/T magnetic strength ratio, CACS, and plaque composition in the LAD, LCX, and RCA. To ensure no clinically important electrophysiological information was missed, additional MCG parameters relevant to ischemia (QRS interval, QT interval, R-wave peak dipole distance, and T-wave peak dipole distance) were also included. In total, 11 features were used for model development: four CCTA-derived variables (LAD, LCX, RCA plaque characteristics, and CACS) and seven MCG-derived variables (PR interval, QRS interval, QT interval, ST-segment duration, R/T magnetic strength ratio, R-wave peak dipole distance, and T-wave peak dipole distance). The three plaque composition variables were treated as categorical predictors, while CACS and all seven MCG variables were modeled as continuous predictors. These variables were selected not only for their statistical associations but also as recognized indicators of ischemia-related structural and electrophysiological changes, with their mechanistic roles further discussed in the “Discussion” section.

Among the five ML methods tested (LR, RF, NB, SVM, and XGBoost), XGBoost consistently outperformed the others across all modeling strategies (MCG-only, CCTA-only, and combined MCG-CCTA). Overall diagnostic metrics for all algorithms and model types are summarized in Supplementary Table 3, and the relative performance profiles are illustrated by radar plots in Fig. 2. For XGBoost models, the MCG-only and CCTA-only configurations achieved AUC values of 0.769 (95% CI: 0.708–0.829) and 0.755 (95% CI: 0.692–0.818), respectively, offering a favorable balance between sensitivity and specificity compared with the other classifiers. Integrating MCG and CCTA further enhanced diagnostic accuracy. The combined XGBoost model delivered the best overall results, with an AUC of 0.829 (95% CI: 0.773–0.885), accuracy of 0.800, sensitivity of 0.704, and specificity of 0.853. This represented relative AUC increases of 7.9% over the MCG model and 9.8% over the CCTA model (DeLong test, all p 0.05). Predictive values were also favorable (PPV = 0.726, NPV = 0.839), resulting in the highest F1-score (0.715). Fig. 3 displays the ROC curves of the MCG-only, CCTA-only, and combined MCG-CCTA XGBoost models, along with their corresponding confusion matrices at the optimal cutoff, which may be helpful as a concise visual summary of discrimination and correct/incorrect classifications. To address the potential impact of confounding by conventional risk factors, we performed subgroup analyses by sex, hypertension, and smoking status. The combined MCG-CCTA model showed similar discrimination across these strata (Supplementary Table 4).

Fig. 2.

Radar plots illustrating the diagnostic performance of three models based on five machine learning (ML) methods. (A) SVM model. (B) RF model. (C) LR model. (D) NB model. (E) XGBoost model. (F) MCG models. (G) CCTA models. (H) Combined models. AUC, area under the receiver operating characteristic curve; F1, F1-score; NPV, negative predictive value; PPV, positive predictive value; SVM, support vector machine; RF, random forest; LR, logistic regression; NB, naive bayes; XGBoost, extreme gradient boosting.

Fig. 3.

ROC curves and confusion matrices for MCG, CCTA, and MCG-CCTA models constructed by XGBoost. (A) ROC curve of the individual MCG model, the individual CCTA model, and the MCG-CCTA fusion model. (B) Confusion matrix analysis for the individual MCG model. (C) Confusion matrix analysis for the individual CCTA model. (D) Confusion matrix analysis for the MCG-CCTA fusion model.

SHAP analysis based on the XGBoost combined model (MCG-CCTA fusion) revealed that both MCG and CCTA features significantly contributed to model predictions, confirming their complementary value (Fig. 4). Based on the mean absolute SHAP values, the most influential features were PR interval, LCX plaque composition, ST segment, CACS, and LAD plaque composition, followed by R/T magnetic strength ratio, RCA plaque composition, QRS interval, QT interval, dipole distance at T-wave peak, and dipole distance at R-wave peak. The top five features accounted for approximately 70% of the total feature importance, with MCG-derived timing indices ranking prominently and CCTA-derived plaque composition and calcification measures also making substantial contributions. These findings provide an interpretable explanation for the improved discrimination of the fusion model observed in the performance comparison.

Fig. 4.

Global interpretability of the MCG-CCTA combined model constructed by XGBoost using SHAP values.