Machine learning derived ECG risk score improves cardiovascular risk assessment in conjunction with coronary artery calcium scoring

Abstract

Background: Precision estimation of cardiovascular risk remains the cornerstone of atherosclerotic cardiovascular disease (ASCVD) prevention. While coronary artery calcium (CAC) scoring is the best available non-invasive quantitative modality to evaluate risk of ASCVD, it excludes risk related to prior myocardial infarction, cardiomyopathy, and arrhythmia which are implicated in ASCVD. The high-dimensional and inter-correlated nature of ECG data makes it a good candidate for analysis using machine learning techniques and may provide additional prognostic information not captured by CAC. In this study, we aimed to develop a quantitative ECG risk score (eRiS) to predict major adverse cardiovascular events (MACE) alone, or when added to CAC. Further, we aimed to construct and validate a novel nomogram incorporating ECG, CAC and clinical factors for ASCVD.

Methods: We analyzed 5,864 patients with at least 1 cardiovascular risk factor who underwent CAC scoring and a standard ECG as part of the CLARIFY study (ClinicalTrials.gov Identifier: NCT04075162). Events were defined as myocardial infarction, coronary revascularization, stroke or death. A total of 649 ECG features, consisting of measurements such as amplitude and interval measurements from all deflections in the ECG waveform (53 per lead and 13 overall) were automatically extracted using a clinical software (GE Muse™ Cardiology Information System, GE Healthcare). The data was split into 4 training (Str) and internal validation (Sv) sets [Str (1): Sv (1): 50:50; Str (2): Sv (2): 60:40; Str (3): Sv (3): 70:30; Str (4): Sv (4): 80:20], and the results were compared across all the subsets. We used the ECG features derived from Str to develop eRiS. A least absolute shrinkage and selection operator-Cox (LASSO-Cox) regularization model was used for data dimension reduction, feature selection, and eRiS construction. A Cox-proportional hazards model was used to assess the benefit of using an eRiS alone (Mecg), CAC alone (Mcac) and a combination of eRiS and CAC (Mecg+cac) for MACE prediction. A nomogram (Mnom) was further constructed by integrating eRiS with CAC and demographics (age and sex). The primary endpoint of the study was the assessment of the performance of Mecg, Mcac, Mecg+cac and Mnom in predicting CV disease-free survival in ASCVD.

Findings: Over a median follow-up of 14 months, 494 patients had MACE. The feature selection strategy preserved only about 18% of the features that were consistent across the various strata (Str). The Mecg model, comprising of eRiS alone was found to be significantly associated with MACE and had good discrimination of MACE (C-Index: 0.7, p = <2e-16). eRiS could predict time-to MACE (C-Index: 0.6, p = <2e-16 across all Sv). The Mecg+cac model was associated with MACE (C-index: 0.71). Model comparison showed that Mecg+cac was superior to Mecg (p = 1.8e-10) or Mcac (p < 2.2e-16) alone. The Mnom, comprising of eRiS, CAC, age and sex was associated with MACE (C-index 0.71). eRiS had the most significant contribution, followed by CAC score and other clinical variables. Further, Mnom was able to identify unique patient risk-groups based on eRiS, CAC and clinical variables.

Conclusion: The use of ECG features in conjunction with CAC may allow for improved prognostication and identification of populations at risk. Future directions will involve prospective validation of the risk score and the nomogram across diverse populations with a heterogeneity of treatment effects.

Keywords: artificial intelligence; atherosclerotic cardiovascular diseases (ASCVD); electrocardiogram (ECG); machine learning; nomogram; risk assessment/classification.

Conflict of interest statement

AM is an equity holder in Picture Health, Elucid Bioimaging, and Inspirata Inc. Currently he serves on the advisory board of Picture Health, Aiforia Inc, and SimBioSys. He also currently consults for Biohme, SimBioSys and Castle Biosciences. He also has sponsored research agreements with AstraZeneca, Boehringer-Ingelheim, Eli-Lilly and Bristol Myers-Squibb. His technology has been licensed to Picture Health and Elucid Bioimaging. He is also involved in 3 different R01 grants with Inspirata Inc. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2022 Siva Kumar, Al-Kindi, Tashtish, Rajagopalan, Fu, Rajagopalan and Madabhushi.

Figures

Figure 1

Overall workflow. The first step involves collecting ECG tests, manual CAC scores and clinical data from eligible patients. The ECG features are then automatically extracted using the commercially available GE MUSE software. Top ECG features were selected using the LASSO feature selection method and used for constructing eRiS. Mnom was constructed using clinical features and eRiS. Mecg and Mnom were validated for prognostic performance and predicting downstream MACE events. LASSO: least absolute shrinkage and selection operator; eRiS: ECG risk score, Mecg: Cox PH model using eRiS alone; Mnom: Nomogram with eRiS, CAC and clinical factors.

Figure 2

ECG risk score predicts MACE events. Patients with higher ECG risk score (eRiS) correlate with occurrence of MACE events, demonstrating the value of considering ECG as a factor in determining probability of a MACE event. X-axis denotes patients arranged in ascending value of eRiS scores.

Figure 3

Kaplan-Meier plot for MACE-free survival according to eRiS-based risk groups in Sv. The eRiS threshold of −0.055 showed two distinct groups of high vs. low MACE-free survival in Sv (4).

Figure 4

Receiver operating characteristic (ROC) curve for CAC only vs. eRiS+CAC shows the benefit of adding eRiS to CAC for better prediction of the probability of a MACE event in (A). CAC+eRiS showed better performance than CAC+PCE (C-index: 0.72 vs. 0.67) for patients who had PCE available in (B).

Figure 5

Precision-Recall (PR) curve for CAC only vs. eRiS+CAC shows the benefit of adding eRiS to CAC for better prediction of the probability of a MACE event in (A) (Average F1 statistic 0.20 vs. 0.21. PR AUC: 0.68 vs. 0.71). (B) CAC+eRiS showed better performance than CAC+PCE for patients who had PCE available (Average F1 statistic 0.28 vs. 0.30. PR AUC: 0.68 vs. 0.71).

Figure 6

Additive benefit of eRiS to CAC: Hazard ratio is not attenuated when eRiS is adjusted by CAC score, indicating a strong relationship with MACE which is not weakened by the addition of CAC. Similar results were seen with other splits.

Figure 7

Kaplan-Meier plot for MACE-free survival according to eRiS+CAC and eRiS only risk groups in Sv (4). (A) The eRiS+CAC threshold showed worse prognosis for patients with high eRiS combined with high CAC score. HR between high and low risk: 6.72 [4.42–10.22]. (B) Similar observation seen in eRiS based segregation. HR between high and low risk: 5.22 [2.54–10.75].

Figure 8

ECG Nomogram (Mnom) demonstrates relative contribution of each covariate in MACE prediction. ECG risk score has the most significant contribution, followed by CAC score and other clinical variables.

Figure 9

Kaplan-Meier plot for MACE-free survival according to Mnom risk groups for Sv (4) [C-index 0.6 (se = 0.023]). HR between high and low risk: 3.24 [1.02–10.30].