Deep neural network trained on surface ECG improves diagnostic accuracy of prior myocardial infarction over Q wave analysis

Background

Detection of prior myocardial infarction (MI) may inform arrhythmia treatment and prognosis, yet cardiac imaging is resource intensive. ECG Q-wave analysis is quick and inexpensive but has poor accuracy for assessing prior MI.

Purpose

To evaluate the ability of a deep neural network (DNN) trained on the surface ECG to identify patients with prior MI.

Methods

We assessed 608 well-characterized patients (61.4±14.5 years, 31.2% female) at 2 academic centers. From one 12-lead ECG, median beats were calculated in 3 orthogonal planes (X, Y, Z; Fig. 1A) and used to train a DNN to identify a history of prior MI. Accuracy was compared to manual assessment of pathologic Q waves, defined as a deflection >25% of the subsequent R wave, >40ms in width, and >0.2mV amplitude in 1 of 3 ECG planes.

Results

Of 608 patients, 175 had history of MI (28.7%). The DNN outperformed the accuracy of pathologic Q waves. In training, DNN converged to >98% accuracy and in testing, its accuracy was 71±5% (Fig. 1B) (k=5-fold cross validation). This outperformed the 62% accuracy of pathologic Q waves in this study (red dotted line, Fig. 1B). In the validation cohort, DNN provided an area under the receiver operating characteristics curve of 0.730 (Fig. 1C).

Conclusion

Deep learning of a 12-lead ECG can identify features of prior myocardial injury more accurately than Q-wave analysis. In attempting to improve these results further, studies should explain what inputs weighted DNN decisions, and identify those that reflect abnormalities detectable clinically or on imaging.

Figure 1

Funding Acknowledgement

Type of funding source: Public grant(s) – National budget only. Main funding source(s): NIH NRSA F32

Comparing machine learning approaches to identify myocardial scar from the ECG

Background

Detection of myocardial infarction (MI) traditionally requires ECG Q waves, which have poor sensitivity, or imaging, which is time consuming. We hypothesized that machine learning (ML) of the ECG could identify prior MI, but its accuracy may depend highly upon the architecture and parameters chosen.

Purpose

To compare ML architectures that predict prior MI from the ECG.

Methods

We curated ECGs in 608 patients seen in cardiology clinics at 2 centers. We transformed 12-lead ECGs to median beats in Frank (X, Y, Z) planes (fig. A). We tested 3 architectures: a 1D deep neural network (DNN), a 3D neural network, and a support vector machine (SVM). The 1D DNN used only temporal convolutions (fig B) while the 3D DNN uses a spatial convolution (fig C) prior to the fully-connected layer (fig. C). Predictive accuracy for history of MI was compared for all architectures (fig. D).

Results

Patients (61.4±14.5 years, 31.2% female) had a 28.7% (175/608) prevalence of prior MI. Optimized SVM of 6 features provided accuracy of 66.1% for identifying prior MI, similar to ECG Q wave analysis. 1D DDN had accuracy of 63.6% with an area under curve (AUC) of 0.625. 3D DNN outperformed 1D DNN and SVM, providing an accuracy of 71±5% (using k=5-fold cross validation), with an AUC of 0.730.

Conclusion

ECG machine learning can identify prior MI better than Q wave analysis, but is sensitive to technical parameters and specific computational architecture. It is important to develop a framework to enable robust comparisons of different ML studies and future refinements.

Funding Acknowledgement

Type of funding source: Public grant(s) – National budget only. Main funding source(s): National Institutes of Health - United States

P5645Lifetime sex-specific sudden cardiac death prediction using ECG global electrical heterogeneity: the atherosclerosis risk in communities (ARIC) study

Background

Sex-based differences in sudden cardiac death (SCD) exist and screening methods for SCD are inadequate.

Purpose

To develop sex-specific lifetime risk prediction models using electrocardiographic (ECG) global electrical heterogeneity (GEH) and clinical characteristics.

Methods

Participants from the Atherosclerosis Risk in Communities study with analyzable ECGs (n=14,725; age, 54.2±5.8 yrs; 55% female, 74% white) were followed up for 24.4 years (median). Traditional ECG and GEH variables were measured on 12-lead ECGs. A Cox regression model was used to develop a prediction model. In women, the final model included race, age, coronary heart disease (CHD), stroke, hypertension, diabetes, smoking, high-density lipoprotein, albumin, uric acid, education level, heart rate, QTc, sum absolute QRST integral, spatial peak QRS-T angle. In men, the final prediction model included age, race, CHD, stroke, hypertension, diabetes, total cholesterol, physical activity, smoking, serum phosphorus, albumin, chronic kidney disease, spatial area QRS-T angle, area spatial ventricular gradient (SVG) elevation and magnitude, and peak SVG magnitude.

Results

There were a total of 530 SCDs. Our prediction models showed robust prediction of SCD in both sexes [(Harrell's C-statistic women 0.863 (95% CI 0.845–0.882), men 0.786 (95% CI 0.786–0.803)]. In women when ECG and GEH variables were added to clinical variables, the net reclassification improved by 9% (P=0.001) (Table). In men there was no significant reclassification improvement.

Net reclassification Lifetime SCD Risk: Clinical + ECG + GEH Variables Women Men <5% 5–15% >15% Total <5% 5–15% >15% Total SCD Cases <5% 82 14 0 96 103 16 0 119 5–15% 7 59 10 76 12 116 12 140 >15% 0 0 20 20 0 5 74 79 Lifetime SCD Risk: Total 89 73 30 192 115 137 86 338 Clinical Variables Only Non-Cases <5% 6,956 131 2 7,089 4,411 264 0 4,675 5–15% 180 509 42 731 210 1,059 48 1,317 >15% 0 28 84 112 0 56 214 270 Total 7,136 668 128 7,932 4,621 1,379 262 6,262

Conclusions

We were the first to develop sex-specific lifetime SCD prediction models. The addition of ECG GEH to clinical variables improved SCD risk reclassification in women, but not in men. Prediction of SCD was more accurate in women as compared to men.