Deep Learning to Predict Cardiac Magnetic Resonance–Derived Left Ventricular Mass and Hypertrophy From 12-Lead ECGs

UK Biobank data are publicly available by application (www.ukbiobank.ac.uk). MGB data contain protected health information and cannot be shared publicly. The code underlying LVM-AI is accessible at https://github.com/broadinstitute/ml4h/tree/master/model_zoo/left_ventricular_mass_from_ecg_student_and_mri_teacher.

The UK Biobank is a prospective cohort of 502 629 participants recruited between 2006 and 2010.¹¹ Briefly, ≈9.2 million individuals aged 40 to 69 living within 25 miles of the 22 assessment centers in England, Wales, and Scotland were invited, and 5.4% participated in the baseline assessment. Extensive questionnaires, physical measures, and biological samples were collected at recruitment, with multimodal imaging obtained in a large subset. All participants are followed for health outcomes through linkage to national datasets. Participants provided written informed consent. The UK Biobank was approved by the UK Biobank Research Ethics Committee (reference number 11/NW/0382). Use of UK Biobank (application 7089) and MGB data were approved by the local MGB Institutional Review Board.

For this analysis, we included all individuals who underwent both resting 12-lead ECG and CMR contemporaneously during the UK Biobank imaging assessment. Demographics including age, sex, and race, and physical measurements including height, weight, and body mass index (BMI) were obtained at the imaging assessment or study visit most closely preceding.

The UK Biobank CMR protocol has been described previously.¹² Briefly, all CMRs were acquired on a clinical wide-bore 1.5 Tesla scanner (MAGNETOM Aera, Syngo Platform VD13A, Siemens Healthcare, Erlangen, Germany) and used balanced steady-state free precession with typical parameters.

Resting 12-lead ECG data were downloaded as XML files and converted into arrays of lead amplitude sequence data. CMR images were downloaded as DICOM slabs and converted into arrays of 3-dimensional voxel data. We utilized a validated deep learning model (ML4H_seg)¹³ to extract LV mass from the CMR images, which served as the ground truth for LVM-AI. A conceptual overview of the study is shown in Figure 1.

After setting aside a random sample (n=4903) as an internal test set (UK Biobank Test, Figure 2), we trained LVM-AI within 32 239 individuals with paired CMR and 12-lead ECG. LVM-AI is a one-dimensional convolutional neural network designed to infer LV mass using 12-lead ECG (Figure I in the Data Supplement). LVM-AI was provided with the entire 10 seconds of the 12-lead ECG waveform as well as participant age, sex, and BMI. Given the clinical importance of diagnosing LVH (ie, elevated LV mass), LVM-AI utilized a loss function giving additional weight to errors at the high extreme of the LV mass distribution.¹⁴

We sought to compare LVM-AI to the lead aVL (R wave in lead aVL >1.1 mV),⁴ Sokolow and Lyon,⁵ and Cornell voltage⁴ rules for diagnosing LVH using 12-lead ECG (Table I in the Data Supplement). Although ECG rules adjusted for additional clinical factors have been proposed,¹⁵ we focused on the original rules since they are commonly used.⁶ However, in secondary analyses we assessed the performance of a Cornell voltage product adjusted for age, sex, BMI, and hypertension.¹⁵

To apply ECG rules, we extracted lead-specific R and S wave amplitudes from the 12-lead ECG XML files. To validate ECG rule calculation, we developed a plotting function¹⁶ to reconstruct 12-lead ECG waveforms for visual interpretation (Figure II in the Data Supplement). Two cardiologists (S.K. and J.P.P.) manually assessed the accuracy of extracted amplitudes and presence of LVH by each rule. Both per-lead (98.7%–100%) and per-rule (98.7%–99.3%) accuracy met prespecified criteria (>90%). Interrater agreement on a per-lead (Gwet’s AC1¹⁷ 0.98) and per-rule (range, 0.91–0.96) basis was excellent (Tables II and III in the Data Supplement). Given that ECG-based LVH rules have not been validated in the setting of left bundle branch block, left anterior fascicular block, and limb lead reversal, we excluded these tracings from ECG rule analyses.⁶

We assessed for associations between LVM-AI predicted LV mass and incident AF, myocardial infarction (MI), HF, and ventricular arrhythmias (VA) within 35 350 participants with follow-up clinical data available after the imaging assessment. Diseases were defined using self-report and inpatient International Classification of Diseases-9/International Classification of Diseases-10 codes (updated through March 31, 2021, Table IV in the Data Supplement). Follow-up started at ECG acquisition and spanned until the earliest of an event, death, or last follow-up. Last follow-up was dependent upon the availability of linked hospital data and was therefore defined as March 31, 2020 for participants enrolled in England (99.6%), October 31, 2016 for participants enrolled in Scotland (0.2%), and February 29, 2016 for participants enrolled in Wales (0.2%).

We tested LVM-AI in the external MGB health care–related dataset (Figure 2). First, we assessed correlation between LVM-AI predicted LV mass and CMR-derived LV mass in a sample of individuals from the MGB Biobank with both ECG and CMR performed within 1 year of each other (MGB test). Second, we compared the accuracy of LVM-AI predicted LVH to traditional ECG-based rules in MGB Test. Third, we assessed for associations between LVM-AI predicted LVH and incident AF, MI, HF, and VA in a previously described MGB-based dataset independent of MGB Test (MGB Outcomes).¹⁸ Disease definitions have been previously validated (Table IV in the Data Supplement).^18,19 External validation methods are described in detail in Methods in the Data Supplement.

We calculated the Pearson correlation and mean absolute error between LVM-AI predicted LV mass and CMR-derived LV mass. We assessed agreement using Bland-Altman plots.²⁰ We quantified calibration using the mean within-individual difference between LVM-AI predicted and CMR-derived LV mass. To compare LVM-AI to traditional ECG rules, we calculated sensitivity, specificity, positive predictive value, negative predictive value, and c-statistic of each using CMR-derived LVH as the reference. We also generated contingency tables and performed net reclassification analyses. In all analyses, LVH was defined as indexed LV mass >72 g/m² (men) and >55 g/m² (women).²¹ The sex-specific 90th percentile of LV mass index was a secondary LVH definition. Indexing for body surface area was performed using the DuBois formula.²² CIs were generated using the exact method and test characteristics were compared using 1000-sample bootstrapping.

To assess LVM-AI behavior, we produced saliency maps depicting areas of the ECG having the largest gradients (ie, greatest influence on LV mass predictions). For each individual, we additionally identified the ECG lead having the highest absolute gradient, as a surrogate for the most influential ECG lead on that individual’s LV mass estimate.

We assessed associations between LVM-AI predicted LVH and incident AF, MI, HF, and VA using Cox proportional hazards models adjusted for age and sex. We built similar models using LVM-AI predicted LV mass index as a continuous exposure. The proportional hazards assumption was assessed by inspecting Schoenfeld residuals. Substantial deviations from proportional hazards (observed for age and sex only) were modeled using interaction terms including strata of person-time. We plotted cumulative risk of events within strata of LVM-AI predicted LVH using the Kaplan-Meier method. Given evidence that anatomic LVH may provide complementary prognostic information to ECG rule-based LVH,²³ we fit analogous Cox proportional hazards models using (1) the ECG rules and (2) both the ECG rules and LVM-AI predicted LVH as exposures of interest.

We performed sensitivity analyses to assess robustness of our results. First, we trained a version of LVM-AI taking ECG alone as input. Second, we trained a version of LVM-AI utilizing an unweighted logcosh function regression loss (a loss function giving equal weight to errors at either extreme of the LV mass distribution). Third, we compared LVM-AI with a modified Cornell voltage product adjusted for age, sex, BMI, and hypertension.¹⁵

Analyses were performed using Python v3.8 (including TensorFlow),^24,25 the ML4H codebase,¹⁶ and R v4.0 (packages data.table, ggplot2, epiR, pROC, nricens).²⁶ P <0.05 were considered statistically significant.

A total of 37 142 individuals had both CMR-derived LV mass and 12-lead ECG available. Setting aside 4903 individuals for UK Biobank Test, we trained LVM-AI on a total of 32 239 participants (Figure 2). Individuals in the training set had a mean age of 64.2±7.5 years and 52% were female. The mean CMR-derived LV mass index was 47.0±9.6 g/m². Other characteristics are shown in Table 1.

Individuals in UK Biobank Test (mean age, 63.6±7.7, 53% female) were similar in composition to the training set, whereas individuals in MGB Test (mean age 55.5±14.6, 46% female) had substantially greater cardiac comorbidity (Table 1). The mean CMR-derived LV mass index was 47.3±9.4 g/m² in UK Biobank Test and 61.1±21.5 g/m² in MGB Test. The prevalence of CMR-derived LVH was 126/4903 (2.6%) in UK Biobank Test and 454/1327 (34%) in MGB test.

In UK Biobank Test, when compared with CMR-derived LV mass, LVM-AI demonstrated good correlation (r=0.79 [95% CI, 0.78–0.80], P<0.001), accuracy (mean absolute error 12.6 g [95% CI, 12.3–12.9]), and calibration (within-individual mean difference −3.1 g [95% CI, −3.5 to −2.6], Figure 3). Within MGB Test, LVM-AI demonstrated moderate correlation (r=0.48 [95% CI, 0.44–0.52], P<0.001) but poor accuracy (mean absolute error 117.0 g [95% CI, 111.7–122.3]), due largely to systematic overestimation (within-individual mean difference −111.8 g [95% CI, −117.3 to −106.2]). After linear recalibration (Methods in the Data Supplement), correlation (r=0.60 [95% CI, 0.57–0.64], P<0.001) and accuracy (mean absolute error, 28.4 g [95% CI, 27.1–29.8]) were improved (Figure 3). In both test sets, correlation was slightly lower for indexed LV mass (UK Biobank r=0.63 [95% CI, 0.61–0.64]; MGB r=0.51 [95% CI, 0.47–0.55], Figure 3). Bland-Altman plots demonstrated greater agreement in UK Biobank Test (95% limits of agreement −15.1 to 18.6 g/m²) than MGB Test (95% limits of agreement −36.3 to 36.3 g/m²), as well as a tendency to make conservative estimation errors in MGB (Figure III in the Data Supplement). Sex-stratified distributions of actual and predicted LV mass are shown in Figure IV in the Data Supplement. Plots depicting learned embeddings from LVM-AI are shown in Figure V in the Data Supplement. Saliency maps demonstrated that components of the ECG waveform plausibly relevant for LV mass estimation (eg, p wave, early portion of QRS complex) had the greatest impact on LV mass estimates (Figure 4 and Figure VI in the Data Supplement). On an individual basis, the ECG lead exerting the greatest influence on LV mass estimates was most frequently V5 (97.4%), followed by V4 (2.4%) then V1 (0.3%).

In total, 4417 (90.0%) in UK Biobank Test and 1062 (77.5%) in MGB Test had a 12-lead ECG suitable for ECG rule calculation (Figure 2). LVM-AI demonstrated moderate LVH discrimination (UK Biobank c-statistic, 0.653 [95% CI, 0.608–0.698]; MGB 0.621 [95% CI, 0.592–0.649]), which was favorable when compared individually to Sokolow-Lyon, Cornell, and lead aVL criteria (P<0.001 for all, Figure 5 and Table V in the Data Supplement). LVH discrimination using LVM-AI was similar to the presence of any ECG rule in UK Biobank Test (c-statistic, 0.618 [95% CI, 0.574–0.663]; P=0.11), but significantly greater than any ECG rule in MGB Test (0.588 [95% CI, 0.564–0.611]; P=0.02).

When compared with any ECG rule, LVM-AI had greater sensitivity and specificity in UK Biobank Test (sensitivity 34% [95% CI, 25–44] versus 32% [95% CI, 24–42]; specificity 96% [95% CI, 96–97] versus 91% [95% CI, 90–92]), and greater sensitivity but lower specificity in MGB Test (sensitivity 41% [95% CI, 36–46] versus 24% [95% CI, 20–29]; specificity 83% [95% CI, 80–86] versus 93% [95% CI, 91–95], Figure 5). When compared with individual ECG rules, LVM-AI had greater sensitivity, with comparable or moderately lower specificity (Figure 5; Table V in the Data Supplement). Net reclassification improvement using LVM-AI versus any ECG rule was 0.071 (95% CI, −0.016 to 0.17) in UK Biobank Test, and 0.067 (95% CI, 0.0072–0.13) in MGB Test (Table VI in the Data Supplement), with increased case detection in both sets (UK Biobank 1.9% [95% CI, −7.6 to 13]; MGB 17% [95% CI, 12–22]).

In the UK Biobank and MGB Outcomes samples, LVM-AI predicted LVH was associated with incident AF (hazard ratio, 1.84 [95% CI, 1.29–2.63] in UK Biobank; 1.34 [95% CI, 1.21–1.50] in MGB), MI (hazard ratio 1.80 [95% CI, 1.09–2.96]; 1.29 [95% CI, 1.15–1.46]), HF (3.97 [95% CI, 2.70–5.84]; 1.49 [95% CI, 1.37–1.62]), and VA (3.16 [95% CI, 1.62–6.18]; 1.71 [95% CI, 1.47–1.99]). Associations were similar using LVH defined as the 90th percentile of LV mass index and using LV mass index as a continuous variable (Table 2 and Figure VII in the Data Supplement). Cumulative risk curves stratified by presence of LVM-AI predicted LVH are shown in Figure 6 and Figure VIII in the Data Supplement.

In secondary analyses, associations between ECG rule-based LVH and incident events varied by specific rule, although the presence of LVH by any ECG rule was consistently associated with incident AF and HF (Table VII in the Data Supplement). In models including both LVM-AI predicted and ECG rule-based LVH, LVM-AI predicted LVH was independently associated with incident events (Table VIII in the Data Supplement). Cumulative risk curves stratified by the presence of LVH using LVM-AI and any ECG rule are shown in Figure IX in the Data Supplement.

An age, sex, BMI, and hypertension-adjusted Cornell voltage product had lower sensitivity and specificity than LVM-AI (Table IX in the Data Supplement and Figure X in the Data Supplement). In the UK Biobank, a version of LVM-AI trained using ECG alone (ie, without age, sex, or BMI) had similar correlation with CMR-derived LV mass (r=0.72 [95% CI, 0.70–0.73]) and diagnostic performance for LVH (c-statistic 0.654 [95% CI, 0.608–0.700]). In contrast, LVM-AI trained using an unweighted loss function had slightly higher correlation (r=0.82 [95% CI, 0.81–0.83]) but substantially worse LVH diagnostic performance (c-statistic 0.528 [95% CI, 0.505–0.552]). Performance of the secondary LVM-AI models is summarized in Figures XI and XII in the Data Supplement.