Document resource
Introduction Left ventricular hypertrophy (LVH) is an established, independent predictor for cardiovascular disease. Indices derived from the electrocardiogram (ECG) have been used to infer the presence of LVH with limited sensitivity. We previously compared supervised machine learning techniques to classify cardiac magnetic resonance (CMR) derived indexed left ventricular mass (iLVM) using ECG biomarkers in 37,534 UK Biobank (UKB) participants with area under the receiver operator curve (AUC) (0.89 [95% CI, 0.87–0.92]). In this study we explored agnostic approaches to improve classification and developed a deep learning (DL) model with external validation in the Study of Health in Pomerania (SHIP).Methods We analysed the median waveform across 8 independent ECG leads (I, II, V1–6), derived from the raw 15 second signals of 48,835 participants from the UKB imaging study. The dataset was split into a training set (70%), validation set (15%) and hold-out test set (15%) for performance evaluation. The model architecture was a modified 34-layer residual network, for which the input was a participant’s median ECG and clinical variables and the predicted iLVM as the output. A subsequent logistic regression model used iLVM predictions and participant sex as independent variables to account for systematic underestimations. The model training procedure included an early stopping criterion to limit overfitting. External validation was sought in 1,423 participants in SHIP. To assist in generalisation, the model was fine-tuned on 75% of the SHIP cohort with the remainder used for performance measurement. Saliency maps visualising the predictive areas of the ECG were produced using approximated SHapley Additive exPlanations (SHAP), with respect to a random baseline sample of ECGs.Results The baseline characteristics of the UKB and SHIP participants are shown in table 1. In UKB, 717 (1.5%) participants had CMR-derived LVH and 83 (5.8%) participants in SHIP. LVH classification performance is reported in table 2. AUC for the DL model was (0.97 [95% CI, 0.96–0.98]), an improvement to our previously developed supervised algorithm, and both models outperformed the current clinical ECG criteria for LVH. There was modest generalisability of the DL model in SHIP (0.77 [95% CI, 0.66–0.89]) (figure 1). Sex, age, ventricular rate and systolic blood pressure were among the most predictive clinical features. Saliency maps of the ECG waveforms using SHAP are shown in figure 2. Broadly, the ECG components most relevant for LVH is the QRS complex.Abstract 6-015 Table 1Baseline characteristics of the UKB and SHIP participants UKB (N=48,835) SHIP (N=1,423) LVH (N=717) Normal LV LVH (N=83) Normal LV Age (years) 64 [58, 70] 65 [59, 71] 52 [45, 61] 52 [42, 62] Sex, Female (%) 353 (49.2) 24,962 (52.0) 32 (38.6) 621 (46.3) BMI ( 27.7 [24.3, 30.6] 26.6 [23.6, 28.8] 28.0 [24.8, 31.4] 27.3 [24.4, 30.0] Systolic BP (mmHg) 161 [145, 178] 144 [129, 158] 138 [125, 150] 127 [115, 138] Diastolic BP (mmHg) 87 [78, 96] 82 [74, 89] 82 [73, 90] 78 [72, 84] High cholesterol (%) 456 (63.4) 30,932 (64.3) 54 (65.1) 885 (66.0) Diabetes (%) 57 (7.9) 2681 (5.6) 7 (8.4) 66 (4.9) Indexed LVM ( 68.1 [57.9, 74.6] 44.7 [38.8, 49.8] 68.8 [59.8, 74.9] 49.3 [42.6, 55.7] Integer variables are presented as a number (percentage) and continuous variables as median [IQR]. BMI: body mass index, BP: blood pressure, LV: left ventricle, LVH: left ventricular hypertrophy, LVM: left ventricular mass, SHIP: Study of Health in Pomerania, UKB: UK Biobank.Abstract 6-015 Table 2LVH classification performance for test partitions of the UKB and SHIP cohorts UKB SHIP Deep learning (N=7,325) *Sokolow-Lyon (N=5,630) *Cornell-voltage (N=5,630) Deep learning (N=356) AUC 0.97 (0.96 - 0.98) 0.51 (0.49 - 0.53) 0.49 (0.47 - 0.52) 0.77 (0.66 - 0.89) Sensitivity 0.83 (0.79 - 0.87) 0.03 (0.01 - 0.11) 0.06 (0.01 - 0.03) 0.75 (0.54 - 0.93) Specificity 0.97 (0.97 - 0.97) 0.99 (0.98 - 0.99) 0.93 (0.92 - 0.94) 0.62 (0.57 - 0.65) F1 0.71 (0.70 - 0.72) 0.51 (0.49 - 0.53) 0.49 (0.47 - 0.51) 0.65 (0.55 - 0.74) 95% Confidence intervals (in parenthesis). SHIP: Study of Health in Pomerania, SVM: support vector machine, UKB: UK Biobank. * Sokolow-Lyon and Cornell-voltage performance was also derived in the smaller dataset, from prior published work (PMID: 37538142).Abstract 6-015 Figure 1Receiver operator characteristic (ROC) curves for LVH classification in the UKB and SHIP cohortsThe operating points for the reported sensitivity/specificity are annotated with crosses, both of which optimise Youden’s J statistic (maximal difference between true positive rate and false positive rate). SHIP: Study in Health of Pomerania, UKB: UK Biobank.Abstract 6-015 Figure 2Approximated SHAP values (integrated gradients) for two sample median ECGs from UK Biobank, with low (left) and high (right) predicted indexed LVM respectivelyColours correspond to morphology effect, red increasing predictions and blue decreasing predictions, and grey having little relative effect.Conclusions DL model combining the ECG and clinical variables was able to classify CMR derived LVH. The model had greater classification performance compared to our previously developed supervised algorithm and current clinical ECG benchmarks. There was some generalisability of the DL model to an external community population with differences in clinical characteristics and ECG acquisition as important factors.