A deep learning phenome wide association study of the electrocardiogram

Abstract

Aims: Deep learning methods have shown impressive performance in detecting a range of diseases from electrocardiogram (ECG) waveforms, but the breadth of diseases that can be detected with high accuracy remains unknown, and in many cases the changes to the ECG allowing these classifications are also opaque. In this study, we aim to determine the full set of cardiac and non-cardiac conditions detectable from the ECG and to understand which ECG features contribute to the disease classification.

Methods and results: Using large datasets of ECGs and connected electronic health records from two separate medical centres, we independently trained PheWASNet, a multi-task deep learning model, to detect 1243 different disease phenotypes from the raw ECG waveform. We confirmed that the ECG can be used to detect chronic kidney disease (AUC = 0.80), cirrhosis (AUC = 0.80), and sepsis (AUC = 0.84), as well as a range of cardiac diseases, and also found new detectable conditions, including respiratory failure (AUC = 0.86), neutropenia (AUC = 0.83), and menstrual disorders (AUC = 0.84). We found that of the 37 non-cardiac strongly detectable conditions, 35 were detectable by the model output for just four diseases, suggesting that they have similar effects on the ECG. We found that high performance in some conditions including neutropenia, respiratory failure, and sepsis can be explained by linear models based on conventional measurements taken from the ECG.

Conclusion: Our study uncovers a range of diseases detectable in the ECG, including many previously unknown phenotypes, and makes progress towards understanding ECG features that allow this detection.

Keywords: Artificial intelligence; Disease screening; Electrocardiogram; Phewas.

Graphical Abstract

Figure 1

Comparison of PheWASNet to baselines. Colours represent categories. All comparisons are performed at Stanford unless noted. (A) Stanford vs. Cedars-Sinai performance. (B–C) Comparison to models based on 555 measurements from the ECG. (D) Comparison to age as a univariate predictor. (E) Comparison to a linear model using age, sex, race/ethnicity, and BMI. (F) Comparison to SEER as a univariate predictor. (G–I) Comparison to three ECG measurements as univariate predictors. (J) Comparison between 12 and 1-lead PheWASNet models. (K) Comparison to a PheWASNet model with randomly initialized weights.

Figure 2

(A) Visualisation of statistical significance of model performance across 1243 phenotypes in the Stanford cohort. Colours represent categories and areas of dots represent the number of examples in the test set. The bold line marks the cutoff for a false discovery rate of 1%. (B) AUCs in the Stanford cohort. The red line marks the cutoff for an effect size cutoff of 0.80. (C–D) The same for the Cedars-Sinai cohort.