Identifying heart failure dynamics using multi-point electrocardiograms and deep learning

Figures

• 

  Figure 1. Patient selection and data distribution. A total of 32 045 electrocardiogram–brain natriuretic peptide pairs from 6892 patients were initially identified. Patients were included if they had at least 2 paired 12-lead electrocardiogram and brain natriuretic peptide tests performed on the same day on separate occasions. Exclusion criteria included sacubitril/valsartan use and history of sepsis, hyperthyroidism, liver cirrhosis, end-stage renal disease, pulmonary embolism, or morbid obesity (body mass index >35). After exclusions, the remaining data were randomly allocated to training (70%), validation (15%), and test (15%) datasets.

• 

  Figure 2. Deep learning models for electrocardiogram analysis and heart failure progression. (A) Electrocardiogram-to-brain natriuretic peptide transformer: estimates brain natriuretic peptide levels from a single 12-lead electrocardiogram. The model processes electrocardiogram signals through a convolutional neural network and transformer architecture to predict brain natriuretic peptide values. (B) Dual-electrocardiogram heart failure trajectory model: classifies heart failure progression using paired baseline and follow-up electrocardiograms. It analyses changes between two time points to directly categorize patients into either improved (≥40% brain natriuretic peptide decrease, baseline ≥ 100 pg/mL), deteriorated (≥40% brain natriuretic peptide increase, follow-up ≥ 100 pg/mL), or no-change groups.

• 

  Figure 3. Prediction of brain natriuretic peptide with the electrocardiogram-to-brain natriuretic peptide transformer. The figure shows a scatter plot with the normalized logarithms of brain natriuretic peptide values derived from the deep learning model on the x-axis and the actual normalized logarithm of brain natriuretic peptide values obtained from blood tests on the y-axis. A substantial correlation was observed between the two variables, with a Pearson correlation coefficient of R = 0.766 (95% confidence interval: 0.740–0.791). The mean absolute error was calculated as 0.688 (95% confidence interval: 0.655–0.722).

• 

  Figure 4. Dual-electrocardiogram heart failure trajectory model performance for heart failure status classification. Receiver operating characteristic curves of the dual-electrocardiogram heart failure trajectory model for classifying heart failure status changes. The model’s area under the receiver operating characteristic curve for each class is: deteriorated (0.886, 95% confidence interval: 0.873–0.900), no-change (0.887, 95% CI: 0.877–0.897), and improved (0.903, 95% confidence interval: 0.891–0.916). The model uses paired electrocardiogram data to classify heart failure progression.

• 

  Figure 5. Diagnostic basis visualization: representative cases. (5A) Twelve-lead baseline and follow-up electrocardiograms of the same patient, with highlighted areas denoting segments that contribute to the detection of a deteriorated class. The model was strongly activated by the QRS complex in Lead V5, as well as by the P wave in Lead V4. (5B) Twelve-lead baseline and follow-up electrocardiograms of the same patient with an applied attention map. The highlighted areas represent the segments that activated the model towards the prediction of an improved class. The model was particularly activated by increased R-wave amplitude in the pre-cordial leads.

• 

  Figure 6. Diagnostic basis visualization: averaged attention mapping. (A) The averaged attention map for the deteriorated class. (B) The averaged attention map for the improved class. We visualized the averaged attention maps for all samples within each class of the test dataset. Given the variability in heart rate among the electrocardiograms, we standardized the RR intervals to be 1 s consistently and adjusted the lengths before averaging the attention map values for each of the 12 leads. The electrocardiograms overlaid on the attention maps are not the average electrocardiograms of all samples. Instead, they are representative electrocardiograms selected specifically to clearly illustrate the positions of the P waves, QRS complexes, and T waves.