A Deep Learning Model for Inferring Elevated Pulmonary Capillary Wedge Pressures From the 12-Lead Electrocardiogram

Figures

• 

  Figure 1. Training and Testing Procedure Data sets used for training and evaluating model 1, model 2, and RHCNet. The model with the best performance on the test set is also evaluated using the holdout set. ECG = electrocardiogram; RHCNet = Right Heart Catheterization Network.

• 

  Figure 2. ECG Interval Logistic Regression Model In model 1, ECG intervals, derived from the machine read of each ECG, are used to train a logistic regression model to predict the probability that cardiac filling pressures are elevated. This model achieves an AUC of 0.50 ± 0.02. AUC = area under the receiver operating characteristic curve; ECG = echocardiogram.

• 

  Figure 3. Randomly Initialized Deep Neural Network In model 2, a deep neural network is trained to detect elevated mPCWP from the 12-lead ECG signal. Initial model parameters are set to random values and then modified to maximize model performance. Convolutional layers use specialized functions to extract information about the shape of the ECG. This yields an AUC of 0.69 ± 0.03. AUC = area under the receiver operating characteristic curve; ECG = echocardiogram; mPCWP = mean pulmonary capillary wedge pressure.

• 

  Figure 4. RHCNet Pre-training, Training, and Performance on Test Data RHCNet is trained in 2 steps. First, a deep neural network model is trained to estimate ECG intervals from a large set of ECGs for which information about cardiac filling pressures was not available. Fine-tuning involves using the parameters from this model as initial values for additional training, using an independent data set of 12-lead ECGs that are matched to hemodynamic measurements. This yields an AUC of 0.80 ± 0.02 on the test set. Pre-training the model for a different task using a large data set results in a set of initial model parameters that give improved results. AUC = area under the receiver operating characteristic curve; ECG = echocardiogram; RHCNet = Right Heart Catheterization Network.

• 

  Figure 5. RHCNet Performance as a Function of Sensitivity and Prevalence (A) Specificity as a function of sensitivity; (B) PPV as a function of sensitivity (x-axis) and prevalence (5%, 10%, 20%, 30%, 40% and 50%); (C) NPV as a function of sensitivity and prevalence. NPV = negative predictive value; PPV = positive predictive value; RHCNet = Right Heart Catheterization Network.

• 

  Figure 6. RHCNet Performance as a Function of Model Unreliability (A) Normalized Brier score (average model error) for inferring an elevated mPCWP in patients with high and low unreliability in the development set (B) and AUC for the same. The unreliable subgroup is defined as the set of all predictions having unreliable scores within the top decile. For simplicity, we refer to all other predictions that are relatively reliable. ∗P < 10−2, ∗∗P < 10−3. AUROC = area under the receiver operating characteristic curve; mPCWP = mean pulmonary capillary wedge pressure; RHCNet = Right Heart Catheterization Network.

• 

  Central Illustration. Evaluating Trends in mPCWP on a Per-Patient Basis RHCNet predictions on serial ECGs for a patient in our holdout set. This patient had HF and then underwent orthotopic cardiac transplantation, as noted in the colored bar at the bottom of the figure. The x-axis represents the time from the patient’s first RHC. At each time point a 12-lead ECG was obtained and RHCNet was used to estimate the probability that the mPCWP was elevated (blue), as well as the unreliability of that prediction (black), using only the 12-lead ECG. The right-hand axis corresponds to the model output and unreliability, both inferred from a 12-lead ECG. The left-hand axis corresponds to the measured mPCWP, acquired via right heart catherization on the same day that the ECG was acquired (red line). Overall, RHCNet tracks the patient’s mPCWP and subsequent improvement after transplant. Moreover, high unreliability scores (near the threshold corresponding to the 10% highest unreliability) are more likely to be incorrect. These data suggest that RHCNet may enable serial non-invasive tracking of central hemodynamics using the 12-lead ECG. ECG = electrocardiogram; HF = heart failure; mPCWP = mean pulmonary capillary…