Read our paper describing this work here. If you use our model, please cite the paper directly:
Schlesinger, Daphne E., et al. "A Deep Learning Model for Inferring Elevated Pulmonary Capillary Wedge Pressures From the 12-Lead Electrocardiogram." JACC: Advances 1.1 (2022): 100003.
RHCnet takes in the 12-lead, 10 second ECG, sampled at 500 Hz. It generates the following inferences of hemodynamic values:
- p(PCWP > 15 mmHg)
- p(mPAP > 20 mmHg)
- p(PVR > 3 Wood's Units)
- p(CO > 4 L/min)
We pre-train on a much larger cohort of ECGs, then fine-tune the model for the downstream task of interest: identifying abnormal hemodynamic values.
We include an example ECG from PhysioNet1,2 in the example notebook. The input leads should be in the following order: I, II, III, aVR, aVL, aVF, V1-V6. Each ECG is normalized by its mean and variance across all samples. An input array should be of shape Nx5000x12 for N ECGs.
The output will be of shape Nx4, where the columns are ordered mPAP, PCWP, PVR, CO.
We are most interested in positive and negative predictive values (PPV and NPV respectively). We compute these quantities using a threshold that produces a sensitivity of 80 percent, where the threshold must be computed separately for each bootstrap. These thresholds are 0.46±0.06 for the PCWP task, and 0.52±0.02 for the mPAP task. The PPVs and NPVs, computed across bootstraps, are included in the table below.
| mPCWP | mPAP | |
|---|---|---|
| PPV | 0.67±0.03 | 0.83±0.02 |
| NPV | 0.78±0.01 | 0.63±0.02 |
In addition, we evaluate the unreliability score by examining the 10% most unreliable predictions, as compared to the 90% least unreliable. Across bootstraps, predictions with an unreliability greater than 0.15±0.03 were among the 10% most unreliable predictions.
- numpy
- pandas
- matplotlib
- tensorflow 2.0 or greater
- tensorflow-addons
See the package file for an exhaustive list.
[1] Bousseljot R, Kreiseler D, Schnabel, A. Nutzung der EKG-Signaldatenbank CARDIODAT der PTB über das Internet. Biomedizinische Technik, Band 40, Ergänzungsband 1 (1995) S 317
[2] Goldberger AL, Amaral LAN, Glass L, Hausdorff JM, Ivanov PCh, Mark RG, Mietus JE, Moody GB, Peng C-K, Stanley HE. PhysioBank, PhysioToolkit, and PhysioNet: Components of a New Research Resource for Complex Physiologic Signals. Circulation 101(23):e215-e220 [Circulation Electronic Pages; http://circ.ahajournals.org/content/101/23/e215.full]; 2000 (June 13).