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

Area of Science:

• Cardiology
• Medical Informatics
• Artificial Intelligence in Medicine

Background:

• Central hemodynamic parameters are typically measured invasively via pulmonary artery catheterization, which carries risks and is not universally available.
• Elevated mean pulmonary capillary wedge pressure (mPCWP) is a critical indicator in various cardiovascular conditions.

Purpose of the Study:

• To develop a noninvasive method for identifying elevated mPCWP using the 12-lead electrocardiogram (ECG).
• To leverage deep learning to infer mPCWP from ECG data, providing a potential alternative to invasive measurements.

Main Methods:

• A deep learning model was developed using 248,955 clinical records from Massachusetts General Hospital.
• The model was trained to infer mPCWP >15 mmHg from ECGs, with a subset of data used for pre-training and direct mPCWP measurements for model development and validation.
• An unreliability score was developed to quantify the trustworthiness of the model's predictions.

Main Results:

• The model achieved an area under the receiver operating characteristic curve (AUC) of 0.80 ± 0.02 on the test set and 0.79 ± 0.01 on the holdout set.
• Model performance was dependent on the unreliability score, with higher scores indicating poorer performance (e.g., AUC of 0.70 ± 0.06 in the highest decile).

Conclusions:

• Mean pulmonary capillary wedge pressure (mPCWP) can be inferred noninvasively from ECGs using deep learning.
• The reliability of these inferences can be quantified, offering valuable clinical information when invasive monitoring is not readily available or feasible.