ECG-guided non-invasive estimation of pulmonary congestion in patients with heart failure

Aniruddh Raghu, Daphne E. Schlesinger, Eugene Pomerantsev, Srikanth Devireddy, Pinak Shah, Joseph Garasic, John V. Guttag, Collin M. StultzScientific 2023 · DOI

TL;DR

Researchers have developed HFNet, a deep learning model that estimates pulmonary congestion—a critical sign of worsening heart failure—using only a standard 12-lead ECG and basic demographics. By incorporating a novel uncertainty score, the system not only predicts elevated heart pressures with high accuracy but also alerts clinicians when its own predictions are likely to be unreliable.

What Problem Does This Paper Address?

In patients with heart failure (HF), "congestion" occurs when fluid backs up into the lungs due to high pressures in the heart's left side. The clinical gold standard for measuring this is the mean Pulmonary Capillary Wedge Pressure (mPCWP).

Currently, measuring mPCWP accurately requires an invasive procedure called right heart catheterization, where a tube is threaded through the veins into the heart. While non-invasive alternatives like cardiac ultrasound (Doppler) exist, they require highly specialized technicians and aren't always available in emergency or routine settings. There is a massive clinical need for a fast, cheap, and non-invasive way to identify patients with an mPCWP > 18 mmHg (the threshold for significant congestion) to prevent hospitalizations and guide therapy.

Methods

The authors developed HFNet, a multi-modal deep learning architecture:

  • ECG Encoder: A convolutional neural network (CNN) that processes the raw 10-second, 12-lead ECG signal.
  • Demographic Encoder: A fully connected network that incorporates age and sex.
  • Uncertainty Scoring: This is a standout feature. The model calculates predictive entropy. If the model’s output is near 0.5 (toss-up), the entropy is high, signaling that the prediction is "untrustworthy."
  • Interpretability Tools: The team used saliency maps to see which parts of the ECG the model prioritized and synthetic ECG generation to systematically perturb waveforms (like P-wave amplitude) to see how the model responded.

Datasets

The study utilized two large-scale, retrospective datasets of patients with a prior diagnosis of heart failure:

  1. Development & Internal Test Set: 7,121 records from 3,767 patients at Massachusetts General Hospital (Hospital 1).
  2. External Validation Set: 2,725 records from 1,249 patients at Brigham and Women’s Hospital (Hospital 2).

Unlike many AI studies that use "clean" data, these datasets represent real-world clinical environments where ECGs and invasive measurements were taken on the same day.

Key Insights and Ideas

  • The "Trust" Factor: The model doesn't just give a "yes/no" answer; it provides a gauge of its own confidence. In the lowest-uncertainty group, the model’s accuracy was significantly higher (AUROC 0.86) than in the high-uncertainty group (AUROC 0.50, essentially random). This allows clinicians to filter out "noisy" AI advice.
  • Physiological Alignment: Through saliency analysis, the researchers found that HFNet focuses primarily on the diastolic portion of the cardiac cycle and the P-wave. This is a major "aha moment" because it aligns perfectly with cardiology theory: mPCWP is a reflection of left atrial pressure at the end of diastole.
  • Dynamic Tracking: For patients who had multiple catheterizations over months or years, the model was able to track the trend of their congestion, accurately reflecting when a patient was getting better or worse.
  • Synergy with Chest X-rays: The model provides the most value when Chest X-rays (CXR) are ambiguous. Even if a CXR doesn't show fluid, a positive HFNet prediction significantly raises the probability that the patient is actually congested.

Key Results

  • Overall Performance: HFNet achieved an AUROC of 0.82 on internal data and 0.81 on external validation, significantly outperforming a baseline logistic regression model (AUROC 0.67–0.71).
  • High Confidence Performance: When looking only at the most certain predictions (bottom 75% of entropy), the AUROC improved to 0.86–0.87.
  • Clinical Utility: Using a threshold for 80% sensitivity, the model achieved a Positive Predictive Value (PPV) of 0.89 when combined with X-ray findings of edema, and a PPV of 0.78 even when the X-ray appeared clear.
  • Temporal Accuracy: In patients with multiple measurements, the model reproduced hemodynamic trends with ~79% average accuracy.

Limitations and Future Work

While promising, the authors note several hurdles:

  • Binary Classification: The model currently only predicts if mPCWP is above or below 18 mmHg. It cannot yet provide a precise numerical pressure value (e.g., "22 mmHg").
  • Timing Gaps: Because the data was retrospective, the ECG and the catheterization happened on the same day but not necessarily at the exact same minute. Capturing these simultaneously could improve training.
  • Arrhythmias: The dataset had a relatively low prevalence of atrial fibrillation (<10%). Future work needs to validate if the model holds up in patients with highly irregular heart rhythms.
  • Population Diversity: The study relied on specific prevalence estimates for patients with reduced ejection fraction; further prospective studies are needed to see how it generalizes to broader, more diverse patient populations.

Summary updated

Abstract

Quantifying hemodynamic severity in patients with heart failure (HF) is an integral part of clinical care. A key indicator of hemodynamic severity is the mean Pulmonary Capillary Wedge Pressure (mPCWP), which is ideally measured invasively. Accurate non-invasive estimates of the mPCWP in patients with heart failure would help identify individuals at the greatest risk of a HF exacerbation. We developed a deep learning model, HFNet, that uses the 12-lead electrocardiogram (ECG) together with age and sex to identify when the mPCWP > 18 mmHg in patients who have a prior diagnosis of HF. The model was developed using retrospective data from the Massachusetts General Hospital and evaluated on both an internal test set and an independent external validation set, from another institution. We developed an uncertainty score that identifies when model performance is likely to be poor, thereby helping clinicians gauge when to trust a given model prediction. HFNet AUROC for the task of estimating mPCWP > 18 mmHg was 0.8 [Formula: see text] 0.01 and 0.[Formula: see text] 0.01 on the internal and external datasets, respectively. The AUROC on predictions with the highest uncertainty are 0.50 [Formula: see text] 0.02 (internal) and 0.[Formula: see text] 0.04 (external), while the AUROC on predictions with the lowest uncertainty were 0.86 ± 0.01 (internal) and 0.82 ± 0.01 (external). Using estimates of the prevalence of mPCWP > 18 mmHg in patients with reduced ventricular function, and a decision threshold corresponding to an 80% sensitivity, the calculated positive predictive value (PPV) is 0.[Formula: see text] 0.01when the corresponding chest x-ray (CXR) is consistent with interstitial edema HF. When the CXR is not consistent with interstitial edema, the estimated PPV is 0.[Formula: see text] 0.02, again at an 80% sensitivity threshold. HFNet can accurately predict elevated mPCWP in patients with HF using the 12-lead ECG and age/sex. The method also identifies cohorts in which the model is more/less likely to produce accurate outputs.
BibTeX
@inproceedings{raghu2023ecg,
  author = {Aniruddh Raghu and Daphne E. Schlesinger and Eugene Pomerantsev and Srikanth Devireddy and Pinak Shah and Joseph Garasic and John V. Guttag and Collin M. Stultz},
  title = {{ECG-guided non-invasive estimation of pulmonary congestion in patients with heart failure}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-30900-9},
}
Figure