ASN's Mission

ASN leads the fight to prevent, treat, and cure kidney diseases throughout the world by educating health professionals and scientists, advancing research and innovation, communicating new knowledge, and advocating for the highest quality care for patients.

learn more

Contact ASN

1401 H St, NW, Ste 900, Washington, DC 20005

email@asn-online.org

202-640-4660

The Latest on Twitter

Kidney Week

Abstract: PO0855

Prediction of Left Ventricular Function Using Electrocardiogram Data in Patients on Hemodialysis

Session Information

Category: Dialysis

  • 701 Dialysis: Hemodialysis and Frequent Dialysis

Authors

  • Vaid, Akhil, Icahn School of Medicine at Mount Sinai, New York, New York, United States
  • Charytan, David M., NYU Langone Health, New York, New York, United States
  • Chan, Lili, Icahn School of Medicine at Mount Sinai, New York, New York, United States
  • Nadkarni, Girish N., Icahn School of Medicine at Mount Sinai, New York, New York, United States
Background

Left ventricular (LV) systolic dysfunction is common in patients on maintenance hemodialysis (HD). Early identification of patients with depressed left ventricular ejection fraction (LVEF) can facilitate disease modifying treatment. Electrocardiograms (ECGs) are routinely performed in patients on HD, however they have not been used for estimating LVEF in this population.

Methods

We analyzed data from five Mount Sinai facilities. Patients on HD with a transthoracic echocardiogram within 7 days of an ECG were identified using diagnostic and procedure codes. ECG data were preprocessed to remove recording artifacts, plotted to an image, and along with patient demographics were analyzed using a model comprised of a Multi-Layer Perceptron and a Convolutional Neural Network. We developed three models; 1) trained from scratch in only HD patients, 2) pre-trained on natural images (Imagenet), and 3) pre-trained on all LVEF:ECG pairs (n=696,890) excluding those for ESRD patients. Models 2 and 3 leverage transfer learning, which reuses knowledge gained from a task to perform a similar task. All models were trained/tested on LVEF:ECG pairs for ESRD patients within a Group Stratified K Fold (K=5) Cross Validation design, and performance was compared per Area Under Receiver Operating Characteristic curve (AUROC) for each category of LVEF, ≤40%, 41 to ≤50%, and >50%.

Results

We extracted 18,626 LVEF:ECG pairs for 2,168 ESRD patients. For detection of LVEF ≤ 40%, models trained from scratch and pre-trained on Imagenet had AUROCs of 0.74 (95% CI: 0.67-0.80) and 0.71 (95% CI: 0.65-0.77) respectively. These were outperformed by the model pre-trained on ECG data [AUROC of 0.91 (95% CI: 0.88-0.93)]. Similar results were seen at detection of LVEF 41 to ≤50% with the AUROC being 0.55 (95% CI: 0.49-0.6) for both the model trained from scratch and the Imagenet model, while the model pre-trained on ECG data achieved an AUROC of 0.82 (95% CI: 0.78-0.87).

Conclusion

A model pre-trained on non-HD LVEF:ECG pairs using transfer learning consistently outperformed models trained from scratch or pre-trained on Imagenet. This model can facilitate identification of LV systolic dysfunction in patients on HD.

ROC curves