Pediatric ECG-Based Deep Learning to Predict Left Ventricular Dysfunction and Remodeling

Pediatric ECG-Based Deep Learning to Predict Left Ventricular Dysfunction and Remodeling

Mayourian J, La Cava WG, Vaid A, Nadkarni GN, Ghelani SJ, Mannix R, Geva T, Dionne A, Alexander ME, Duong SQ, Triedman JK. Circulation. 2024 Mar 19;149(12):917-931. doi: 10.1161/CIRCULATIONAHA.123.067750. Epub 2024 Feb 5.PMID: 38314583

Take Home Points:

  1. Using 92,377 Boston Children’s Hospital ECG-echocardiogram pairs, an artificial intelligence-enhanced pediatric model was generated which was highly predictive of LV dysfunction and remodeling.
  2. Internal and external validation of the model, demonstrated negative predictive values of 99.0% and 99.2% respectively for the composite outcome of LV systolic dysfunction, hypertrophy or dilation.
  3. Saliency maps highlighting the regions of the ECG having the greatest and least influence on the primary outcomes of the model, allow for a visual comparison to existing criteria used when interpreting an ECG.
Dr Jeremy P Moore

Commentary by Dr. Jeremy Moore (Los Angeles, CA, USA) Congenital and Pediatric Cardiac EP chief section editor:  

Deep learning-based artificial intelligence-enhanced ECG (AI-ECG) algorithms have recently been shown to reliably predict several adult cardiovascular echocardiographic-confirmed findings such as ventricular dysfunction, hypertrophy, and dilation. This paper by Mayourian et al. sought to explore this novel technology in a large pediatric population.

The study was divided into two major components – a convolutional neural network was trained on ECG-echocardiogram pairs from patients at Boston Children’s Hospital (92,377 pairs, 46,261 patients; median age 8.2 years), and subsequently internally validated on BCH patients that were randomly partitioned and not included in the training cohort (12,631 patients, median age 8.8 years), BCH emergency department patients (2,830, median age 7.7 years), as well as on an external cohort of patients from Mount Sinai Hospital (5,088 patients, median age 4.3 years).

Children £ 18 years of age with an ECG-echocardiogram pair (£ 2 days apart) between January 1, 2002, and December 31, 2021, were included for this study. Echocardiograms performed in the medical and cardiac intensive care units, or operating rooms were excluded, as were patients with known congenital heart disease. External ECG-echocardiogram pairs for validation testing used the same exclusion criteria, but pairs £ 7 days apart were included.

The primary outcomes used to train the AI-ECG model were LV systolic dysfunction, LV hypertrophy, and LV dilation. Qualitative measures greater than mild were considered positive for these outcomes, as well as quantitative z -score cutoffs for LV EF, LV mass and LV end-diastolic volume of £ -4, ³ +4 and ³ +4 respectively. A composite outcome was considered positive if any of the three metrics was positive.

The results of this AI-ECG model were impressive. Both on internal and external validation, the negative predictive value of the model for the composite quantitative outcome was 99.0% and 99.2%. The model even outperformed the pediatric cardiologist expert ECG-based diagnosis of LV hypertrophy. See table below for a summary of the validation results:

To better understand the model, the authors performed saliency mapping and median waveform analysis. This sought to provide a better visual insight for comparison to existing criteria used when interpreting an ECG. Their findings are highlighted in this key image:

The saliency maps demonstrate the regions of the ECG having the greatest and least influence on the primary outcomes. Features which were considered high-risk for predicting LV dysfunction included T-wave inversion in the lateral precordial leads (V4-V6), and S-waves in V1-V2. High-risk salient features to predict LV hypertrophy were deep S-waves in V1-V2, or high amplitude R-wave in limb lead 1.  Lastly, for LV dilation, high-amplitude R-waves in V4-V6 were considered high risk features by the model.

Conclusion

This work highlights the possible role that artificial intelligence may play on how healthcare is delivered. The ECG is a relatively inexpensive, very accessible diagnostic test, with high variance on interpretation based on the level of the provider’s experience and expertise. The authors point out, that an AI-ECG predictive model could be used to guide the need and urgency for pediatric cardiology referrals, diminish the costs of unnecessary referrals and echocardiograms, but most importantly avoid missed diagnoses that could result in an adverse clinical outcome.

This paper by Mayourian et al. aimed to create an AI-ECG model based on the human expert interpretation of echocardiograms, hence the inclusion of both qualitative and quantitative assessments of LV size and function. They point out that qualitative classification is subject to more variability, hence the model generally performed better with quantitative cutoffs. This paper is just the start of a new technology with great potential to transform how pediatric cardiology is practiced, and potentially reduce disparities by improving access to highly predictive and affordable testing.

Pediatric Cardiac Professionals