The Digital Diagnostician: Deep Learning to Detect HCM vs. Amyloidosis

As novel therapeutic agents such as Mavacamten and Tafamidis are developed and become clinically useful for patients with hypertrophic cardiomyopathy (HCM) and cardiac amyloidosis respectively (1-2), the identification of these entities becomes ever more important.

For these novel therapies to be optimally useful, clinicians must be able to detect cardiac hypertrophy and uncover its underlying etiology quickly and accurately. LV hypertrophy is widely thought to be underdiagnosed, and thus a robust diagnostic tool would help get these new drugs into the hands of patients who need them.

Researchers at Stanford and Cedars-Sinai developed a deep-learning algorithm that may be able to detect increases in left ventricular (LV) thickness and classify its underlying etiology, be it hypertrophic cardiomyopathy (HCM) or cardiac amyloidosis (3).

The authors developed a two-step deep-learning workflow: the first step involved a beat-to-beat assessment of LV dimensions using a 2-dimensional convolutional neural network, and then the second step used a spatiotemporal convolutional model to assess the underlying cause of hypertrophy.

The input data came from over 12,000 clinician-curated echocardiograms from 2008 to 2020 compiled by the Stanford Amyloid Center and the Cedars-Sinai Advanced Heart Disease Clinic. For each echocardiogram, the research team selected the parasternal long-axis (PLAX) and apical 4-chamber 2-dimensional views. The model “learned” by reading intraventricular septum (IVS), LV internal dimension (LVID), and LV posterior wall (LVPW) measurements that had been entered by a human clinician for 9600 images. For disease detection, the model was trained against LVH-matched negative controls. The model was then validated and tested on the remaining images that were not used during the “learning” process.

After some refinement, the model was able to accurately predict LV dimensions and underlying etiologies. When tested on an external dataset of domestic data, the overall R2 of the model was 0.96, and the mean absolute error (MAE) was 1.7 mm for IVS thickness, 3.8 mm for LVID, and 1.8 mm for LVPW thickness. On the same dataset, the model was able to accurately diagnose HCM vs. cardiac amyloidosis. For amyloidosis, the AUC of the ROC plot was 0.79, and for HCM it was 0.89. Additionally, the model seems to be broadly applicable as it performed well on international datasets and across various BMIs and image qualities.

The model built in this study, EchoNet-LVH, provides a glimpse into how a symbiotic relationship between clinician and computer might improve diagnostics in the near future. As to whether this relationship will have a measurable impact on patient outcomes, a clinical trial is now underway to answer that question.

Duffy G, Cheng PP, Yuan N, et al. High-Throughput Precision Phenotyping of Left Ventricular Hypertrophy With Cardiovascular Deep Learning. JAMA Cardiol. Published online February 23, 2022. doi:10.1001/jamacardio.2021.6059

References.

  1. Spertus, J. A., Fine, J. T., Elliott, P., Ho, C. Y., Olivotto, I., Saberi, S., Li, W., Dolan, C., Reaney, M., Sehnert, A. J., & Jacoby, D. (2021). Mavacamten for treatment of symptomatic obstructive hypertrophic cardiomyopathy (EXPLORER-HCM): health status analysis of a randomised, double-blind, placebo-controlled, phase 3 trial. The Lancet, 397(10293), 2467–2475. https://doi.org/10.1016/S0140-6736(21)00763-7
  2. Maurer MS, Schwartz JH, Gundapaneni B, Elliott PM, Merlini G, Waddington-Cruz M, Kristen AV, Grogan M, Witteles R, Damy T, Drachman BM, Shah SJ, Hanna M, Judge DP, Barsdorf AI, Huber P, Patterson TA, Riley S, Schumacher J, Stewart M, Sultan MB, Rapezzi C; ATTR-ACT Study Investigators. Tafamidis Treatment for Patients with Transthyretin Amyloid Cardiomyopathy. N Engl J Med. 2018 Sep 13;379(11):1007-1016. doi: 10.1056/NEJMoa1805689. Epub 2018 Aug 27. PMID: 30145929.
  3. Duffy G, Cheng PP, Yuan N, et al. High-Throughput Precision Phenotyping of Left Ventricular Hypertrophy With Cardiovascular Deep Learning. JAMA Cardiol. Published online February 23, 2022. doi:10.1001/jamacardio.2021.6059