December 16, 2024Open Access
International Validation of Echocardiographic AI Amyloid Detection Algorithm
Key Result
EchoNet-LVH accurately detected cardiac amyloidosis from echocardiogram videos with an AUC of 0.896 and a specificity of 0.988 across multiple international healthcare systems.
Study Design
Type
Case-Control (n=1,423)
Multicenter
Yes
Structured PICO
Does the EchoNet-LVH deep learning algorithm accurately detect cardiac amyloidosis on echocardiograms?
P
Population
Cardiac amyloidosis patients and controls from multiple international sites
I
Intervention
EchoNet-LVH (computer vision deep learning algorithm based on parasternal long axis and apical-4-chamber view videos)
C
Comparator
Controls (echocardiogram studies of non-cardiac amyloidosis patients)
O
Outcome
Discrimination performance measured by area under the receiver operating characteristic curve (AUC)surrogate
The EchoNet-LVH deep learning algorithm demonstrates high specificity and positive predictive value for detecting cardiac amyloidosis from standard echocardiogram views.
Main Result
Effect estimate: AUC 0.896 (95% CI 0.875 - 0.916)
Limitations
- Prospective trialing of CA screening approaches has not yet completed.
- Incomplete data on the true population prevalence of CA, which significantly impacts the positive predictive value.
- Significant observer variability in most echocardiographic measurements.
Abstract
Background: Diagnosis of cardiac amyloidosis (CA) is often missed or delayed due to confusion with other causes of increased left ventricular wall thickness. Conventional transthoracic echocardiographic measurements like global longitudinal strain (GLS) has shown promise in distinguishing CA, but with limited specificity. We conducted a study to investigate the performance of a computer vision detection algorithm in across multiple international sites. Methods: EchoNet-LVH is a computer vision deep learning algorithm for the detection of cardiac amyloidosis based on parasternal long axis and apical-4-chamber view videos. We conducted a multi-site retrospective case-control study evaluating EchoNet-LVH's ability to distinguish between the echocardiogram studies of CA patients and controls. We reported discrimination performance with area under the receiver operating characteristic curve (AUC) and associated sensitivity, specificity, and positive predictive value at the pre-specified threshold. Results: EchoNet-LVH had an AUC of 0.896 (95% CI 0.875 - 0.916). At pre-specified model threshold, EchoNet-LVH had a sensitivity of 0.644 (95% CI 0.601 - 0.685), specificity of 0.988 (0.978 - 0.994), positive predictive value of 0.968 (95% CI 0.944 - 0.984), and negative predictive value of 0.828 (95% CI 0.804 - 0.850). There was minimal heterogeneity in performance by site, race, sex, age, BMI, CA subtype, or ultrasound manufacturer. Conclusion: EchoNet-LVH can assist with earlier and accurate diagnosis of CA. As CA is a rare disease, EchoNet-LVH is highly specific in order to maximize positive predictive value. Further work will assess whether early diagnosis results in earlier initiation of treatment in this underserved population.