Machine learning-enabled screening for aortic stenosis with handheld ultrasound

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Neural network classifiers show promise in detecting aortic stenosis (AS) using cardiac ultrasound.
• Existing models are trained on cart-based imaging and have not been validated on focused cardiac ultrasound (FoCUS) from handheld devices.

Purpose of the Study:

• To evaluate the performance of a cart-based AS classifier on FoCUS images acquired with handheld ultrasound devices.
• To fine-tune the classifier using handheld data to improve AS detection accuracy.

Main Methods:

• A prospective study included 160 patients (≥65 years) undergoing transthoracic echocardiography.
• A pre-trained cart-based AS classifier was tested on FoCUS images from a handheld device.
• The classifier underwent last-layer fine-tuning on handheld data and was externally validated.

Main Results:

• The cart-based model achieved an AUROC of 0.87 on FoCUS images.
• Fine-tuning improved the AUROC to 0.94, with external validation reaching 0.97.
• The fine-tuned model demonstrated high positive and negative predictive values in simulated screening environments.

Conclusions:

• A cart-based AI model for AS detection experienced reduced performance on handheld FoCUS images.
• Fine-tuning the model with handheld data significantly enhanced its diagnostic accuracy.
• This approach shows potential for automated AS detection using handheld ultrasound devices.