From ultrasound video to cardiac-assessment support

01

What the workflow needs

The workflow begins with a cardiac ultrasound examination performed by skilled personnel. The phone records the display and sends the clip, so preparation includes equipment, the operator, and the clinician who will review the result.

• An ultrasound machine that can acquire the required cardiac view, and a smartphone that can record the display clearly.
• The EasyEF app and an internet connection for clip submission and result delivery.
• Personnel trained in the workflow, with a clinician to interpret results in the patient’s context.
• An agreed evaluation scope and local data-governance arrangements before a joint evaluation begins.

02

Capture the ultrasound view

The user records the ultrasound display with a smartphone. The approach evaluated in 2024 used five-second parasternal long-axis clips, showing the heart along its long axis from beside the sternum.

Framing the complete ultrasound region, reducing reflections, and holding the phone steady help produce a suitable input.

03

Check quality before assessment

• Complete image region
• Minimise reflections
• Steady, sufficient recording

Quality screening checks for problems such as reflections, a small or incomplete ultrasound region, multiple image regions, short duration, and excessive movement.

In the 2024 study, 923 of 1,336 submitted clips remained after screening. The study results describe performance on clips meeting those inclusion criteria.

04

Prepare a consistent video input

1. 01Identify the ultrasound region
2. 02Stabilise the video
3. 03Adjust the lighting

Before classification, the pipeline identifies the ultrasound region, stabilizes the video, and adjusts lighting to reduce variation introduced by recording a screen.

The 2024 pipeline used an adapted DAMO-YOLO detector, ORB features for stabilization, and CLAHE with thresholding for lighting adjustment. Each system version needs evaluation tied to its actual pipeline and model.

05

Classify ventricular function

AI analyzes the video sequence to classify left ventricular function. The 2024 system used an adapted ResNet-101, initially predicting five categories before combining them into Reduced EF, Mildly Reduced EF, and Preserved LV.

The output is a functional category rather than a precise continuous EF measurement. Test accuracy was 68.5% for five categories and 96.2% for three, reflecting different levels of classification detail.

06

Return the result to the care team

AI output + clinical context

The clinician brings the information together to make care decisions.

The app submits the clip over the internet to the processing service and displays the returned assessment. The clinician combines the AI output with symptoms, history, and other findings when considering care or further assessment.

The next platform phase connects submission, processing, and a Dashboard / Clinical Report for care teams. Timing evaluation will cover the full journey from recording to result delivery, including network and device conditions.

07

Questions before a service evaluation

• Does it work with every machine? Recording the display reduces reliance on a direct machine connection, but each equipment setup still needs image-quality and compatibility testing.
• Does the phone scan the heart? The ultrasound machine produces the image. The phone records and submits it.
• What does training cover? The planned 1–2-hour session covers the app workflow and clip capture. Cardiac ultrasound acquisition skills require training and supervision appropriate to the user’s role.

08

Extending devices and research questions

Next steps · Development and planned research

We are exploring handheld ultrasound integration and an Echo AI Box concept for existing machines, to study acquisition methods suited to different clinical settings.

Future research questions include valve and pericardial abnormalities. These capabilities need separate development and testing beyond the functional classification already studied.