01
Testing the image-to-classification approach
The study compared parasternal long-axis videos, a lengthwise view of the heart, with reference LVEF values from cardiologist reports. LVEF is a measure of left ventricular pumping function. The question was how well AI could classify function from clips recorded with a phone.
The work was published in the Journal of Prapokklao Hospital Clinical Medical Education Center in 2024, volume 41(2), pages 123–132. Its principal outcome is video-clip classification accuracy.
References and further reading
02
From submitted clips to a test set
Between 1 May and 31 July 2023, 1,336 clips were submitted. After screening, 923 five-second clips with reference LVEF values were included.
The data was divided into 739 training clips and 184 test clips. These counts refer to video clips, not unique patients.
| Stage | Count | Meaning |
|---|---|---|
| Submitted videos | 1,336 clips | Before screening |
| Included dataset | 923 clips | After screening |
| Training | 739 clips | Data used for learning |
| Testing | 184 clips | Denominator of the 96.2% overall result |
References and further reading
03
Which data the results cover
The study excluded unsuitable clips and certain conditions, including valvular abnormalities, pericardial disease, congenital heart disease, and atrial fibrillation.
This scope matters when applying the findings. Other conditions and image-quality settings need additional testing, including consideration of variation in reference LVEF measurement.
References and further reading
04
Evaluating five and three categories
The pipeline screens quality, preprocesses images, classifies with a neural network, and consolidates five categories into Reduced EF, Mildly Reduced EF, and Preserved LV.
Accuracy was 68.5% for five categories and 96.2% after consolidation into three. Reducing the number of categories changes the level of detail the system must distinguish.
References and further reading
05
Overall results and remaining weaknesses
Three-category classification was correct for 177 of 184 clips. Mildly Reduced EF was correct for 12 of 17 clips, approximately 71%. This small subgroup is an important focus for improvement.
Preserved LV details differ between the Thai and English abstracts. We present the consistent figures here and link to the original article for further methodological review.
| Measure | Correct / tested | Reported value |
|---|---|---|
| Overall three-category accuracy | 177 / 184 clips | 96.2% |
| Mildly Reduced EF | 12 / 17 clips | Approximately 71% |
References and further reading
06
What we learned and will build on
The findings support the feasibility of screening from smartphone recordings of ultrasound displays, while exposing the intermediate-category challenge. Further work needs more diverse data and better ways to help users understand AI outputs.
This study provides a foundation for evaluation in more settings. Effects on time, referrals, and patient outcomes remain separate research questions.
References and further reading
SOURCE NOTES
Sources and published project summaries
Project content is compiled and published by the EasyEF team from development records, presentations and study reports. Internal links lead to the team’s summaries, not additional independent evidence. Published research links to its original journal; study status and limitations are stated in the relevant articles.
Explore the source library →