Understanding EasyEF research results

01

Clips and participants are different units

The 2024 study included 923 clips, of which 184 were tested. Rayong studied 218 participants. A clip does not necessarily equal one person, so these counts cannot be added as a patient total.

02

Learning data and test data

The 2024 model learned from 739 training clips and was evaluated on 184 test clips. The 96.2% result comes from that test set, addressing a different question from performance on learning data.

Future evaluation needs clear data separation, including participant and site relationships, so results address the intended question.

03

Why subgroup results matter

Larger groups contribute more to the overall result. The 2024 study can therefore have 96.2% overall accuracy alongside approximately 71% for Mildly Reduced EF, just as Rayong’s 93.6% coexists with 27.3% for fair LV.

Looking at each group reveals where the system needs improvement and which outputs require particular attention.

04

Reading the category names

The 2024 study uses Reduced EF, Mildly Reduced EF, and Preserved LV; Rayong uses poor, fair, and good LV. We retain each study’s labels so the results stay tied to its methods.

Both studies identify difficulties in an intermediate category, but similar labels do not establish identical criteria or populations. Collaboration decisions should use the methods of the relevant study.

05

Compare results that answer the same question

Five-category classification is more detailed than three-category classification, so the 2024 results of 68.5% and 96.2% must be read in that context. Rayong uses a different population and data unit.

We do not average 96.2% and 93.6% or treat their difference as a performance trend, because the methods and conditions differ.

06

Know what has been evaluated and what is planned

The 2024 study is published; Rayong’s peer-review status is unverified. Network activities and awards show different dimensions of progress from clinical evaluation results.

Evaluation in at least five hospitals and SaMD-readiness work are development plans. Accuracy testing does not establish mortality or service-workload benefits.

07

Questions that help shape collaboration

Start with intended users and patients, then consider clip quality, system version, reference assessment, and success criteria. Include subgroup performance and unassessable cases.

Then define service questions such as time, workload, and report understanding. Considering both dimensions helps the evaluation meet the organization’s needs.