EchoAgent: a vision–language ai model for guideline-based detection and grading of left ventricular diastolic dysfunction from echocardiographic images

Background Assessment of left ventricular diastolic dysfunction (DD) by echocardiography is essential for the diagnosis of heart failure with preserved ejection fraction and other cardiovascular conditions. However, interpretation remains complex, requiring integration of multiple parameters according to current European Society of Cardiology and American Society of Echocardiography / European Association of Cardiovascular Imaging recommendations. Interobserver variability and limited access to expert echocardiographers may hinder standardized assessment, particularly in resource-limited settings. Purpose To develop and evaluate a vision–language artificial intelligence model capable of detecting and grading diastolic dysfunction directly from echocardiographic images according to contemporary guideline criteria. Methods EchoAgent was developed using the vision–language model Qwen2.5-VL (32B parameters). Fine-tuning was performed using Low-Rank Adaptation (LoRA) (rank 128, α=32, dropout=0.05) on NVIDIA A100 GPUs for 20 epochs (learning rate 3×10⁻⁵; AdamW 8-bit optimizer). The dataset consisted of 450 echocardiographic studies acquired at our center, each including four key diagnostic views relevant to diastolic assessment. Ground-truth grading was determined by expert cardiologists according to ASE/EACVI guideline criteria. Data were split into 70% training and 30% validation sets. Results The model successfully identified echocardiographic features relevant to diastolic function and classified diastolic dysfunction grades with an accuracy ranging from 78% to 82% in the validation cohort. Performance demonstrated consistent agreement with expert interpretation across standard views, suggesting the feasibility of automated guideline-based diastolic assessment from routine echocardiographic images. Agreement with expert cardiologists showed a Cohen’s κ of 0.71. In the validation cohort, the model classified diastolic dysfunction grades with an accuracy of 78–82%, with a sensitivity of 84% and specificity of 79% for detecting diastolic dysfunction. Conclusion A vision–language AI model can detect and grade diastolic dysfunction directly from echocardiographic images with promising accuracy. Such systems may support standardized echocardiographic interpretation and extend advanced cardiac imaging expertise to settings with limited access to experienced echocardiographers.