Artificial Intelligence-Enabled Cardiac Function Estimation from Phone Videos of Echocardiograms

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Mobile phone echocardiogram videos are increasingly used in clinical settings.
• Existing AI models for LVEF estimation are typically trained on native DICOM videos.
• Performance of AI models on mobile phone videos requires evaluation.

Purpose of the Study:

• To assess the performance of AI models for LVEF estimation using mobile phone-recorded echocardiographic videos.
• To determine if AI models retain accuracy when applied to non-native video formats.
• To evaluate the impact of fine-tuning AI models with mobile phone video data.

Main Methods:

• A multicenter retrospective study involving 6209 mobile phone-recorded echocardiographic videos from 2648 studies.
• Comparison of AI-estimated LVEF with clinician-reported LVEF.
• Fine-tuning of pretrained AI models using a dedicated training cohort of mobile phone videos.

Main Results:

• Without fine-tuning, the AI model achieved a mean absolute error (MAE) of 7.00 percentage points on phone videos, compared to 6.08 on DICOM.
• Fine-tuning the AI model with phone videos improved performance, reducing MAE to 6.96 percentage points.
• Progressive central zoom preprocessing negatively impacted model performance.

Conclusions:

• AI-assisted LVEF estimation from mobile phone echocardiograms is feasible, especially when native DICOM export is not possible.
• Fine-tuning AI models enhances their accuracy for mobile phone video analysis.
• Further prospective evaluation is recommended before widespread clinical implementation.