David Ouyang, MD
@davidouyang.bsky.social
Cardiologist, Data Scientist, AI researcher
EchoPrime is a contrastive vision-language model. Trained on over 12M video-repprt pairs, it learns the relationship between what cardiologists write and what the ultrasound images show.
If you feed it a video or echo reports, it can find similar historical studies that match those concepts.
If you feed it a video or echo reports, it can find similar historical studies that match those concepts.
November 17, 2025 at 2:41 AM
EchoPrime is a contrastive vision-language model. Trained on over 12M video-repprt pairs, it learns the relationship between what cardiologists write and what the ultrasound images show.
If you feed it a video or echo reports, it can find similar historical studies that match those concepts.
If you feed it a video or echo reports, it can find similar historical studies that match those concepts.
EchoPrime is the culmination of a long roadmap for AI echo.
Substantial time and effort was put into curating the largest existing echocardiography database.
EchoPrime is trained on 1000x the data of EchoNet-Dynamic, our first model for AI-LVEF, and 10x the data of existing AI echo models.
Substantial time and effort was put into curating the largest existing echocardiography database.
EchoPrime is trained on 1000x the data of EchoNet-Dynamic, our first model for AI-LVEF, and 10x the data of existing AI echo models.
November 17, 2025 at 2:39 AM
EchoPrime is the culmination of a long roadmap for AI echo.
Substantial time and effort was put into curating the largest existing echocardiography database.
EchoPrime is trained on 1000x the data of EchoNet-Dynamic, our first model for AI-LVEF, and 10x the data of existing AI echo models.
Substantial time and effort was put into curating the largest existing echocardiography database.
EchoPrime is trained on 1000x the data of EchoNet-Dynamic, our first model for AI-LVEF, and 10x the data of existing AI echo models.
We are excited to announce EchoPrime is published in Nature. EchoPrime is the first echocardiography AI model capable of evaluating a full transthoracic echocardiogram study, identify the most relevant videos, and produce a comprehensive interpretation!
November 17, 2025 at 2:38 AM
We are excited to announce EchoPrime is published in Nature. EchoPrime is the first echocardiography AI model capable of evaluating a full transthoracic echocardiogram study, identify the most relevant videos, and produce a comprehensive interpretation!
Congratulations to Sarnoff Fellow Victoria Yuan on her new preprint 'Automated Deep Learning Pipeline for Characterizing Left Ventricular Diastolic Function'!
May 4, 2025 at 4:12 AM
Congratulations to Sarnoff Fellow Victoria Yuan on her new preprint 'Automated Deep Learning Pipeline for Characterizing Left Ventricular Diastolic Function'!
3/n Without additional prompting, our #echofirst model naturally gravitates towards the vena contracta in color Doppler videos to assess AR severity.
March 25, 2025 at 10:08 PM
3/n Without additional prompting, our #echofirst model naturally gravitates towards the vena contracta in color Doppler videos to assess AR severity.
2/n Accurate assessment of AR severity and sequalae is critical for surveillance and timing of surgery. Using over 40k studies from @smidtheart.bsky.social and validated on 1.5k studies from @stanforddeptmed.bsky.social, we show that AI can accurately assess AR severity across a range of patients.
March 25, 2025 at 10:08 PM
2/n Accurate assessment of AR severity and sequalae is critical for surveillance and timing of surgery. Using over 40k studies from @smidtheart.bsky.social and validated on 1.5k studies from @stanforddeptmed.bsky.social, we show that AI can accurately assess AR severity across a range of patients.
1/n Excited to announce work by Dr. @BinderRodriguez
from @MedUni_Wien on the AI automated assessment of Aortic Regurgitation (AR) on #echofirst using 59,500 videos from @smidtheart.bsky.social. Lead by Dr. Bob Siegel.
from @MedUni_Wien on the AI automated assessment of Aortic Regurgitation (AR) on #echofirst using 59,500 videos from @smidtheart.bsky.social. Lead by Dr. Bob Siegel.
March 25, 2025 at 10:07 PM
1/n Excited to announce work by Dr. @BinderRodriguez
from @MedUni_Wien on the AI automated assessment of Aortic Regurgitation (AR) on #echofirst using 59,500 videos from @smidtheart.bsky.social. Lead by Dr. Bob Siegel.
from @MedUni_Wien on the AI automated assessment of Aortic Regurgitation (AR) on #echofirst using 59,500 videos from @smidtheart.bsky.social. Lead by Dr. Bob Siegel.
March 20, 2025 at 10:35 PM
3/n Across a wide range of image quality, patient characteristics, and study types, EchoNet-Measurements perform well, improving upon sonographers' and cardiologists' precision.
March 20, 2025 at 10:34 PM
3/n Across a wide range of image quality, patient characteristics, and study types, EchoNet-Measurements perform well, improving upon sonographers' and cardiologists' precision.
2/n Lead by Yuki Sahashi, EchoNet-Measurements provide automated annotations for the 18 most common #echofirst (both B mode and Doppler) measurements.
Excellent performance in test cohorts (overall R2 of 0.967 in the held-out CSMC dataset and 0.987 in the SHC dataset)
Excellent performance in test cohorts (overall R2 of 0.967 in the held-out CSMC dataset and 0.987 in the SHC dataset)
March 20, 2025 at 10:33 PM
2/n Lead by Yuki Sahashi, EchoNet-Measurements provide automated annotations for the 18 most common #echofirst (both B mode and Doppler) measurements.
Excellent performance in test cohorts (overall R2 of 0.967 in the held-out CSMC dataset and 0.987 in the SHC dataset)
Excellent performance in test cohorts (overall R2 of 0.967 in the held-out CSMC dataset and 0.987 in the SHC dataset)
1/n We are thrilled to present EchoNet-Measurements, an open source, comprehensive AI platform for automated #echofirst measurements.
Using more than 1,414,709 annotations from 155,215 studies from 78,037 patients for training, this is the most comprehensive #echofirst segmentation model.
Using more than 1,414,709 annotations from 155,215 studies from 78,037 patients for training, this is the most comprehensive #echofirst segmentation model.
March 20, 2025 at 10:33 PM
1/n We are thrilled to present EchoNet-Measurements, an open source, comprehensive AI platform for automated #echofirst measurements.
Using more than 1,414,709 annotations from 155,215 studies from 78,037 patients for training, this is the most comprehensive #echofirst segmentation model.
Using more than 1,414,709 annotations from 155,215 studies from 78,037 patients for training, this is the most comprehensive #echofirst segmentation model.
New preprint by @SarnoffCardio and @dgsomucla Victoria Yuan:
There are new therapies for obstructive HCM, however obstruction is frequently missed. Extra #echofirst work is required to eval for obstruction.
We develop an AI model on standard A4C videos to identify patients w/ obstruction.
There are new therapies for obstructive HCM, however obstruction is frequently missed. Extra #echofirst work is required to eval for obstruction.
We develop an AI model on standard A4C videos to identify patients w/ obstruction.
March 7, 2025 at 9:03 PM
New preprint by @SarnoffCardio and @dgsomucla Victoria Yuan:
There are new therapies for obstructive HCM, however obstruction is frequently missed. Extra #echofirst work is required to eval for obstruction.
We develop an AI model on standard A4C videos to identify patients w/ obstruction.
There are new therapies for obstructive HCM, however obstruction is frequently missed. Extra #echofirst work is required to eval for obstruction.
We develop an AI model on standard A4C videos to identify patients w/ obstruction.
Consistency - one of the most useful heuristics of whether a medical AI model is good or not.
I've noticed that AI models trained on small datasets tend to jitter - jumping a lot from frame to frame - while robust models tend to have consistent measurements across the entire video.
Coming soon...
I've noticed that AI models trained on small datasets tend to jitter - jumping a lot from frame to frame - while robust models tend to have consistent measurements across the entire video.
Coming soon...
March 5, 2025 at 12:04 AM
Consistency - one of the most useful heuristics of whether a medical AI model is good or not.
I've noticed that AI models trained on small datasets tend to jitter - jumping a lot from frame to frame - while robust models tend to have consistent measurements across the entire video.
Coming soon...
I've noticed that AI models trained on small datasets tend to jitter - jumping a lot from frame to frame - while robust models tend to have consistent measurements across the entire video.
Coming soon...