Weston Hughes
@jwestonhughes
Followers
201
Following
15K
Media
24
Statuses
112
Career update: this week I’m starting as a senior data scientist with @PierreEliasMD at Columbia/NYP Cardiology! I’ll be both continuing academic research and exploring pathways to widespread deployment of our work in the clinic. Excited to keep building!
1
0
29
1/6 🏥 Structural heart disease (SHD) is common and often missed. What if a standard ECG could spot it—accurately, at scale, across hospitals, and patient types? Meet EchoNext, our AI model trained on 1.2M ECG-echo pairs. 👉 Published in Nature 🔗
nature.com
Nature - EchoNext, a deep learning model for electrocardiograms trained and validated in diverse health systems, successfully detects many forms of structural heart disease, supporting the...
1
3
13
🧵1/Today, we published a key milestone towards AI based cardiac screening in Nature. https://t.co/Lr3ymIrgz5 EchoNext outperformed cardiologists and found thousands of high-risk patients missed in routine care. We also made a version available to the world.
14
45
195
Excited to share our new paper in @NatureCVR!! We decoded how genes team up through epistasis to drive cardiac hypertrophy. 🧬❤️ 🙏👏@euanashley @bbiinnyyuu
@Stanford @UCBerkeley Link: https://t.co/PEjECVeqya NCVR news: https://t.co/N7GTuGb6Gd
3
6
14
🧵 AI for Valvular Heart Disease: DELINEATE-Regurgitation Study Excited to share our latest work on using deep learning to improve diagnosis and prediction in valve regurgitation. Let's dive in! #CardioTwitter #EchoAI @PierreEliasMD
https://t.co/bAGY3lEM25 (1/8)
academic.oup.com
AbstractBackground and Aims. Classification and risk stratification in aortic (AR), mitral (MR), and tricuspid regurgitation (TR) remains a significant cli
1
5
15
So excited to welcome @jwestonhughes, all-star alum of @euanashley @james_y_zou's labs, to Columbia/NYP! The future of cardiac AI is bright.
Career update: this week I’m starting as a senior data scientist with @PierreEliasMD at Columbia/NYP Cardiology! I’ll be both continuing academic research and exploring pathways to widespread deployment of our work in the clinic. Excited to keep building!
0
2
23
@PierreEliasMD I'll be working remotely from the SF with frequent visits to NYC. Let me know if you're in either city and want to catch up!
0
0
0
Excited to share 'KardioNet' published in @JACCJournals! Our proof-of-concept #AI model to detect hyperkalemia from the #AppleWatch #ECG . A huge thanks to @David_Ouyang for all the advice and support! paper: https://t.co/4TzDRTVvpc code:
Excited to share our new work #KardioNet, an AI-enabled smartwatch ECG algorithm for serum potassium monitoring. Trained on ~300k ECG and potassium pairs at @CedarsSinai, Kardio-Net is designed to enhance potassium monitoring via smartwatch for ESRD patients. #digitalhealth
2
5
19
Glad to share this work is now out in European Heart Journal – Digital Health! Thanks again to @mvperez92 @euanashley @james_y_zou @David_Ouyang @PierreEliasMD! Paper here:
Exciting new preprint out this week on the tradeoff between complexity and performance in machine learning for ECG models! Here, we walk through a range of models for detecting low EF at @StanfordMed, from simple and interpretable to complex and highly accurate. 1/n
1
8
22
Excited to share our new work #KardioNet, an AI-enabled smartwatch ECG algorithm for serum potassium monitoring. Trained on ~300k ECG and potassium pairs at @CedarsSinai, Kardio-Net is designed to enhance potassium monitoring via smartwatch for ESRD patients. #digitalhealth
1
5
31
Excited to present work out today at @NEJM_AI with @amey_vrudhula @jwestonhughes @yuanneal. We show that question formulation is important medical AI. In developing AI-ECG models, we show that choosing a REGRESSION task outperforms a CLASSIFICATION task for for low EF.
9
29
141
New work from @jwestonhughes extending his already comprehensive approach to #AI for the ECG by addressing complexity and performance. Although we love the power and non-linearity of neural network models, we often wonder just how well more traditional statistical models stack
Exciting new preprint out this week on the tradeoff between complexity and performance in machine learning for ECG models! Here, we walk through a range of models for detecting low EF at @StanfordMed, from simple and interpretable to complex and highly accurate. 1/n
0
2
17
Very important result which comports with our experience as well. Using standard ECG features can do surprisingly well for many prediction tasks, including reduced LVEF!
Exciting new preprint out this week on the tradeoff between complexity and performance in machine learning for ECG models! Here, we walk through a range of models for detecting low EF at @StanfordMed, from simple and interpretable to complex and highly accurate. 1/n
0
1
5
As always, a project like this takes a village. Thanks to the amazing team: Sulaiman Somani @PierreEliasMD @JamesETooley AJ Rogers @timpotsMD Chris Haggerty @David_Ouyang @euanashley @james_y_zou and Marco Perez 9/n
1
0
3
Takeaways: DL may give the highest AUC, but simpler models can also yield interesting insights. The biggest, most accurate model won’t always be the best for every situation. And we should spend time evaluating the whole range of models, not just the shiniest new neural net. 8/n
1
0
0
We also found that single measurements and our small linear model perform well across multiple sites (Stanford, Columbia, UK Biobank), while the DL model can see a significant drop in performance. Simple models can be easier to port to different medical systems vs DL models. 7/n
1
0
1
And what’s more, these single measurements are strong indicators of low EF on their own. The T-wave amplitude in aVR is as strong a predictor of low EF as BNP is! This even as the aVR lead is often ignored during ECG analysis. https://t.co/2uG0gkmHZ6 6/n
1
0
1
These 5 measurements in a linear model achieve an AUC 0.86 in detecting low EF. There’s a performance drop vs the DL application, but the model is completely interpretable and can be computed by hand, allowing for new insights and application in low-compute scenarios. 5/n
1
0
0
But 555 measurements is a lot to wrap your head around, so we next tried to find a small group of measurements which can detect low EF with reasonable accuracy. We came up with a list of five: the QTc, heart rate, aVR T amplitude, V3 QRS duration, and V3 minimum deflection. 4/n
1
0
1