Shino Kany
@shinokany.bsky.social
Cardiology and EP at UKE Hamburg, AI and Genetics at Broad Institute of MIT and Harvard.
Many thanks to my co-lead Sam Friedman, and co-authors Mostafa Al-Alusi, Shaan Khurshid, @joelramo.bsky.social, Daniel Pipilas, @jamespirruccello.com, Christopher Reeder, Anthony Philippakis, Jennifer Ho, Mahnaz Maddah, and the co-senior authors of this project, Patrick Ellinor, and Akl Fahed.
August 2, 2025 at 12:49 PM
Many thanks to my co-lead Sam Friedman, and co-authors Mostafa Al-Alusi, Shaan Khurshid, @joelramo.bsky.social, Daniel Pipilas, @jamespirruccello.com, Christopher Reeder, Anthony Philippakis, Jennifer Ho, Mahnaz Maddah, and the co-senior authors of this project, Patrick Ellinor, and Akl Fahed.
In the BWH primary care cohort, individuals in the top quintile of ECG2CAD risk faced a 5–10× higher hazard of myocardial infarction or heart failure, and nearly 3× higher all-cause mortality over ~8 years.
August 2, 2025 at 12:49 PM
In the BWH primary care cohort, individuals in the top quintile of ECG2CAD risk faced a 5–10× higher hazard of myocardial infarction or heart failure, and nearly 3× higher all-cause mortality over ~8 years.
Performance remained consistent across age, sex, and race/ethnicity subgroups, and even on ECGs read as “normal” by cardiologists or in patients without classical risk factors.
August 2, 2025 at 12:49 PM
Performance remained consistent across age, sex, and race/ethnicity subgroups, and even on ECGs read as “normal” by cardiologists or in patients without classical risk factors.
We observed a enhanced performance versus clinical risk scores such as PCE. This was particular seen in better average precision, which tells us how good a model flags high-risk individuals.
August 2, 2025 at 12:49 PM
We observed a enhanced performance versus clinical risk scores such as PCE. This was particular seen in better average precision, which tells us how good a model flags high-risk individuals.
Across three test sets, ECG2CAD achieved AUROCs of 0.78 (MGH), 0.75 (BWH) and 0.76 (UK Biobank), outperforming models based on aage and sex. When combined with age and sex, the minimal information a clinican would have, performance improved further.
August 2, 2025 at 12:49 PM
Across three test sets, ECG2CAD achieved AUROCs of 0.78 (MGH), 0.75 (BWH) and 0.76 (UK Biobank), outperforming models based on aage and sex. When combined with age and sex, the minimal information a clinican would have, performance improved further.
We developed ECG2CAD, a convolutional neural network trained on 764,670 12-lead ECGs from 137,036 patients at MGH, to predict prevalent CAD at the time of ECG. Validation was performed on independent cohorts from @MGH, Brigham & Women’s Hospital, and the UK Biobank.
August 2, 2025 at 12:49 PM
We developed ECG2CAD, a convolutional neural network trained on 764,670 12-lead ECGs from 137,036 patients at MGH, to predict prevalent CAD at the time of ECG. Validation was performed on independent cohorts from @MGH, Brigham & Women’s Hospital, and the UK Biobank.
Coronary artery disease remains the leading cause of death globally, yet many individuals lack timely, noninvasive screening options. Deep learning applied to routine ECGs offers a scalable path to earlier detection, as shown in other diseases such as Afib or heart failure.
August 2, 2025 at 12:49 PM
Coronary artery disease remains the leading cause of death globally, yet many individuals lack timely, noninvasive screening options. Deep learning applied to routine ECGs offers a scalable path to earlier detection, as shown in other diseases such as Afib or heart failure.
Thanks, yeah they are wild. But thats also a function of directly measuring AV function instead of prediction. So was nice to see.
April 1, 2025 at 2:02 PM
Thanks, yeah they are wild. But thats also a function of directly measuring AV function instead of prediction. So was nice to see.
I want to thank our collaborators from NEDA, Simon Stewart, David Playford, Geoff Strange and Yih-Kai Chan. The most enormous thanks go to my mentors @jamespirruccello.com and Patrick Ellinor. James is the brain behind all this and has guided me through this effort with great patience and support.
April 1, 2025 at 3:31 AM
I want to thank our collaborators from NEDA, Simon Stewart, David Playford, Geoff Strange and Yih-Kai Chan. The most enormous thanks go to my mentors @jamespirruccello.com and Patrick Ellinor. James is the brain behind all this and has guided me through this effort with great patience and support.
I want to thank all our coauthors and collaborators: @joelramo.bsky.social , Cody Hou, Sean Jurgens, Victor Nauffal, @jonwcunningham.bsky.social, @emilyswlau.bsky.social, @atulbutte.bsky.social, Jeffrey Olgin, Sammy Elmariah, @markelindsay.bsky.social , and all participants at UK Biobank and NEDA.
April 1, 2025 at 3:31 AM
I want to thank all our coauthors and collaborators: @joelramo.bsky.social , Cody Hou, Sean Jurgens, Victor Nauffal, @jonwcunningham.bsky.social, @emilyswlau.bsky.social, @atulbutte.bsky.social, Jeffrey Olgin, Sammy Elmariah, @markelindsay.bsky.social , and all participants at UK Biobank and NEDA.
Our findings suggest that current definitions may miss a large group with prognostically significant aortic valve disease. Adopting our purely hemodynamic “mild ASproposed” thresholds could improve early identification and influence surveillance and prevention strategies.
April 1, 2025 at 3:31 AM
Our findings suggest that current definitions may miss a large group with prognostically significant aortic valve disease. Adopting our purely hemodynamic “mild ASproposed” thresholds could improve early identification and influence surveillance and prevention strategies.
But what if this is only related to MRI and does not hold for the clinical standard diagnostic (Echo)? To answer this, we looked at the National Echo Database Australia (NEDA, N=365,870) where ASproposed criteria predicted increased all-cause mortality (HR 1.25) and CV mortality.
April 1, 2025 at 3:31 AM
But what if this is only related to MRI and does not hold for the clinical standard diagnostic (Echo)? To answer this, we looked at the National Echo Database Australia (NEDA, N=365,870) where ASproposed criteria predicted increased all-cause mortality (HR 1.25) and CV mortality.
Crucially, this definition identified almost all people at risk for AV replacement. The cumulative incidence of AV surgery by year 6 among those without AS (n = 46,988) was 0.032% compared with 1.8% with mild ASproposed , 17.4% with moderate AS (n = 271), and 56.2% in severe AS (n = 30).
April 1, 2025 at 3:31 AM
Crucially, this definition identified almost all people at risk for AV replacement. The cumulative incidence of AV surgery by year 6 among those without AS (n = 46,988) was 0.032% compared with 1.8% with mild ASproposed , 17.4% with moderate AS (n = 271), and 56.2% in severe AS (n = 30).
Applying these thresholds, 5.8% of participants were classified with mild ASproposed. They had a higher risk for AV procedures (HR 31.7), a risk of atrial fibrillation (HR 1.86), and heart failure (HR 2.37) compared with individuals without AS over ~4 years of follow-up.
April 1, 2025 at 3:31 AM
Applying these thresholds, 5.8% of participants were classified with mild ASproposed. They had a higher risk for AV procedures (HR 31.7), a risk of atrial fibrillation (HR 1.86), and heart failure (HR 2.37) compared with individuals without AS over ~4 years of follow-up.
In a healthy subcohort (n=41,859), we derived reference ranges for AV function by age&sex. We defined “mild ASproposed” (outside the 95th percentile >any age group) as any 1 of these criteria: peak velocity >1.65 m/s, mean gradient >4.8 mmHg, or AVA <2.1 cm² (men) / <1.7 cm² (women).
April 1, 2025 at 3:31 AM
In a healthy subcohort (n=41,859), we derived reference ranges for AV function by age&sex. We defined “mild ASproposed” (outside the 95th percentile >any age group) as any 1 of these criteria: peak velocity >1.65 m/s, mean gradient >4.8 mmHg, or AVA <2.1 cm² (men) / <1.7 cm² (women).
However, the UK Biobank MRIs don't come with clinical reports. So, we developed a deep learning model to segment the blood pool above the asc.aorta which allowed us to quantify aortic valve area, peak velocity, and mean gradient from velocity-encoded MRI data in 62,902 UK Biobank participants.
April 1, 2025 at 3:31 AM
However, the UK Biobank MRIs don't come with clinical reports. So, we developed a deep learning model to segment the blood pool above the asc.aorta which allowed us to quantify aortic valve area, peak velocity, and mean gradient from velocity-encoded MRI data in 62,902 UK Biobank participants.
So, we wanted to see what “normal” aortic valve (AV) function is and what clinical consequences would be associated with abnormal AV function. However, we would need a population-based cohort with imaging without clinical indication. Enter the UK Biobank, a prospective cohort study from the UK.
April 1, 2025 at 3:31 AM
So, we wanted to see what “normal” aortic valve (AV) function is and what clinical consequences would be associated with abnormal AV function. However, we would need a population-based cohort with imaging without clinical indication. Enter the UK Biobank, a prospective cohort study from the UK.
Aortic stenosis (AS) is the most common valve lesion. There is no medical treatment and major trials focus on moderate/severe AS. But what if we should intervene earlier for meds to work? Issue is “mild AS” is only defined for peak velocity and w/ varying definitions (Vmax 2-2.9m/s or 2.6-2.9m/s).
April 1, 2025 at 3:31 AM
Aortic stenosis (AS) is the most common valve lesion. There is no medical treatment and major trials focus on moderate/severe AS. But what if we should intervene earlier for meds to work? Issue is “mild AS” is only defined for peak velocity and w/ varying definitions (Vmax 2-2.9m/s or 2.6-2.9m/s).
I extend my heartfelt gratitude to my mentors Patrick Ellinor and Shaan Khurshid whose unwavering patience and guidance continue to inspire and support me every day. And of course all institutions: UKE Hamburg, @broadinstitute.org and Mass General.
February 18, 2025 at 8:08 PM
I extend my heartfelt gratitude to my mentors Patrick Ellinor and Shaan Khurshid whose unwavering patience and guidance continue to inspire and support me every day. And of course all institutions: UKE Hamburg, @broadinstitute.org and Mass General.
This work was possible thanks to prior work and amazing co-authors Mostafa Al-Alusi, @joelramo.bsky.social, @jamespirruccello.com, @ajufoezim.bsky.social , Tim Churchill, Steven Lubitz, Mahnaz Maddah, @jsawallagusehmd.bsky.social and all @ukbiobank.bsky.social participants.
February 18, 2025 at 8:08 PM
This work was possible thanks to prior work and amazing co-authors Mostafa Al-Alusi, @joelramo.bsky.social, @jamespirruccello.com, @ajufoezim.bsky.social , Tim Churchill, Steven Lubitz, Mahnaz Maddah, @jsawallagusehmd.bsky.social and all @ukbiobank.bsky.social participants.