What does your electric heart tell you about the entire human disease landscape? Our ECG PheWAS was just published in @npjDigitalMed! A phenome wide association study of a deep learned latent space from 100K+ ECGs meta analyzed in 3 external cohorts...

NEW from our group @broadinstitute @MGHHeartHealth @patrick_ellinor led by @ShinoKany and @lucidtronix ECG2CAD: An AI model for detection of coronary artery disease using 12-lead ECGs @JACCJournals https://t.co/wUKbjRxURI

A great collaboration with @ShinoKany @joeltramo @DPipilasMD @pulkitsng @reeder_ml @shaan_khurshid @jpirruccello @MahnazMaddah @JenHoCardiology @patrick_ellinor and @broad_ml4h

Median waveforms computed from high risk and low-risk individuals have distinct morphology and the saliency maps highlighted many parts of the ECG

Even ECG's read as normal conveyed information about diverse cardiovascular outcomes, suggesting latent information about disease risk that is currently flying under the radar.

The hazard ratio from the ECG-based hypertension score (HTN-AI) was higher than that from systolic blood pressure or pulse pressure: Co-led with @alalusim_md https://t.co/r2n63eEq30

Is the ECG secretly a sphygmomanometer? We trained a neural net to predict hypertension from ECGs and found it associated with many cardiovascular outcomes.

In a new study, researchers from @mghhearthealth and colleagues show that with #AI, the ECG contains information that can help detect over 600 diseases. The study was published in @NaturePortfolio. Read more: https://t.co/6D9iEntaQT  @lucidtronix @shaan_khurshid @rvpacing

@shaan_khurshid Endless thanks to coauthors: @xxinxwang, @RVpacing , Nate Diamant, @LuChenWeng, Seung Hoan Choi, @reeder_ml, @jpirruccello, Pulkit Singh, @emilyswlau, Anthony Philippakis, @CDAndersonMD, @MahnazMaddah, @gpbatra, @patrick_ellinor, @JenHoCardiology, & @steven_lubitz!

Co-lead @shaan_khurshid has this excellent thread summarizing the results

Across the 1600+ Phecode diagnoses we tested, the ECG latent space was significantly informative for 645 prevalent and 606 future diseases (after Bonferroni correction)

But there are strong comorbid correlations across many disease categories. One interesting exception is the mostly negative correlation shown by the blue-ish stripe corresponding to Pregnancy Complications which occur in a relatively healthier part of our ECG latent space.

Beautiful, informative structure can be seen in the cross correlation of the 256-dimensional ECG Disease vectors, Circulatory System diseases in particular light up with comorbid correlation. https://t.co/bWxdnEOZze

Just how many diseases can you identify with a good old 12-lead #ECG? NEW @npjDigitalMed, co-led w/ @lucidtronix, @RVpacing & @xwang, we use an autoencoder to map ECG to phenome, finding associations btwn ECG & >600 conditions @broadinstitute @broad_ml4h @MGHHeartHealth LINK⬇️

What does a 1-parameter autoencoder of cardiac MRIs learn? If you use that 1 number to parameterize a circle you learn the heart beat, if not, you learn the exposure value of the image -- It depends on the inductive bias! Deets in: https://t.co/As3i2spWKM

Can ECG #AI predict who will develop #afib after stroke (eg, to prioritize for ILR)? Yes! New in @CirculationEP our @broad_ml4h ECG #AI #afib risk model discriminates cardioembolic stroke (>> clinical factors) & stratifies risk of new #afib after stroke: https://t.co/bfIiX9wXC5

A very fun collaboration with @gemma_e_moran, Marianne Rakic and Anthony Phillipakis! @broad_ml4h

The punchline of this paper is that registration, spatial or temporal, guided by domain knowledge or learned from scratch, helps uncover the Biological information – whether that info is genotypic as measured by GWAS or phenotypic as measured by PheWAS.

Nowadays, neural networks can learn to register data modalities together, in flexible ways, for example DeepCycle, VoxelMorph, and Dropfuse all learn to register images using diverse inductive biases.

Registration has ancient roots – at least as far back as ancient Egypt where land was surveyed with regularly knotted ropes that were gathered together—physically implementing isomorphic scaling.