Thread by @JFutoma

Mihaela van der Schaar kicking off day 2 of #MLHC2019 with work on learning individualized risk predictions in cancer

Shots fired...Mihaela claims most current disease trajectory models are overly simplistic and not useful, due to the Markov assumption #MLHC2019

Deep learning disease trajectory models may predict better, but is the lack of interpretability a problem? #MLHC2019

Shots fired, again #MLHC2019

The relationship between the proposed PASS and other related classes of trajectory models #MLHC2019

Moving on to a different application, on how often measurements should be made / personalized screening #MLHC2019

Moving on to causal inference...is Mihaela on team potential outcomes? No DAGs? #MLHC2019

Ahhh there it is, that common assumption: ignorability (is this ever actually true?) #MLHC2019

Better causal inference method(s) incoming #MLHC2019

Bringing BNP to ITE estimation. The presented theory shows the minimax rate depends on the underlying level of sparsity and smoothness #MLHC2019

The claim is made that in the large sample regime, selection bias is of minimal importance (but what I really want to know is how tight is this proposed bound? Is it useful practically?) #MLHC2019

Interesting...GANs for ITE. Curious how this performs vs more traditional methods #MLHC2019

Sadly we didn’t get to this last section of the talk, but would have been neat to see the plan for how to scale things up #MLHC2019

Moving on to the first round of clinical abstract spotlights now - looking forward to it! #MLHC2019

Starting us off with work from folks at Michigan on using ML to help patients improve faster from occupational injury. They use RL! Can’t wait to see how at their poster #MLHC2019

First abstract from @DukeHealth / @DukeInnovate! On how to best combine data from different hospital sites to yield the strongest predictions: pooling and ensembling appears to improve performance across the board #MLHC2019

And now: it’s Bean Time! @kdpsinghlab presenting super interesting (and scary!) work on privacy. Hard to believe this is actually real, tons of PHI to be pulled directly from tree-based models...incredible #MLHC2019

Another UM abstract: FIDDLE, a useful data processing pipeline #MLHC2019

And that’s followed up by another pipeline, specifically for MIMIC. Much needed for this community! Compared to past work, better includes background knowledge, especially around grouping raw features into clinical concepts #MLHC2019

Back to UM again, on predicting extubation failure #MLHC2019

And now, an abstract on looking into the data integrity of eICU, a very commonly used data source in this community. Neat! #MLHC2019

And now work from Toronto / Sick Kids on pediatric applications of ML! Goal to improve triage and reduce pediatric wait times #MLHC2019

Same presenter from Sick Kids, on a sepsis detection tool in the ED, also in pediatrics. #MLHC2019

Another sepsis abstract, this time on learning phenotypes! Seems super useful since sepsis is ridiculously heterogeneous and poorly understood #MLHC2019

Back to UM, on a neurosurgical imaging application to improve ease of diagnosis. But they ran an RCT, it sounds like? Wild!
“This is like insane” - @MarkSendak #MLHC2019

And now a project from NYU Langone on improving impairment after stroke #MLHC2019

Another abstract, same presenter now talking about refining movement quantitation in stroke #MLHC2019

A more methodological abstract on predicting ICU length of stay with competing risks & deep learning from Philips Healthcare #MLHC2019

And closing out this section, a former @DukeInnovate Scholar presenting on a model to predict in-hospital mortality - but more importantly, a thoughtful dive into how this model will actually be used in clinical practice & deployed. Proud to have been a part of this one!#MLHC2019