MIT's class on Machine Learning in Healthcare is now available for free on MIT's OpenCourseWare! All videos, slides, and lecture notes can be found here:
We've launched Layer Health
@layerhealth
, a new AI startup solving healthcare's information problem, with large language models. I'm CEO, working with an amazing team and backed by $4 million from Google Ventures
@GVteam
@generalcatalyst
&
@inceptionhealth
In a first for ML virtual conferences,
#UAI2021
early-bird registration is FREE for students🎓😎🎉. Conference will be July 27-30. Details here:
@UncertaintyInAI
New dataset & benchmark for clinical NLP and unsupervised learning! With a team of clinicians, we annotated 718 discharge summaries, 100K sentences, for patient instructions, appointments, medications, labs, procedures, & imaging:
Can we improve patient and provider follow-up after a hospital discharge?
Our
#ACL2021NLP
paper, led by
@david_sontag
, uses ML to extract action-items from clinical notes, and releases one of the largest annotated datasets for clinical NLP. (1/3)
How to beat linear models for ML on health data? Pre-train using them! See our
#AAAI2021
paper on Reverse Distillation & a new transformer-based model (). And... (drum roll) our new open-source code for ML on OMOP:
@RBoiarsky
@OHDSI
Applications NOW OPEN for MIT's undergraduate summer research program, which seeks to increase the # of underrepresented minorities and underserved (eg low socio-economic bg, 1st gen) students in graduate research. Come work with me on ML for health care!
Research in health care is often very personal. My lab's research on
#MultipleMyeloma
started when my mom was diagnosed with it in 2014. It is too late to help her, but I believe machine learning can help millions of other patients. Thanks to
@TheMarkFdn
for their support.
May is National Cancer Research Month. In honor of it, we are proud of highlight the work of our partner
@david_sontag
of
@MIT_CSAIL
/
@MIT_IMES
, a computer scientists who is tackling myeloma
I've been increasingly worried about the fairness of machine learning in healthcare. ML is being used to prioritize care, and bias leads to disparities in who benefits. Here's my answer: "Why Is My Classifier Discriminatory?" w/
@irenetrampoline
@frejohk
MIT machine learning just got a whole lot stronger with new hires Pulkit Agrawal (
@pulkitology
), Jacob Andreas (
@jacobandreas
) and Cathy Wu (
@wucathy
). All starting in Fall 2019. All coming from my alma mater UC Berkeley :)
Mike finished his Ph.D. with me at MIT a year ago, and is now recruiting students in CS
@JohnsHopkins
. Consider applying to work with him - he's a great mentor and top researcher.
I'm recruiting PhD students for my lab at Johns Hopkins!
Please apply if you're interested in reliable ML / causal inference for decision-making in healthcare. See my website () for more info.
Deadline 12/15. Retweets welcome :)
Best paper award for
#UAI2021
goes to
@awnihannun
, Chuan Guo, and Laurens van der Maaten for their paper
Measuring Data Leakage in Machine-Learning Models with Fisher Information
Congratulations!
It *is* possible to get high-quality structured data from electronic medical records AND to save time; one just needs to re-think the user interface (and use machine learning). Here's a taste of the future:
(w/ Steven Horng
@bethisraellahey
@BIDMCEM
)
Attention all clinicians interested in machine learning & AI... work with MIT students! We are recruiting mentors for course projects in our Machine Learning for Healthcare Spring 2021 class. Details & sign up here:
@willieboag
@rayruizhiliao
@ckbjimmy
Want to apply ML to real-world EHR data? This Partners Biobank challenge is a great opportunity to compare your algorithms to other researchers' algorithms on a common health care dataset:
Jonas Peters and I are the program chairs for
#UAI2020
. Paper deadline is Feb 20, 2020, and conference will be in Toronto, Aug 3-6, 2020. General chairs are
@ryan_p_adams
and
@VibhavGogate
. Consider submitting your best work!
We have at least 5 postdoctoral positions at MIT CSAIL for first-generation college students, individuals underrepresented in graduate education at MIT, and others overcoming significant challenges in path toward graduate school -- applications due June 24th
MIT CSAIL is excited to announce the Mentored Opportunity in Research Postdoctoral Fellowship (Meteor), to broaden participation in computing and artificial intelligence:
Wow!
CHiQA: an experimental AI system for answering health-related questions for patients
See paper & new (publicly available) datasets just published in JAMIA '19 by Demner-Fushman and colleagues
@NLM_LHC
Presenting MedKnowts... a new way of writing clinical notes for electronic medical records, developed by AI and HCI researchers at
@MIT
and clinicians at
@BIDMChealth
.
"Simple, Distributed, and Accelerated Probabilistic Programming". The
#NIPS2018
paper for Edward2. Scaling probabilistic programs to 512 TPUv2 cores and 100+ million parameter models.
Our new paper "Fair Regression for Health Care Spending" is out:
We build fairness into the objective function for continuous outcomes & see large improvements in group undercompensation
Coauthored w/PhD student Anna Zink
Code:
Overparameterization seems to be very helpful for supervised learning of deep neural networks -- but what about for unsupervised learning (e.g. using variational auto-encoders)? w/Rares Buhai,
@risteski_a
, and Yoni Halpern:
Are you a medical student, resident or fellow with an interest in ML or NLP? MIT's clinical machine learning group needs your help with a user study on a new tool for rapid annotation of clinical notes! It'll take 15 hours and is paid. Details here:
Monica finished her Ph.D. with me at MIT a year ago and is starting a research group at Duke in Sept 2024. As a leader both in industry
@layerhealth
and in academia, her LLM work is transforming healthcare. Consider applying to
@dukecompsci
and
@DukeBiostats
I’m recruiting PhD students
@Duke
for fall 2024! Consider applying if you’re interested in reimagining healthcare by developing novel ML/NLP methodology. I can advise students through the CS dept and the Biostats & Bioinformatics dept.
Info here:
Great tutorial on causal inference! I particularly like the sections on checking for violations of assumptions, natural experiments, and the framing of potential outcomes within the broader language of causal graphs.
at
#KDD2018
? Interested in estimating effects of algorithms or applying ML to societal domains like healthcare? Check out our tutorial on causal inference and counterfactual reasoning at 1pm today
@emrek
@kdd_news
MIT's free Machine Learning for Healthcare consulting sessions begin March 5th at 5pm. All clinician researchers from the Boston area are welcome! Sign up here:
All of Us research platform opens up for beta testing, with data on 225,000 participants, enabled by cloud-based Jupyter notebooks and an OMOP common data model
Congrats to my Ph.D. student, Yoni Halpern, for the AMIA Doctoral Dissertation Award second place prize, for his thesis "Semi-Supervised Learning for Electronic Phenotyping in Support of Precision Medicine"
AMIA Doctoral Dissertation Award Winners announced!
@blpercha
and Yonatan Halpern will receive awards and present their doctoral work at
#AMIA2018
. Read more about their winning dissertations, additional finalists, and the award program
Amazon web services just launched a product to extract medical concepts and perform de-identification in unstructured text in electronic medical records! This is incredibly exciting (1/3)
We've released the largest-ever open-source dataset to speed up MRIs using
#AI
. Read how we've collaborated with
@facebookai
to provide researchers access to MR imaging and improve patient care worldwide
Double/debiased machine learning
Individual-level treatment
Regulatory oversight
From development to deployment
Teaching yourself about structural racism
As the editors of
@biostatistics
,
@drizopoulos
and I are thrilled to share this free access multidisciplinary collection of commentaries on machine learning for causal inference.
All 5 pieces are linked in our editorial about the series:
We used machine learning on EHR audit logs to learn to predict -- so that we can automatically surface -- patient notes that clinicians may want to read. Published today at
#mlhc2023
, joint work between
@MIT
and
@BIDMChealth
.
Paper:
#NoMoreThousandClicks
Clinicians have significant + shifting information needs over the course of a patient visit. In our new paper, we characterize + predict the notes relevant for clinicians to read, based on the current clinical context. Presenting this work, led by Sharon Jiang, today at
#MLHC2023
As Copilot becomes more popular, we need to understand how programmers interact with it. We built a model of interaction between Copilot and Programmers named 'CUPS' and predict programmer behavior in our latest paper
I'll be hosting a mentorship session Tues 4:30pm EST as part of
@NeurIPSConf
's Mementor Beta. Hoping to chat with junior researchers working on machine learning in healthcare. May host more sessions later in the week depending on how it goes.
New dataset alert!
AMR-UTI: Antimicrobial Resistance in Urinary Tract Infections
De-identified data derived from 80,000 patients with urinary tract infections (UTI) treated at Massachusetts General Hospital and Brigham & Women’s Hospital in Boston
New book on "Artificial Intelligence in Medical Imaging: Opportunities, Applications and Risks" with chapters by
@DrHughHarvey
and
@DrLukeOR
. Note: check your academic library, e.g. MIT has free access to PDF version of book.
Call for abstracts for NeurIPS 2019 causality workshop, "Do the right thing": machine learning and causal inference for improved decision making. Due Sept 9th. More details here:
The 6 strategic areas MIT will be hiring in in the College of Computing (not all this year): social, economic, & ethical implications; natural intelligence; human health; *planetary* health; human experience; quantum computing
Working in the area of medical data science and want to be a tenure-track Assistant Professor at MIT? We've got a new opening joint between EECS and the Institute for Medical Engineering and Science. Apply by Feb 28th!
@MIT_IMES
@MITEECS
@MIT_CSAIL
ML faculty team “Punsupervised Learning” for scavenger hunt at Friday evening’s
@MITEECS
Ph.D. visit day, w/ Tamara Broderick,
@jacobandreas
, and Song Han
Paper deadline for
#UAI2021
is Feb 19. Program chairs are Marloes Maathuis &
@cassiopc
, general chairs
@david_sontag
& Jonas Peters. Virtual conference will be July 27-30, single track, with assigned discussants for select papers
@MichaelOberst
and I tackle the following question in our upcoming ICML 2019 paper, motivated by our lab's research of ML in healthcare: How do you build trust in a new policy learned by reinforcement learning from observational data?
Runner-ups for best paper award
#UAI2021
go to
Batya Kenig
Approximate Implication with d-Separation
Francisco Ruiz, Michalis Titsias
@TaylanCemgilML
@ArnaudDoucet1
Unbiased Gradient Estim. for VAEs using Coupled Markov Chains
@ShalitUri
The title of Google's paper and the abstract's emphasis on deep learning is indeed misleading, given that a simple regression on a reasonable feature set gets similar performance (as I predicted back in February: ). /1
@eigenhector
@zacharylipton
@Google
@MarzyehGhassemi
@davekale
@JeffDean
Where are the results reported for this l1-regularized linear model, using the same features? The title and abstract emphasize the term "deep". Academic papers shouldn't be marketing tools: we should avoid hype unless it is actually warranted by the results.
Practical problem: you have a dataset and want to do causal inference. How to report the validity region, i.e. where the new policy should be used? Joint w/
@frejohk
, Dennis Wei,
@MichaelOberst
, Tian Gao,
@bratogram
,
@krvarshney
(part of the
@MITIBMLab
):
@ShalitUri
I encourage folks to see slides 8-13 of my March talk at the AMIA summit "The (Current) Limits of Deep Learning in Health Care" () tl;dr: RNNs for risk stratification on longitudinal clinical data don't yet outperform simpler models, and for good reasons /2
Pete Szolovits and I are looking for a TA for our MITx course Machine Learning for Healthcare, which will launch in early 2021 (~5 hr/week commitment). Please DM or email me if interested! Looking for PhD students or recent PhD graduates with research and/or industry experience.
Still uncommon in unsupervised machine learning to have an algorithm that we both understand theoretically and which works well in practice -- our work on topic modeling is one example:
Thanks to
@blei_lab
for perspective:
At MIT we are developing an AI-based tool to make doctors’ notes easier to read for patients. We are recruiting breast cancer patients for a user study with their notes. Participants will get $30 compensation. If interested, see . Please help advertise!
Benchmarks for Bayesian deep learning. E.g., one can use this to quickly evaluate algorithms that "predict diabetic retinopathy, and use their uncertainty for prescreening (sending patients the model is uncertain about to an expert for further examination)"
Really excited to release Bayesian Deep Learning Benchmarks - please share with others who you think might like this, and have a look at the blog/repo/colab:
This work was done over a period of a year and a half by many collaborators
@OATML_Oxford
Our
#MLHC2021
paper, latest work in collaboration between
@MIT
and
@sloan_kettering
. We show we can reduce 80% of model errors with less than 15% of the manual annotation effort.
Many variables needed to construct timelines for clinical research are trapped in notes 🗒️ Manual extraction can be expensive, and ML is still error-prone. In our
#MLHC2021
paper, we explore a hybrid approach that can extract clinical events accurately with minimal oversight!
Reminiscing about what ICML and NeurIPS workshops used to be like in the early 2000's, when they were focused and had actual discussion & brainstorming about research in progress? Let's recreate that experience. Submit a workshop proposal for
#UAI2021
!
Best Treatment for the Coronavirus? Paid Sick Leave
("The government could defray the cost of emergency sick leave for employers, for example by allowing businesses to claim a one-time tax credit")
I'm teaching a 3-day Machine Learning for Healthcare class (in person, at MIT) from June 27-29 through
@MITProfessional
incl. labs applying ML to health data, a sneak peak at the state-of-the-art research, and Q&As with health AI leaders in industry.
So excited to see this first BIG step in the right direction. This is what can happen when we invest in health IT, informatics research, and common data models such as i2b2 and OMOP.
Great by
@zakkohane
and colleagues: Consortium for Clinical Characterization of COVID-19 by EHR (4CE). Harmonized data sets analyzed locally and shared as aggregate data for rapid analysis and visualization. Great teamwork. Like
@OHDSI
.
@medrxivpreprint
Looking for a postdoc and passionate about conducting research in foundational ML and disease biology? Come join the new Eric & Wendy Schmidt Center
@broadinstitute
!
I'm affiliate faculty and would be thrilled to collaborate - mention my name
We've had previous success with clinicians in Boston taking MIT's Machine Learning for Healthcare class - come join next year's cohort! Apply here:
Some background in Python & machine learning needed, but we will help with resources to guide learning.
When deciding what treatment to prescribe, we don't always need exact counterfactual estimates - bounds might suffice. And we can estimate accurate bounds using less data!
#ICML2020
paper with
@Maggiemakar
,
@frejohk
, and John Guttag
Slide from my sister
@LauraBKleiman
's
@CWR4C
talk
@kochinstitute
@MIT
yesterday on drug repurposing for cancer
Real-world data + observational data -> real-world evidence
ML, causal inference, health data sets
This could have been one of my talks!
@overleaf
Could you push your maintenance forward by a day? There's a major machine learning conference (ICML) paper deadline a few hours after your maintenance period, and many folks will be using Overleaf during that time.
Boston-area postdocs, want to devote one day per week (funded) to participate in a new approach to biomedical research? Join the MIT Catalyst program! In past projects, 75% went on to receive follow-on funding; 44% transitioned to commercial development.
@zacharylipton
@Google
@MarzyehGhassemi
@davekale
@JeffDean
@JeffDean
, while you are revising the paper, please consider adding a "simple" machine learning baseline of l1-regularized logistic regression using the same inputs as your deep models and backward time windows (see , "Enhanced model", for an example).
I've recently starting using the phrase "target deployment" to characterize how ML in health care should be evaluated, in analogy to
@_MiguelHernan
's "target trial"
Typically, we evaluate models, then quantify potential net-benefit, and realize some of it with real world operational constraints. I believe we can do a lot better in taking actions in response to a prediction. See our JAMA viewpoint
@StanfordDeptMed
Public service announcement for folks with links on their websites to previously recorded
@icmlconf
/
@NeurIPSConf
/
@aclmeeting
talks
@TechTalksTV
- you will want to remove the links. The website has been taken over by spammers.
We then built machine learning algorithms to automatically extract or highlight these action items from hospital discharge notes.
Too often things fall between the cracks during transitions in care. Time to change.
@JeffDean
@zacharylipton
@Google
@MarzyehGhassemi
@davekale
Agreed. However, please consider for v3. I often find that the gap is quite small between cleverly designed deep models and variants of the simple approach I suggested -- on precisely the same problems that you've looked at.
Elated to share that I've been named a 2019
@PDSoros
fellow! I'm very honored and blessed to be part of such an incredible community which highlights the immigrant experience and American dream. 🇺🇸🇺🇸
MIT's free Machine Learning for Healthcare consulting sessions begin March 5th at 5pm. All clinician researchers from the Boston area are welcome! Sign up here: