1/9 Excited to introduce the LCD Benchmark! A new benchmark dataset for 30-day out-of-hospital mortality prediction (predicting the future!) using long clinical documents. Designed for testing the capabilities of LMs on lengthy documents! medrxiv.org/content/10.1101/… #ClinicalNLP
9/9 Most importantly, I would like to thank our esteemed co-authors: @shan23chen, @Serena_pancakes, @dmitriydligach, @dbittermanmd, @Majeans2011, and @tmills(my postdoc PI)! I am very grateful to you all. Thanks to your valuable contributions, this project has reached this point!
8/9 Attention Analysis:
Using the hierarchical transformer model and norm-based analysis, we provide insights into which sections of the discharge summaries contribute most to predictions, enhancing model interpretability.
2
100
7/9 Models Evaluated:
* Bag-of-Words
* CNN
* Hierarchical Transformers
* Mixtral
* GPT-4
Our results show that long clinical document processing remains a challenging yet exciting area. Please submit your model's output through CodaBench and compare it with others.
1
89
6/9 Get the Data:
Access the Long Clinical Document (LCD) Benchmark dataset through GitHub (github.com/Machine-Learning-…).
Note: You will need to download notes from PhysioNet.org after fully executed MIMIC-IV data use agreement.
#MIMIC4 #MIMICIV
2
83
5/9 Impact:
We expect this benchmark to become a valuable resource for developing and evaluating advanced NLP models tailored for clinical text. It will also facilitate timely end-of-life conversations and improve patient care.
#LCDbenchmark #ClinicalNLP
1
83
3/9 Dataset Highlights:
* Document classification task (Median word count: 1687) - predicting the future
* Sourced from MIMIC-IV discharge notes and statewide death data
* Evaluated with models from BoW and CNN to Hierarchical Transformers and GPT-4
(Figure: # of tokens)
2
109
2/9 Clinical documents are a rich source of information. Yet, most LMs/benchmarks struggle with their length. Our benchmark helps bridge this gap, ensuring models are assessed on real-world, lengthy clinical texts. Leaderboard (#CodaBench) available.
codabench.org/competitions/2…
2
169
We are thrilled to announce a new shared task 😃: Chemotherapy Timelines Extraction from Clinical Text.
Registration (& DUA) is open and training data will be released at the end of this week. sites.google.com/view/chemot…
The task will be hosted at #NAACL - #ClinicalNLP 2024.
Exciting news
We are organizing a shared task.
Chemotherapy Treatment Timelines Extraction from the Clinical Narrative (text mining task)
Collocated with NAACL 2024, Mexico City, Mexico. Do LLMs solve the task?
Check out chemotimelines2024.healthnlp…
4
669
We were delighted to host Prof. @TMills from @harvardmed to our invited talk session this morning! Thank you very much for the fantastic talk on the ClinicalNLP tasks! All of our lab members (and other participants from the department) enjoyed your talk very much!!
It was great to return to visit Prof. Jaewoo Kang's group at Korea University after three years to talk biomedical NLP. Thanks to Prof. Kang and his students (esp. @wonjin_info) for hosting me.
Glad to hear this awesome news!
This reminds me of my teenage memories... I used to play with Wolfram Alpha and it was one of my favorite toys.
Back then, I used to have a dream of becoming a mathematician (This dream ended immediately after the first year in Univ 😢😂)
New in the #WolframNetRepo: Extract text features with a BioBERT Trained on PubMed and PMC Data wolfr.am/OJ0cmwPm
Thanks @leejnhk, @wonjin_info and others for creating this model wolfr.am/T2TSCviO #NLP #neuralnetwork #ML #AI
Here is our way to combat COVID-19. Today we released the Open-domain QA system trained on #COVID19 related articles. CHECK covidsearch.korea.ac.kr We hope our work can help us to overcome this global crisis. #Team_DMIS #NLP #bionlp
covidsearch.korea.ac.kr/
Our first effort to build a real-time QA system on 31K COVID-19 articles with biomedical entities highlighted. It provides answers to all questions in the CORD-19 dataset by @allen_ai. Get your questions answered in just a second! #COVID19
6
WonJin Yoon retweeted
ICML 2020 (icml.cc/Conferences/2020) will be a virtual conference. We've been having discussions with ICLR folks about their plans, and will be learning from their experience (they are up first). We hope to enable as much of the normal ICML experience as possible, virtually.
Really excited to see BioASQ 8 is on the way. Our team will also participate in this team with a new strategy. Cheers to all the participants, and I look forward to meeting you all at the workshop! #bioasq #bionlp #team_dmis #dmis
The @BioASQ Task 8b is on!
Find the 1st batch of Phase A (#biomedical #InformationRetrieval) here participants-area.bioasq.org…
--------
New participants click here to find out how you can join in the #BioASQ #clef2020 Lab! participants-area.bioasq.org
#eHealth #BigData #QuestionAnswering
4
Our challenge-winning model (#sota for BioASQ task b -exact answers) is available now.
#github github.com/dmis-lab/bioasq-b…
#arxiv arxiv.org/abs/1909.08229
Pic) Me at #BioASQ workshop. Had a wonderful conversations! Thanks for the organizers and participants!!
Thanks for our team too!
My second paper on biomedical text processing, BioBERT is now available in arXiv and github!
arxiv.org/abs/1901.08746
Pre-trained model : github.com/naver/biobert-pre…
Fine-tuning codes : github.com/dmis-lab/biobert
#arxiv #BioBERT #BERT #nlp #bionlp #ner #re #qa
Our paper BioBERT (arxiv.org/abs/1901.08746) is now available in arxiv with pre-training weights and fine tuning codes! Pre-training on biomedical corpus helps a lot on multiple biomedical text mining tasks including NER, RE, QA, etc. :)

















