Twitxy Saved

Frazier Huo

@Zepeng_Huo

Joined November 2015
280 Following 23 Followers 14 Posts

Latest posts 5

Frazier Huo @Zepeng_Huo
๐Ÿš€ Public releases ๐˜๐—ต๐—ฟ๐—ฒ๐—ฒ ๐—ฑ๐—ฒ-๐—ถ๐—ฑ๐—ฒ๐—ป๐˜๐—ถ๐—ณ๐—ถ๐—ฒ๐—ฑ ๐—น๐—ผ๐—ป๐—ด๐—ถ๐˜๐˜‚๐—ฑ๐—ถ๐—ป๐—ฎ๐—น ๐—˜๐—›๐—ฅ ๐—ฑ๐—ฎ๐˜๐—ฎ๐˜€๐—ฒ๐˜๐˜€ (EHRSHOT, INSPECT, MedAlign) with 25,991 patients & 295M clinical events (1997-2023)โ€”now freely available for ๐š—ฬฒ๐š˜ฬฒ๐š—ฬฒโ€“ฬฒ๐šŒฬฒ๐š˜ฬฒ๐š–ฬฒ๐š–ฬฒ๐šŽฬฒ๐š›ฬฒ๐šŒฬฒ๐š’ฬฒ๐šŠฬฒ๐š•ฬฒโ€‚ฬฒ๐™ฐฬฒ๐™ธฬฒโ€‚ฬฒ๐š›ฬฒ๐šŽฬฒ๐šœฬฒ๐šŽฬฒ๐šŠฬฒ๐š›ฬฒ๐šŒฬฒ๐š‘ฬฒ in healthcare!

These datasets support ๐—บ๐˜‚๐—น๐˜๐—ถ๐—บ๐—ผ๐—ฑ๐—ฎ๐—น ๐—น๐—ฒ๐—ฎ๐—ฟ๐—ป๐—ถ๐—ป๐—ด, ๐˜๐—ถ๐—บ๐—ฒ-๐˜๐—ผ-๐—ฒ๐˜ƒ๐—ฒ๐—ป๐˜ ๐—บ๐—ผ๐—ฑ๐—ฒ๐—น๐—ถ๐—ป๐—ด, and ๐—ณ๐—ฒ๐˜„-๐˜€๐—ต๐—ผ๐˜ ๐˜๐—ฎ๐˜€๐—ธ๐˜€ to advance responsible AI in chronic disease and cancer care.
Jason Alan Fries Verified account @jasonafries Feb 12, 2025
๐ŸŽ‰ We're thrilled to announce the general release of three de-identified, longitudinal EHR datasets from Stanford Medicineโ€”now freely available for non-commercial research-use worldwide! ๐Ÿš€

Read our HAI blog post for more details: hai.stanford.edu/news/advancingโ€ฆ

๐——๐—ฎ๐˜๐—ฎ๐˜€๐—ฒ๐˜ ๐—ฆ๐˜‚๐—บ๐—บ๐—ฎ๐—ฟ๐—ถ๐—ฒ๐˜€
๐Ÿ“Š 3 longitudinal EHR datasets
๐Ÿ‘ฅ Scale: 25,991 patients, 441,680 visits, and 295M clinical events (median: 4,882 events per patient)
โณ Timeframe: Patient trajectories from 1997 to 2023 (median: 10 years per patient)

๐—ฆ๐˜๐—ฎ๐—ป๐—ฑ๐—ฎ๐—ฟ๐—ฑ๐—ถ๐˜‡๐—ฒ๐—ฑ ๐—•๐—ฒ๐—ป๐—ฐ๐—ต๐—บ๐—ฎ๐—ฟ๐—ธ ๐—ง๐—ฎ๐˜€๐—ธ๐˜€
๐ŸŽฏ Few-shot Learning
๐Ÿค– Multimodal Learning & Time-to-Event Modeling
โŒ› Long Context Instruction Following & Temporal Reasoning

Thanks to @MichaelWornow, Ethan Steinberg, @Zepeng_Huo, @HennyJieCC , @BediSuhana42170 , @AlyssaUnell, @drnigam @StanfordMed
View quoted post on X
Frazier Huo @Zepeng_Huo
๐ŸŽ‰ Excited to share that our latest research, ๐˜›๐˜ช๐˜ฎ๐˜ฆ-๐˜ต๐˜ฐ-๐˜Œ๐˜ท๐˜ฆ๐˜ฏ๐˜ต ๐˜—๐˜ณ๐˜ฆ๐˜ต๐˜ณ๐˜ข๐˜ช๐˜ฏ๐˜ช๐˜ฏ๐˜จ ๐˜ง๐˜ฐ๐˜ณ 3๐˜‹ ๐˜”๐˜ฆ๐˜ฅ๐˜ช๐˜ค๐˜ข๐˜ญ ๐˜๐˜ฎ๐˜ข๐˜จ๐˜ช๐˜ฏ๐˜จ, has been accepted at ๐—œ๐—–๐—Ÿ๐—ฅ 2025! ๐Ÿš€

๐Ÿ” ๐—œ๐—บ๐—ฝ๐—ฟ๐—ผ๐˜ƒ๐—ถ๐—ป๐—ด ๐— ๐—ฒ๐—ฑ๐—ถ๐—ฐ๐—ฎ๐—น ๐—œ๐—บ๐—ฎ๐—ด๐—ฒ ๐—ฃ๐—ฟ๐—ฒ๐˜๐—ฟ๐—ฎ๐—ถ๐—ป๐—ถ๐—ป๐—ด ๐˜„๐—ถ๐˜๐—ต ๐—ง๐—ถ๐—บ๐—ฒ-๐˜๐—ผ-๐—˜๐˜ƒ๐—ฒ๐—ป๐˜ (๐—ง๐—ง๐—˜) ๐— ๐—ผ๐—ฑ๐—ฒ๐—น๐—ถ๐—ป๐—ด
While self-supervised methods in medical imaging have significantly enhanced diagnostic accuracy and image segmentation performance, they struggle with ๐—ฝ๐—ฟ๐—ผ๐—ด๐—ป๐—ผ๐˜€๐—ถ๐˜€โ€”the prediction of future health outcomes. Reliable prognosis is essential for assessing disease progression and guiding clinical decision-making. Our framework addresses this gap by combining time-to-event (TTE) modeling with longitudinal EHRs to pretrain a 3D image encoder for outcome prediction, significantly improving prognostic prediction using only imaging data.

๐Ÿ”ฅ Key Highlights
โ€ข Time-to-event (TTE) pretraining: Predicts time until critical clinical events by learning 3D imaging biomarkers.
โ€ข Massive scale: Trained on 8,192 TTE tasks and 18,945 CT scans linked to EHR data containing 225M clinical eventsโ€”the largest paired EHR+3D imaging research dataset currently available
โ€ข Superior prognostic performance: ๐Ÿฎ๐Ÿฏ.๐Ÿณ% ๐—”๐—จ๐—ฅ๐—ข๐—– ๐—ฏ๐—ผ๐—ผ๐˜€๐˜, ๐Ÿฎ๐Ÿต.๐Ÿฐ% ๐—›๐—ฎ๐—ฟ๐—ฟ๐—ฒ๐—น๐—นโ€™๐˜€ ๐—–-๐—ถ๐—ป๐—ฑ๐—ฒ๐˜… ๐—ถ๐—บ๐—ฝ๐—ฟ๐—ผ๐˜ƒ๐—ฒ๐—บ๐—ฒ๐—ป๐˜, ๐—ฎ๐—ป๐—ฑ ๐Ÿฑ๐Ÿฐ% ๐—ฏ๐—ฒ๐˜๐˜๐—ฒ๐—ฟ ๐—ฐ๐—ฎ๐—น๐—ถ๐—ฏ๐—ฟ๐—ฎ๐˜๐—ถ๐—ผ๐—ป without sacrificing diagnostic accuracy.

๐Ÿ› ๏ธ How It Works
โ€ข Transform patient EHR timelines into large-scale TTE pretraining tasks, predicting time distributions until key medical events.
โ€ข A 3D vision encoder processes CT scans, generating embeddings for a task-specific TTE head (e.g., Cox models, survival networks).
โ€ข Enables AI to capture long-term health trajectories for better risk stratification and survival analysis.

๐Ÿ“– Read more
๐Ÿ“„ Paper: lnkd.in/gfnGrNgS
๐Ÿ—‚๏ธ Dataset: EHR lnkd.in/gAxwvXdw Imaging lnkd.in/gWKPrPir
๐Ÿ’ป Code: Coming Soon!
๐Ÿค— Hugging Face Models: Coming Soon!
๐Ÿ™Œ Huge thanks to my amazing co-leads: @jasonafries , @Ale9806_ and co-authors @jeyamariajose,ย Ethan Steinberg, @loublanks, @Dr_ASChaudhari, @curtlanglotz, @drnigam

Excited to push medical imaging AI forwardโ€”stay tuned for tutorials & deep dives! ๐Ÿฅ๐Ÿ”ฌ hashtag#ICLR2025 hashtag#AI hashtag#MedicalImaging hashtag#TTE hashtag#SurvivalAnalysis hashtag#MultiModalAI
1 replies 4 reposts 16 likes 3.3K views View on X
Frazier Huo @Zepeng_Huo
Longitudinal EHR data in such large scale is perfect to build your own EHR foundation model!
Michael Wornow @MichaelWornow Jun 10, 2024
Excited to announce the official full release of ๐Ÿ‘‚๐Ÿ’‰EHRSHOT -- a dataset of 6,739 deidentified longitudinal EHRs for few-shot eval of foundation models!

๐ŸŒWebsite: ehrshot.stanford.edu
๐ŸฅDataset: redivis.com/datasets/53gc-โ€ฆ
๐Ÿค—Model: huggingface.co/StanfordShahLaโ€ฆ
View quoted post on X
Frazier Huo @Zepeng_Huo
Excited to publish dataset for #OpenScience ๐Ÿค—
Jason Alan Fries Verified account @jasonafries Dec 12, 2023
We're excited to introduce #INSPECT a large-scale โœจ3D multimodalโœจmedical imaging dataset #NeurIPS2023

19,402 Stanford Medicine Patients
๐Ÿฉป23,248 CT Scans + ๐Ÿ“„Paired Radiology Notes
๐Ÿ“ˆLongitudinal EHRs
๐ŸฉบClinician-validated task labels
#DataCentricAI #Multimodal #3Dimaging

1/
View quoted post on X