Open Life Science AI
@OpenlifesciAI
Followers
2K
Following
773
Media
2K
Statuses
2K
Daily tweets: Latest Medical LLMs, Datasets, Benchmarks, and Top Research Papers โก Life Science AI https://t.co/162AiaLdXn
Near to Nature
Joined May 2022
Open Medical -LLM Leaderboard with Hugginface collaboration โก๐๐งฌ.
Exciting news! ๐ข In collaboration with Hugging Face ๐ค, we are launching the Medical-LLM Leaderboard. This leaderboard aims to provide a standardized platform for evaluating large language models in the medical domain. It's encouraging to see tech companies like Google,
1
1
17
11/11 For daily insights into Medical AI breakthroughs:. ๐ฌ Discord: ๐ Substack: ๐บ YouTube: ๐ง Spotify: #MedicalAI #Innovation #HealthTech #FutureOfMedicine.
0
0
0
10/11 ๐งโโ๏ธ ๐ค๐๐ฎ๐น๐ถ๐ญ๐ฎ๐ญ๐ถ๐ฏ๐ฒ ๐๐ฑ๐ฎ๐บ๐ฝ๐น๐ฒ๐.The model justifies diagnoses like:. Hypertension*: based on ECG LVH + X-ray enlargement. Pneumonia*: CXR opacities + lab inflammation markers. ๐ก All responses use structured <think>. </think> and <answer>. </answer> formats.
1
0
0
9/11 ๐ฌ ๐๐ฏ๐น๐ฎ๐ญ๐ถ๐ผ๐ป ๐ฆ๐ญ๐๐ฑ๐ ๐๐ป๐๐ถ๐ด๐ต๐ญ๐.๐ง Removing any modality hurts performance.๐ Skipping pretraining or RFT reduces scores.โ
CMHA + CAO = best cross-modal integration.Result: multi-modal integration is ๐ฆ๐ด๐ด๐ฆ๐ฏ๐ต๐ช๐ข๐ญ for best outcomes in diagnosis.
1
0
0
8/11 ๐ ๐ฆ๐ญ๐ฟ๐ผ๐ป๐ด ๐๐ฒ๐ป๐ฐ๐ต๐บ๐ฎ๐ฟ๐ธ ๐ฅ๐ฒ๐๐๐น๐ญ๐.Compared against 8 top MLLMs:. โข ๐๐ถ๐ด๐ต๐ฒ๐๐ญ ๐๐๐๐จ, ๐ ๐๐ง๐๐ข๐ฅ, ๐ฅ๐ข๐จ๐๐, ๐๐๐ฅ๐ง๐ฆ๐ฐ๐ผ๐ฟ๐ฒ.โข ๐๐ฒ๐๐ญ ๐ฐ๐น๐ถ๐ป๐ถ๐ฐ๐ฎ๐น ๐ฒ๐ฏ๐ฎ๐น๐๐ฎ๐ญ๐ถ๐ผ๐ป ๐บ๐ฒ๐ญ๐ฟ๐ถ๐ฐ๐: F1 Score, Precision, AUC. Outperforms models like
1
0
0
7/11 ๐งช ๐ฅ๐ฒ๐ถ๐ป๐ณ๐ผ๐ฟ๐ฐ๐ฒ๐บ๐ฒ๐ป๐ญ ๐๐ถ๐ป๐ฒ-๐ง๐๐ป๐ถ๐ป๐ด ๐ฐ๐ถ๐ญ๐ต ๐๐ฅ๐ฃ๐ข.To boost clinical reasoning, MedTVT-R1 uses:. โข ๐ Group Relative Policy Optimization (GRPO).โข โ
๐๐ผ๐ฟ๐บ๐ฎ๐ญ ๐ฅ๐ฒ๐ฐ๐ฎ๐ฟ๐ฑ for structured responses.โข ๐ ๐๐ฎ๐ฐ๐ฐ๐ฎ๐ฟ๐ฑ ๐ฅ๐ฒ๐ฐ๐ฎ๐ฟ๐ฑ to optimize
1
0
0
6/11 ๐ง ๐๐ฟ๐ฐ๐ต๐ถ๐ญ๐ฒ๐ฐ๐ญ๐๐ฟ๐ฒ ๐๐ถ๐ด๐ต๐น๐ถ๐ด๐ต๐ญ๐.MedTVT-R1 introduces:. โข ๐ ๐๐๐ฐ๐น๐ถ๐ฐ ๐ ๐๐น๐ญ๐ถ-๐๐ฒ๐ฎ๐ฑ ๐๐ญ๐ญ๐ฒ๐ป๐ญ๐ถ๐ผ๐ป (๐๐ ๐๐) to allow rich modality interaction.โข โ๏ธ ๐๐ผ๐ป๐ญ๐ฟ๐ถ๐ฏ๐๐ญ๐ถ๐ผ๐ป-๐๐ฐ๐ฎ๐ฟ๐ฒ ๐ข๐ฝ๐ฒ๐ฟ๐ฎ๐ญ๐ผ๐ฟ (๐๐๐ข) to adaptively weight importance
1
0
0
5/11 ๐งฉ ๐๐ถ๐๐ฒ๐ฎ๐๐ฒ-๐๐ฒ๐ฏ๐ฒ๐น ๐๐ถ๐ฎ๐ด๐ป๐ผ๐๐ญ๐ถ๐ฐ ๐ฅ๐ฒ๐ฎ๐๐ผ๐ป๐ถ๐ป๐ด.MedTVT-R1 supports diagnoses of 7+ conditions including:. โข ๐ซ Coronary Artery Disease.โข ๐ง Hypertension.โข ๐ฆ Sepsis.โข ๐ซ Pneumonia.โข ๐ฉบ Atrial Fibrillation.โข ๐ Diabetes Mellitus.โข ๐งช Acute Renal
1
0
1
4/11 ๐งฌ ๐ฃ๐ต๐๐๐ถ๐ผ๐น๐ผ๐ด๐ถ๐ฐ๐ฎ๐น-๐๐ฒ๐ฏ๐ฒ๐น ๐จ๐ป๐ฑ๐ฒ๐ฟ๐๐ญ๐ฎ๐ป๐ฑ๐ถ๐ป๐ด.MedTVT-R1 is pretrained to deeply understand signals from each modality:. โข ECG rhythm types & conduction blocks.โข X-ray structural findings (e.g., opacities, enlargement).โข Lab panel abnormalities (e.g.,
1
0
0
3/11 ๐ฅ ๐ ๐ฒ๐ฑ๐ง๐ฉ๐ง-๐ค๐ ๐๐ฎ๐ญ๐ฎ๐๐ฒ๐ญ. โข Built from the MIMIC-IV family (ECG, CXR-JPG, LAB).โข Includes 8,706 ๐บ๐๐น๐ญ๐ถ๐บ๐ผ๐ฑ๐ฎ๐น ๐ฝ๐ฎ๐ญ๐ถ๐ฒ๐ป๐ญ ๐๐ฎ๐บ๐ฝ๐น๐ฒ๐.โข First dataset to offer ๐ค๐ ๐ฝ๐ฎ๐ถ๐ฟ๐ ๐ฎ๐ญ ๐ฝ๐ต๐๐๐ถ๐ผ๐น๐ผ๐ด๐ถ๐ฐ๐ฎ๐น ๐ฎ๐ป๐ฑ ๐ฑ๐ถ๐ฎ๐ด๐ป๐ผ๐๐ญ๐ถ๐ฐ ๐น๐ฒ๐ฏ๐ฒ๐น๐
1
0
0
2/11 ๐ง ๐ช๐ต๐ฎ๐ญ'๐ ๐ก๐ฒ๐ฐ ๐ถ๐ป ๐ ๐ฒ๐ฑ๐ง๐ฉ๐ง-๐ฅ1?. โข A ๐ ๐๐น๐ญ๐ถ๐บ๐ผ๐ฑ๐ฎ๐น ๐๐๐ (๐ ๐๐๐ ) that combines:. * ๐ซ ECG (time series). * ๐ท Chest X-rays (images). * ๐งช Lab Tests (tabular data).โข Enables ๐น๐ผ๐ป๐ด-๐ณ๐ผ๐ฟ๐บ, ๐บ๐๐น๐ญ๐ถ-๐ฑ๐ถ๐๐ฒ๐ฎ๐๐ฒ ๐ฑ๐ถ๐ฎ๐ด๐ป๐ผ๐๐ญ๐ถ๐ฐ
1
0
0
๐จ ๐ ๐ฒ๐ฑ๐ถ๐ฐ๐ฎ๐น ๐๐ ๐ฅ๐ฒ๐๐ฒ๐ฎ๐ฟ๐ฐ๐ต ๐๐น๐ฒ๐ฟ๐ญ! ๐จ. ๐๐ฎ๐ป ๐ฐ๐ฒ ๐ฏ๐๐ถ๐น๐ฑ ๐ฎ๐ป ๐๐ ๐ญ๐ต๐ฎ๐ญ ๐ถ๐ป๐ญ๐ฒ๐ด๐ฟ๐ฎ๐ญ๐ฒ๐ ๐๐๐, ๐ซ-๐ฟ๐ฎ๐๐, ๐ฎ๐ป๐ฑ ๐น๐ฎ๐ฏ ๐ญ๐ฒ๐๐ญ๐ ๐ณ๐ผ๐ฟ ๐ถ๐ป๐ญ๐ฒ๐ฟ๐ฝ๐ฟ๐ฒ๐ญ๐ฎ๐ฏ๐น๐ฒ ๐บ๐๐น๐ญ๐ถ-๐ฑ๐ถ๐๐ฒ๐ฎ๐๐ฒ ๐ฑ๐ถ๐ฎ๐ด๐ป๐ผ๐๐ถ๐?. Researchers from @HKUSTGuangzhou
2
6
14
23/23 For daily insights into Medical AI breakthroughs:. ๐ฌ Discord: ๐ Substack: ๐บ YouTube: ๐ง Spotify: #MedicalAI #Innovation #HealthTech #FutureOfMedicine.
1
0
0
22/23 ๐ฆ๐๐๐๐ฒ๐บ๐-๐ง๐ต๐ฒ๐ผ๐ฟ๐ฒ๐๐ถ๐ฐ ๐ฎ๐ป๐ฑ ๐๐ฎ๐๐ฎ-๐๐ฟ๐ถ๐๐ฒ๐ป ๐ฆ๐ฒ๐ฐ๐๐ฟ๐ถ๐๐ ๐๐ป๐ฎ๐น๐๐๐ถ๐ ๐ถ๐ป ๐ ๐-๐ฒ๐ป๐ฎ๐ฏ๐น๐ฒ๐ฑ ๐ ๐ฒ๐ฑ๐ถ๐ฐ๐ฎ๐น ๐๐ฒ๐๐ถ๐ฐ๐ฒ๐. This paper addresses serious cybersecurity risks in AI/ML-enabled medical devices, analyzing public data on recalls,
1
0
0
21/23 ๐๐ป๐ผ๐๐น๐ฒ๐ฑ๐ด๐ฒ-๐๐๐ด๐บ๐ฒ๐ป๐๐ฒ๐ฑ ๐ ๐๐น๐๐ถ๐บ๐ผ๐ฑ๐ฎ๐น ๐๐น๐ถ๐ป๐ถ๐ฐ๐ฎ๐น ๐ฅ๐ฎ๐๐ถ๐ผ๐ป๐ฎ๐น๐ฒ ๐๐ฒ๐ป๐ฒ๐ฟ๐ฎ๐๐ถ๐ผ๐ป ๐ณ๐ผ๐ฟ ๐๐ถ๐๐ฒ๐ฎ๐๐ฒ ๐๐ถ๐ฎ๐ด๐ป๐ผ๐๐ถ๐ ๐๐ถ๐๐ต ๐ฆ๐บ๐ฎ๐น๐น ๐๐ฎ๐ป๐ด๐๐ฎ๐ด๐ฒ ๐ ๐ผ๐ฑ๐ฒ๐น๐. This paper proposes ClinRaGen, a framework enhancing small
1
0
0
20/23 ๐๐ป ๐๐,๐๐๐-๐ฆ๐๐๐ฑ๐ ๐ข๐ฝ๐ฒ๐ป-๐๐ฐ๐ฐ๐ฒ๐๐ ๐๐ฎ๐๐ฎ๐๐ฒ๐ ๐ผ๐ณ ๐๐ผ๐ป๐ด๐ถ๐๐๐ฑ๐ถ๐ป๐ฎ๐น ๐ ๐ฎ๐ด๐ป๐ฒ๐๐ถ๐ฐ ๐ฅ๐ฒ๐๐ผ๐ป๐ฎ๐ป๐ฐ๐ฒ ๐๐บ๐ฎ๐ด๐ฒ๐ ๐ผ๐ณ ๐๐ฟ๐ฎ๐ถ๐ป ๐ ๐ฒ๐๐ฎ๐๐๐ฎ๐๐ฒ๐. This paper introduces an open-access dataset of 11,884 longitudinal brain MRI studies from
1
0
1
19/23 ๐๐ผ๐ต๐ผ๐ฟ๐ ๐๐ถ๐๐ฐ๐ผ๐๐ฒ๐ฟ๐: ๐ ๐ฆ๐๐ฟ๐๐ฒ๐ ๐ผ๐ป ๐๐๐ -๐๐๐๐ถ๐๐๐ฒ๐ฑ ๐๐น๐ถ๐ป๐ถ๐ฐ๐ฎ๐น ๐ง๐ฟ๐ถ๐ฎ๐น ๐ฅ๐ฒ๐ฐ๐ฟ๐๐ถ๐๐บ๐ฒ๐ป๐. This survey addresses the limited adoption of LLMs in clinical trial recruitment, specifically for trial-patient matching, despite their
1
0
0
18/23 ๐ฃ๐๐๐ฐ๐ต๐๐ฒ๐ป๐ฐ๐ต: ๐ ๐ฐ๐ผ๐บ๐ฝ๐ฟ๐ฒ๐ต๐ฒ๐ป๐๐ถ๐๐ฒ ๐ฎ๐ป๐ฑ ๐ฝ๐ฟ๐ผ๐ณ๐ฒ๐๐๐ถ๐ผ๐ป๐ฎ๐น ๐ฏ๐ฒ๐ป๐ฐ๐ต๐บ๐ฎ๐ฟ๐ธ ๐ณ๐ผ๐ฟ ๐ฒ๐๐ฎ๐น๐๐ฎ๐๐ถ๐ป๐ด ๐๐ต๐ฒ ๐ฝ๐ฒ๐ฟ๐ณ๐ผ๐ฟ๐บ๐ฎ๐ป๐ฐ๐ฒ ๐ผ๐ณ ๐๐๐ -๐ฎ๐๐๐ถ๐๐๐ฒ๐ฑ ๐ฝ๐๐๐ฐ๐ต๐ถ๐ฎ๐๐ฟ๐ถ๐ฐ ๐ฐ๐น๐ถ๐ป๐ถ๐ฐ๐ฎ๐น ๐ฝ๐ฟ๐ฎ๐ฐ๐๐ถ๐ฐ๐ฒ. To address the lack of a
1
0
1
17/23 ๐ฅ๐ฎ๐ฑ๐๐ฎ๐ฏ๐ฟ๐ถ๐ฐ: ๐๐ด๐ฒ๐ป๐๐ถ๐ฐ ๐๐ ๐ฆ๐๐๐๐ฒ๐บ ๐๐ถ๐๐ต ๐ฅ๐ฒ๐ฎ๐๐ผ๐ป๐ถ๐ป๐ด ๐๐ฎ๐ฝ๐ฎ๐ฏ๐ถ๐น๐ถ๐๐ ๐ณ๐ผ๐ฟ ๐ฅ๐ฎ๐ฑ๐ถ๐ผ๐น๐ผ๐ด๐. This paper proposes RadFabric, a multi-agent, multimodal reasoning framework using the Model Context Protocol (MCP) to address limitations in
1
0
0
16/23 ๐จ๐ป๐ถ๐๐ฒ๐ฟ๐๐ฎ๐น ๐๐ฎ๐ฏ๐ผ๐ฟ๐ฎ๐๐ผ๐ฟ๐ ๐ ๐ผ๐ฑ๐ฒ๐น: ๐ฝ๐ฟ๐ผ๐ด๐ป๐ผ๐๐ถ๐ ๐ผ๐ณ ๐ฎ๐ฏ๐ป๐ผ๐ฟ๐บ๐ฎ๐น ๐ฐ๐น๐ถ๐ป๐ถ๐ฐ๐ฎ๐น ๐ผ๐๐๐ฐ๐ผ๐บ๐ฒ๐ ๐ฏ๐ฎ๐๐ฒ๐ฑ ๐ผ๐ป ๐ฟ๐ผ๐๐๐ถ๐ป๐ฒ ๐๐ฒ๐๐๐. This paper addresses predicting abnormal clinical laboratory values from performed diagnostic tests,
1
0
0
15/23 ๐จ๐ป๐น๐ผ๐ฐ๐ธ๐ถ๐ป๐ด ๐ก๐ฒ๐๐ฟ๐ฎ๐น ๐ง๐ฟ๐ฎ๐ป๐๐ฝ๐ฎ๐ฟ๐ฒ๐ป๐ฐ๐: ๐๐ฎ๐ฐ๐ผ๐ฏ๐ถ๐ฎ๐ป ๐ ๐ฎ๐ฝ๐ ๐ณ๐ผ๐ฟ ๐๐
๐ฝ๐น๐ฎ๐ถ๐ป๐ฎ๐ฏ๐น๐ฒ ๐๐ ๐ถ๐ป ๐๐น๐๐ต๐ฒ๐ถ๐บ๐ฒ๐ฟ'๐ ๐๐ฒ๐๐ฒ๐ฐ๐๐ถ๐ผ๐ป. This paper proposes a novel pre-model approach using Jacobian Maps (JMs) within a multi-modal framework to
1
0
0