Neurology resident interested in the intersection of health and artificial intelligence: digital biomarkers, pragmatic trials, wearables, reproducible research
Joined January 2011
- Tweets 130
- Following 93
- Followers 154
- Likes 388
31 Photos and videos
Pinned Tweet
9 Jun 2024
#LocalCitationNetwork now allows you to retrieve All References and All Citations from a given set of Input Articles with both @OpenAlex_org and @SemanticScholar! For example, these 11 input articles (via @vamrhein) have 406 references & 7143 citations: localcitationnetwork.github.… 1/
1
10
14
1,539
🚀 Excited to share our latest Journal of Neurology publication on dreaMS app.
Six gamified, adaptive cognitive tests (<10 min) improve sensitivity to change by addressing floor/ceiling & practice effects.
Big thanks to our team & partners!
Read more: link.springer.com/article/10…
3
5
188
Tim Woelfle retweeted
12 Sep 2024
Human-AI collaboration may save time for a second human rater for reporting and bias assessments.
We tested Claude-3-Opus, Claude-2, GPT-4, GPT-3.5, Mixtral-8x22B.
Wonderful work led by @timwoelfle published in @JClinEpi
jclinepi.com/article/S0895-4…
1
4
16
1,657
12 Sep 2024
Our work "Benchmarking Human-AI Collaboration for Common Evidence Appraisal Tools" is published in @JClinEpi! doi.org/10.1016/j.jclinepi.2…
Evidence appraisal tools are very resource intensive but LLMs may assist human raters. Wonder how @OpenAI's o1 & @Meta's Llama 3.1 will perform?
23 Apr 2024
Check out our work on LLMs for systematic reviews of medical literature: Benchmarking Human-AI Collaboration for Common Evidence Appraisal Tools. We used @AnthropicAI's Claude-3-Opus, @OpenAI's GPT-4, @MistralAI's open-source Mixtral-8x22B: medrxiv.org/content/10.1101/… @LGHemkens 1/6
5
6
879
Tim Woelfle retweeted
9 Jun 2024
Shall we please stop worrying about rogue AI and instead worry about the Atlantic Overturning Circulation crossing a tipping point. It seems close and would make Europe basically unlivable. (Thanks to @jonkhler for the link) youtu.be/ZHNNW8c_FaA?si=hzPW…
9 Jun 2024
Just finished reading Aschenbrenner's manifesto (165p) about the impending intelligence explosion. I'm now rethinking my life plans. (Summary to follow on YT) situational-awareness.ai/
9
28
216
43,455
Tim Woelfle retweeted
7 Jun 2024
Why were so few RCTs done to find out optimal COVID control strategies (masks, isolation)? Why so few RCTs of educational strategies? We conduct uncontrolled experiments over & over, remain in the dark. Cultural change to accept RCTs outside conventional medicine urgently needed.
3
17
57
10,643
9 Jun 2024
#LocalCitationNetwork now allows you to retrieve All References and All Citations from a given set of Input Articles with both @OpenAlex_org and @SemanticScholar! For example, these 11 input articles (via @vamrhein) have 406 references & 7143 citations: localcitationnetwork.github.… 1/
1
10
14
1,539
9 Jun 2024
#LocalCitationNetwork will always remain free & open source, meaning 100% transparency! There are many other great literature mapping tools like @Inciteful_xyz (also open source), but most are closed source: @RsrchRabbit, @LitmapsApp, @ConnectedPapers 6/ x.com/RaziaAliani/status/179…
13 May 2024
You enter keywords on Google Scholar.
Then bam! Thousands of hits.
Instead, use AI literature mapping tools!
SAVE this guide to choose the right one for you ⤵
Sifting through the Google Scholar/ PubMed noise takes hours.
DAYS even.
And don't get me started on the compilations.
Endless datasheets.
Do yourself a favor and
Use AI literature mapping tools
They analyze and visualize scientific literature for you.
Just input your seed paper or collection.
The AI recommends similar papers.
Ones ACTUALLY relevant to your search.
You can see it all on an interactive map or graph.
BUT..
How to decide the right tool for your use case?
⤴ That's why I created this comparison table for you
---------------------------------------------------------
#aiinresearch #literaturereview #ai #literaturemapping
@RsrchRabbit @LitmapsApp @Inciteful_xyz
1
1
5
1,150
9 Jun 2024
Finally, check out this recent guidance on citation searching in the @bmj_latest: bmj.com/lookup/doi/10.1136/b… Direct citation searching is now fully implemented in #LocalCitationNetwork & we're working on indirect citation searching: doi.org/10.17605/OSF.IO/NPM2… Stay tuned! 7/7
3
66
Tim Woelfle retweeted
9 May 2024
Recommendations for researchers on when and how to conduct citation searching and how to report it
bmj.com/content/385/bmj-2023…
2
12
12
9,638
Tim Woelfle retweeted
27 Apr 2024
As long as AI systems are trained to reproduce human-generated data (e.g. text) and have no search/planning/reasoning capability, performance will saturate below or around human level.
Furthermore, the amount of trials needed to reach that level will be far larger than the amount of trials needed to train humans.
LLMs are trained with 200,000 years worth of reading material and are still pretty dumb.
Their usefulness resides in their vast accumulated knowledge and language fluency. But they are still pretty dumb.
26 Apr 2024
Interesting how in all these domains AI is asymptoting at roughly human performance - where's the AI zooming past us to superintelligence that Kurzweil etc. predicted/feared?
234
732
4,160
824,633
26 Apr 2024
Great study benchmarking LLMs on clinical oncology questions! They employ some similar techniques as we do, in particular the consistency approach on repeated prompts. The self-assessed confidence is a very interesting approach I'd like to see more in the future.
Original Article: Comparative Evaluation of LLMs in Clinical Oncology nejm.ai/4aJWOAY
ALT Figure 3 from the NEJM AI Original Article "Comparative Evaluation of LLMs in Clinical Oncology": Self-Assessed Confidence Has Discriminatory Power in High-Performing Models. In each question prompt, the models were asked to evaluate their confidence (from 1 to 4) in the response, in which 1 represented minimal confidence (i.e., a random guess) and 4 represented maximal confidence. The self-assessed confidence score had discriminatory power for PaLM 2, generative pretrained transformer 3.5 (GPT-3.5), Claude-v1, and GPT-4. B denotes billion.
2
129
Tim Woelfle retweeted
23 Apr 2024
The study I was waiting for (and knew would be done).
Echoes my (disappointing) experience of Cochrane RoB using GPT4
And don’t ask it to do anything around data integrity checks!
23 Apr 2024
We tested how we can best collaborate with AI to do systematic reviews, meta-research or asses study designs - fantastic team and teamwork, thank you @timwoelfle et al!!
1
3
9
1,637
Tim Woelfle retweeted
24 Apr 2024
”Current LLMs alone appraised evidence worse than humans. Human-AI collaboration may reduce workload for the second human rater for the assessment of reporting (PRISMA) and methodological rigor (AMSTAR) but not for complex tasks such as PRECIS-2.” #EBM #AI
23 Apr 2024
We tested how we can best collaborate with AI to do systematic reviews, meta-research or asses study designs - fantastic team and teamwork, thank you @timwoelfle et al!!
1
4
1,001
Tim Woelfle retweeted
23 Apr 2024
We tested how we can best collaborate with AI to do systematic reviews, meta-research or asses study designs - fantastic team and teamwork, thank you @timwoelfle et al!!
23 Apr 2024
Check out our work on LLMs for systematic reviews of medical literature: Benchmarking Human-AI Collaboration for Common Evidence Appraisal Tools. We used @AnthropicAI's Claude-3-Opus, @OpenAI's GPT-4, @MistralAI's open-source Mixtral-8x22B: medrxiv.org/content/10.1101/… @LGHemkens 1/6
2
4
16
4,256
Tim Woelfle retweeted
23 Apr 2024
Fantastic study (and great research methodology) on the abilities of LLMs to perform evidence appraisal. Certainly something a lot of us have been hoping for.
TL;DR: humans outperform LLMs alone, but human AI performs quite well in some settings.
23 Apr 2024
Check out our work on LLMs for systematic reviews of medical literature: Benchmarking Human-AI Collaboration for Common Evidence Appraisal Tools. We used @AnthropicAI's Claude-3-Opus, @OpenAI's GPT-4, @MistralAI's open-source Mixtral-8x22B: medrxiv.org/content/10.1101/… @LGHemkens 1/6
4
12
3,350
23 Apr 2024
Our >2000 API calls made full use of context lengths >16k tokens. Unfortunately, the current 8k context length of @metaAI 's promising Llama3 is too short. For @AnthropicAI's multimodal Claude-3-Opus, we converted PDFs to >1500 PNGs (one per page), uploading ~2 GB of images. 5/
2
2
160
23 Apr 2024
Our code & data are fully open source and the framework is easily extendable! Check out our streamlined pipeline to add new LLMs and our interactive dashboards using @rmarkdown: github.com/timwoelfle/Eviden… 6/6
2
119