Krithik Vishwanath

AI Researcher in healthcare

[Krithik Vishwanath @krithikvish]

@manoelribeiro @hayounggjung @korolova Loved reading this, cool work!

[Krithik Vishwanath @krithikvish]

@beaulieujones @AdamRodmanMD @EricTopol @openevidence @UpToDate We did have clinicians filter non-clinical queries. I agree the inability to release RCQ limits independent verification of composition; so readers have to evaluate the process we report. Happy to chat offline; we all have the same goal of improving patient care.

[Krithik Vishwanath @krithikvish]

@beaulieujones @AdamRodmanMD @EricTopol @openevidence @UpToDate Example aside, I personally believe that queries from real clinical use are way more representative of clinical impact @ POC than other popular benchmarks in medicine (E.g., MedQA and HealthBench are the most popular evaluation tools out there in the literature).

[Krithik Vishwanath @krithikvish]

@beaulieujones @AdamRodmanMD @EricTopol @openevidence @UpToDate Totally understand, but preserving the ecological usage of models was a priority for us, as the evidence base is thin on this. Stratifying by certain use-cases should be done too and am looking forward to reading analyses. Difficult to do in 100 Qs.

[Krithik Vishwanath @krithikvish]

@VivekSubbiah @NaturePortfolio @Nature_NPJ @Nature @weldeiry @DrChoueiri @neerajaiims @OncBrothers @oncodaily @crisbergerot @PTarantinoMD Great paper! This is an excellent utility point for LLM systems in healthcare

[Krithik Vishwanath @krithikvish]

@beaulieujones @AdamRodmanMD @EricTopol @openevidence @UpToDate Hm, RCQ was built from queries physicians actually asked during live patient care. Is your point that if they used OE instead of the HIPPA compliant GPT-instance, they would have asked different questions?

[Krithik Vishwanath @krithikvish]

@katy_beckermann @iScienceLuvr 1) Not possible for our independent dataset (pulled from a live clinical environment, we had to scrub it of PHI)
2) Hospital systems can choose to use API for frontier models and set up their own UI... they can't for OE/U2D

[Krithik Vishwanath @krithikvish]

@rajivmovva Honestly not sure! The AI companies have an army; e.g., see OpenAI's HealthBench briefing which stated they had input from 262 physicians across the world. The most important metric is the impact to patient health, not necessarily doctor preference. Everything else is a proxy

[Krithik Vishwanath @krithikvish]

@MFGensheimer The questions varied greatly in scope/kind (as was expected from a live clinical environment). Qualitatively, clinician evaluators who were grading responses felt that many of these questions were "google-able", though the example presented is not fully of this class.

[Krithik Vishwanath @krithikvish]

@EricTopol @ekoermann @nyulangone Honored by the shoutout! Hoping that this work encourages/motivates building a stronger evidence base when deploying tools in medicine.

[Krithik Vishwanath @krithikvish]

@rajivmovva I would argue that case can be made for any of these companies (they all have tons of MDs in house)? The amount of experts on your team isn't a replacement for evidence-based practice in healthcare, though it might improve your odds at building a calibrated product.

[Eric Topol @EricTopol]

For medical information, general AI frontier models (Google, OpenAI, Anthropic) outperformed specialized @openevidence and @UpToDate as assessed by 12 US clinicians, randomized and blinded to which model and extensive testing/benchmarks. This was not anticipated. @NatureMedicine
https://t.co/KCH1ADfQWz

[The Daily Texan @thedailytexan]

Krithik Vishwanath, a @UTAustin computational engineering, chemistry and mathematics senior, received the Churchill Scholarship, the first awarded to a UT student in 36 years, for his research in computational medicine models. The scholarship funds a year of master’s study in the United Kingdom for students demonstrating excellence in STEM.

https://t.co/Y8jyXko4aL

[Claude @claudeai]

Computer use is now in Claude Code.

Claude can open your apps, click through your UI, and test what it built, right from the CLI.

Now in research preview on Pro and Max plans.