Value Health Voices

There are a handful of genuinely useful data points - only 7% of NCCN recommendations derive from Phase 3 RCTs, MedNet AI surfaces community Q&A 80% of the time, LLMs answer complex oncology questions accurately ~30% of the time - but these are spaced far apart across 56 minutes of personal anecdotes, restatements, and affirmations. The insight-per-minute rate is low.

“only like, I think it's like 7% of the recommendations are come from phase three randomized trials”

“80% of them actually result in Q and A from MedNet. So 80% of the time your question is pulling from the experience and the knowledge of the community”

The comparison of pediatric vs. adult oncologist default mindset on clinical trials is a genuinely non-obvious insight, and hard-coding 'the answer is not known' into their AI is a meaningful design distinction. However, the broader framing - LLMs hallucinate on complex cases, medicine is art and science, AI should assist not replace - is thoroughly recycled discourse.

“the first question a pediatric oncologist asks themselves when they see a new patient is what trial do I have for this patient? That's like maybe first question adult oncologist asks is what's first line, what's second line, what's third line?”

“we actually designed it to say like that the answer is not known like not to extrapolate”

Dr. Husri is a legitimate operator-practitioner: a radiation oncologist who co-founded and has run MedNet for 12 years, received NSF and NCI grants, went through Y Combinator, and was invited to the ASTRO Presidential Symposium. She speaks from genuine build experience, not punditry, though the company remains relatively small and niche.

“we went through Y Combinator and had funding from, from them”

“we have a team of. Right now we're six physicians at MedNet. So it's. Everything that we develop is product engineering, medical”

Several concrete metrics are cited - 7% NCCN Phase 3 evidence rate, 80% community Q&A pull rate, ~30% LLM accuracy on complex oncology queries, 73-day knowledge doubling time - but sourcing is loose throughout, the 30% figure is introduced by the host without a named study, and many claims about AI limitations and physician behavior are asserted without grounding.

“large language models performing in terms of completely accurate responses for difficult oncologic questions. Something like 30% or less”

“80% of them actually result in in Q and A from MedNet”

The hosts did real preparation - they surface the ASTRO 2022 predictions, cite a specific study, and use a structured three-prediction framework - but the personal friendship dynamic produces consistent softballing, the hosts frequently complete the guest's sentences or answer questions themselves, and there is no meaningful pushback on any claim throughout the episode.

“Were you worried at the time or even now, I guess that with all these companies doing these investments in LLMs in this space, like you said, they're kind of going almost head to head a little bit against what you are providing”

“That's. It's a great example”