LLMs Stumble on Skin Cancer, but the Real Pain in Healthcare Is Paperwork

I stumbled on this piece from the Lynch Lab making the rounds on Hacker News, and the headline does a lot of work: a 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis. The author, sungam, lays out some stark numbers—fine-tuned ResNet-50 hits 92% accuracy, while GPT-4V manages 72% and misses one in five melanomas. That's a real gap. And it's not just about accuracy: the LLMs also generate plausible-sounding but wrong explanations, which is arguably worse in a clinical setting.

But here's the thing. If you're a builder eyeing healthcare, the takeaway isn't "LLMs are useless in medicine." It's that we keep trying to make LLMs do something they're bad at—fine-grained image recognition—while ignoring the mountain of pain they could actually solve. PainSignal tracks 1,209 problems in healthcare right now. The most severe ones aren't "AI can't read a dermoscopy image." They're about documentation, workflow, and safety. That's where the opportunity is.

The article's core claim checks out. CNNs are built for image classification; LLMs are built for language. No surprise there. But the framing that LLMs aren't useful in healthcare at all is where I'd push back. Yes, they hallucinate. Yes, they're not ready to be your dermatologist. But they're already good enough to draft a consult note, summarize a patient's chart, or flag a missing lab order. And the market is screaming for that.

Look at the severity data. Documentation and workflow problems consistently rate 5/5. Problems like 'MedLog Sync' and 'VitalLink Alert' aren't about diagnostic accuracy—they're about manual data entry, slow communication, and fear of missing a critical change. A nurse spending two hours a shift on paperwork isn't helped by a better image classifier. She's helped by an LLM that turns a rambling voice note into a structured note. That's a 10x improvement in quality of life, and it's achievable today.

Then there are the app ideas already emerging from this demand. 'ChartWise AI' and 'NurseFlow: Admin Reducer' are exactly the kind of LLM-powered tools that attack the documentation burden. They don't compete with a ResNet. They complement it. The dermatologist still needs an accurate image model to catch the melanoma, but she also needs to write the follow-up letter, schedule the biopsy, and update the patient's record. Right now, she does all that by hand.

So the real story isn't "old CNN vs new LLM." It's that we have two tools optimized for different jobs. One is great at pixels; the other is great at prose. The healthcare system is drowning in prose—and in the absence of tools to handle it, clinicians drown too. That's why we track 614 healthcare app ideas, most of them aimed at reducing friction, not replacing diagnosis.

Builders take note. If you're a vibe coder looking for a high-impact project, don't try to train a dermatology model. You'll lose to a 2015 architecture. Instead, pick one of those 5/5 severity workflow problems and build a boring, reliable LLM wrapper that saves a nurse 30 minutes a shift. It's less sexy, sure. But it'll actually get used. And in healthcare, usage is the only metric that matters.

The article from Lynch Lab is worth a read—it's a solid reminder that specialized models still dominate niche tasks. But the conclusion shouldn't be doom for LLMs. It should be a redirection. Stop trying to make the generalist do the specialist's job, and start handing it the paperwork. That's where the real pain lives.