Our Take
This is a patient using an LLM as a literature-search and pattern-matching tool for rare disease, not as a replacement for oncology; the real story is how fragmented medical data and low-frequency diagnoses create a gap that general-purpose AI can fill today.
Why it matters
One-third of American adults now use chatbots for health information (per March public opinion poll). Rare-disease patients—especially those seeing specialists who encounter a condition once yearly—are finding AI useful where the system leaves blind spots.
Do this week
Healthcare leaders: audit your end-of-treatment imaging workflows for high false-positive rates and flag them to your clinical informatics team before AI tools do it for your competitors.
A founder diagnosed with non-Hodgkin's lymphoma used Claude to catch a diagnostic error
Connor Christou, 35, founder of Keragon (an AI platform for medical practice administration), was diagnosed in 2025 with an aggressive, fast-growing non-Hodgkin's lymphoma, a rare cancer affecting roughly one in 420,000 people. He caught it by accident: arm swelling prompted a visit for blood clots, but pre-op imaging revealed an 11-by-11-by-8 centimeter tumor behind his sternum. A biopsy confirmed the diagnosis. He was within three weeks of stage four.
Faced with conflicting advice from two world-class oncologists—one recommending a lighter chemotherapy regimen (60% success rate for his pathology), the other the aggressive option (85% success rate)—Christou sought 12 opinions over two days. Eleven of twelve voted for the harder path: continuous in-hospital infusion across six monthly cycles.
During treatment, Christou logged every variable he could measure: sleep (via Whoop band), blood work, scan data, symptom journal entries via voice transcription, and medication adjustments. He fed all of it into Claude.
The critical moment came at the end of treatment. His final PET scan (positron emission tomography, used to detect active disease) came back ambiguous. His oncologist discussed a second line of therapy, potentially radiotherapy near his heart and lungs. Christou researched the false-positive rate for end-of-treatment PET scans in his specific lymphoma type and found it was around 60%.
He uploaded his three PET scans and MRI into Claude. The model flagged a known but often-overlooked phenomenon: in patients under 40 recovering from this lymphoma subtype, the thymus gland (which sits near the heart) can reactivate after chemotherapy, appearing on imaging as active disease. Given his age and scan characteristics, Claude estimated the probability of thymus rebound at roughly 90%.
Christou sought three additional opinions. The fourth doctor confirmed it: thymus rebound, no active disease, no radiotherapy needed.
Rare disease diagnoses expose the limits of individual specialist knowledge
A oncologist might encounter this specific lymphoma once per year. An individual doctor's cumulative caseload, no matter how experienced, cannot match the breadth of medical literature an LLM has absorbed. Christou's case is not an isolated use of AI for health information: a public opinion poll released in March found that one-third of American adults now use chatbots for health information and advice.
Experts flag the danger. Danielle Bitterman, clinical lead for data science and AI at Mass General Brigham, told the New York Times that general-purpose chatbots are frequently wrong and "have not been thoroughly evaluated" for personalized diagnoses. Christou agrees: "It didn't replace the doctors," he says, but it "helped me ask the right questions."
The second-order effect is structural. A patient equipped with Claude and the discipline to cross-reference it against expert opinion can operate as a more informed consumer of rare-disease care than the system was designed to support. This scales poorly as long as it depends on individual initiative, but it signals where the medical system has systematic blind spots.
Medical organizations should audit false-positive rates on imaging and diagnostic protocols
Christou's case turned on a 60% false-positive rate on end-of-treatment PET scans for his lymphoma subtype—a statistic he called astonishing. "It's 2026," he said. "Sixty percent." Yet this rate, while documented in the literature, is often not front-of-mind during clinical decision-making, especially when treating rare presentations.
Healthcare leaders should conduct a systematic review of high-false-positive imaging workflows and flag them to clinical informatics teams. The alternative is watching patients and their advocates use general-purpose LLMs to catch what the institution's own protocols missed.
Christou is now running Keragon with new urgency. He watched nurses and doctors buried under administrative tasks unrelated to care. He received the same chemotherapy protocol as an 80-year-old patient, with side effects managed by a cascading chain of additional drugs. He is convinced, he says, that "we will look back at this era of treatment and cringe."