Vibe Coding My Mom’s 4-Year Medical Mystery

My mom has been getting horribly sick every spring starting in 2022. Same pattern each time — fever, joint pain, angry red rashes that spread across her skin, malaise.

During the worst flares, she couldn’t get out of bed for days at a time to feed herself. I’d get messages from her at 3 AM Bulgarian time — photos of new rashes, swollen hands and limbs, messages filled with panic, countless tears.

Over those years, we’ve seen 30+ doctors across 10+ hospitals and 3+ Bulgarian cities. Rheumatologists, dermatologists, internists, specialists in Sofia, local physicians in her Bulgarian city. We easily spent months of time and thousands of Euros chasing answers that never came.

Some would shrug and shuffle her along to the next specialist. Others would allude to scary possibilities and still shuffle her along to yet another appointment a few weeks or months down the road.

At first, we had remarkably little diagnostic clarity to show for all these consultations. Later on, as it turned out, even when we thought we had clarity, we were wrong.

Timeline #

2022: The first episode. Complete mystery. Dark urine, systemic inflammation, months of illness before it mysteriously resolved. No answers. We visited over fifteen doctors across three cities that year.

2023: This was the one uneventful year we had. 🙂 She had some modest symptoms but nothing that made her bedridden.

2024: Another spring flare lasting 4-6 weeks. This time skin inflammation and sky-high blood pressure dominated. We discovered additional medical constraints but still no clear diagnosis. The flare eventually subsided mostly on its own.

2025: Another severe episode. For weeks, she couldn’t move in the mornings due to stiffness and joint pain. Two specialists both declared it a textbook case of rheumatoid arthritis. The reasoning seemed straightforward — her rheumatoid factor was through the roof, her joints hurt, she was inflamed. Must be RA. At this point, it was clear that what we were dealing with was an autoimmune condition, and after a short course of corticosteroids, she felt better.

2026: Another brutal flare. This time: widespread skin rashes, the highest fever yet, and crushing fatigue. This time, I had a Claude Pro subscription.

2026 - April update: The Claude Pro subscription was wrong. She got hospitalized in an academic medical center.

Why RA never fit #

RA doesn’t cause full-body skin rashes. RA doesn’t follow predictable seasonal patterns. RA doesn’t cause episodic high blood pressure. And RA doesn’t cause dark urine.

The prescribed treatment was methotrexate — standard RA therapy. But she couldn’t take it due to other medical constraints. MTX is a powerful immunosuppressant with serious side effects that her medical history made too risky.

She hadn’t taken it even though she was prescribed it by two separate rheumatologists eight months prior — ironically one of them was a leading vasculitis specialist.

Enter Claude #

During her March flare — when she was running 38.3°C fever and couldn’t leave her apartment — I decided to try something.

As a Machine Learning Engineer, I use AI models daily for work. Instead of just feeding Claude code and training metrics, I started feeding it every detail about my mom’s condition — lab results, symptoms, photos of rashes, the complete medical timeline spanning 30+ doctors and mountains of diagnostic paperwork.

Claude didn’t just analyze isolated symptoms or test results. It held the entire four-year narrative in context simultaneously — something no single doctor had ever done. It could reference her 2022 dark urine episode while analyzing her 2024 lab work while considering her current symptoms, identifying patterns across time that dozens of human specialists had missed.

Claude’s immediate impact #

The first breakthrough was surprisingly practical. Claude could confidently recommend immediate symptom management — choosing between ibuprofen and acetaminophen, evaluating stronger NSAIDs, suggesting cold compresses, telling her whether she could self-administer modest corticosteroid doses, and predicting exactly how her fever would respond. And it worked. Her fever came down precisely as the model predicted, building my confidence in what followed.

Our diagnostic hypothesis #

When I sent photos of her rashes, Claude immediately recognized the reticular (net-like) pattern and distinguished it from what the dermatologist had misidentified. When I shared her Bulgarian lab reports, Claude read them directly — no translation needed — and caught the significance of results that had been overlooked.

Most impressively, Claude identified what seemed like a diagnostic trap that two specialist rheumatologists had fallen into: RF-positive vasculitis masquerading as RA. This is a well-documented phenomenon where 20-30% of people with systemic vasculitis have positive rheumatoid factor, leading to systematic misdiagnosis.

We became convinced her condition was cryoglobulinemia — a rare condition where abnormal proteins in the blood cause inflammation, often triggered by viral infections history. It fit her pattern perfectly: episodic flares, RF-positive but ANA-negative labs, the seasonal triggers, even the geographic correlation with her viral history.

The March crisis and hospitalization #

By late March 2026, her condition had deteriorated beyond anything we could manage at home. Fever spiking to 38.9°C, extensive rashes, treatment-resistant inflammation. After years of fighting for proper care, we finally got her admitted to Medical University Varna — Bulgaria’s premier academic medical center and the region’s only facility authorized to prescribe certain biologic treatments.

This was our shot at definitive answers from the top experts in the country.

What the experts found #

After more than a week of comprehensive testing by academic physicians, the diagnosis wasn’t what Claude expected.

The critical finding was her ferritin level: 4,776 ng/mL — massively elevated (normal for women is <300). Combined with her characteristic salmon-pink rash during this hospitalization, high fever, systemic illness, and steroid-responsive inflammation, this pointed definitively toward Adult-onset Still’s Disease (AOSD).

AOSD is an extremely rare systemic autoinflammatory condition affecting perhaps 1-2 people per 100,000 annually. In Bulgaria’s population of 6.9 million, there might be only 5-10 new cases diagnosed each year. Most general practitioners will never encounter a case in their entire career, and a rheumatologist might only see one or two. It is genuinely one of the hardest diagnoses to make in medicine because there are so many conditions you need to exclude. No wonder it took so long to identify.

Had we had the ferritin data point sooner, with AI we likely would have figured it out, I think. But it took us years to get the measurement.

Where we went wrong (and right) #

Our AI-assisted analysis was sophisticated but ultimately incorrect about the specific diagnosis. We had identified genuine patterns and real clinical features, but we had assembled them into the wrong diagnostic picture.

What we got right:

Where we went wrong:

The cryoglobulin test came back negative, while the ferritin elevation was so dramatic it essentially confirmed Still’s Disease when combined with her clinical presentation.

The computational advantage (and its limits) #

What Claude accomplished that 30+ human doctors couldn’t was maintaining perfect recall of every detail across years while simultaneously cross-referencing medical literature about rare conditions. It processed our scattered conversations, formal lab reports, photo evidence, and symptom timelines as a unified dataset.

Healthcare fragments care across specialists and time. Each doctor saw a slice — dermatologist focused on skin, rheumatologist on joints, internist on general symptoms. No one had ever assembled the complete picture that Claude could analyze as a whole.

But Claude also led us down the wrong diagnostic path. The ability to synthesize vast amounts of information and identify patterns doesn’t guarantee correct conclusions. Even with perfect recall and extensive medical knowledge, diagnostic reasoning requires clinical judgment that remains challenging for AI systems.

What this reveals about diagnosis #

We spent years seeking help from human experts who were individually competent but collectively missed the pattern. Not because they were bad doctors, but because our healthcare system isn’t designed for complex cases spanning years, multiple specialties, and requiring synthesis of scattered information.

But we also learned that even sophisticated AI analysis can reach confident but incorrect conclusions. The academic physicians at Medical University Varna succeeded where both community doctors and AI analysis had failed — through systematic clinical evaluation, comprehensive laboratory testing, and direct patient examination.

The emotional toll of medical uncertainty doesn’t appear in any diagnostic manual — the sleepless nights, constant worry, the way chronic illness radiates outward, consuming the energy and peace of mind of everyone who loves the patient.

Current status #

My mom is now home with a definitive diagnosis of Adult-onset Still’s Disease and a clear treatment plan: methotrexate plus colchicine for long-term disease control, with careful monitoring by academic specialists. Her ferritin has started dropping. The treatment is working but we’re not out of the woods yet.

The fever that had dominated every spring flare has been absent since hospital discharge. Now, she’s dealing with significant joint pain and morning stiffness — classic features of Still’s — these should improve as her body continues to recover over the coming weeks.

Still’s Disease has an excellent prognosis when properly treated. Most patients achieve good disease control and can transition to milder maintenance therapies once stable. I am feeling hopeful that she will make a full recovery.

The role of AI in complex diagnosis #

AI systems like Claude can serve as extraordinary diagnostic partners for pattern recognition, information synthesis, and medical literature review. They can help patients and families advocate more effectively and ask better questions.

But our experience also reveals the limitations of AI diagnosis. Clinical pattern recognition by experienced physicians, comprehensive laboratory testing, and direct patient evaluation remain irreplaceable components of complex diagnosis.

The greatest value may be in AI’s ability to help navigate healthcare systems, maintain comprehensive medical histories, and support informed medical decision-making.

When I started writing this blog post, I thought it would replace medical doctors, but the plot twist in the university hospital convinced me that clinical expertise is necessary with a complex condition.

The model was confidently wrong. Maybe less wrong than a general practitioner, but wrong nonetheless.

Lessons learned #

After four years of uncertainty, 30+ medical consultations, and both AI analysis and academic evaluation, we finally have a diagnosis and treatment plan. Sometimes solving medical mysteries requires the right combination of patient advocacy, comprehensive data collection, AI-assisted analysis, and access to specialized clinical expertise.

The greatest products solve people’s biggest problems. AI models have tremendous potential as diagnostic support tools, but they work best when integrated with — not replacing — expert medical care.

My mom’s recurring mystery illness has been one of my biggest sources of worry in recent years. Through a combination of AI assistance and expert medical care, I now have more confidence than ever that she’ll be okay.

This blog post was written with assistance from Claude.

 
7
Kudos
 
7
Kudos

Now read this

On Iterating at Facebook

During my time at Facebook, I learned a whole lot about A/B testing. This post is a brain dump in which I go over the importance of A/B testing, the gist of the components and infrastructure necessary to support an A/B testing framework,... Continue →