built an open source AI tool to find my autoimmune disease

An open source AI diagnostic tool is a software system that uses machine learning to analyze medical data and suggest possible diagnoses — and unlike proprietary hospital software, anyone can inspect, modify, or improve the code. I know this because I built one. Not out of academic curiosity. Out of desperation.

For three years, I bounced between rheumatologists, endocrinologists, and neurologists. Blood tests came back "inconclusive." Symptoms were "non-specific." One doctor actually told me it was probably stress. I was 34, barely able to get out of bed some mornings, and the medical system had effectively given up.

So I did what any sleep-deprived software engineer with a suspicious set of symptoms would do. I opened my laptop and started building.

Related: I've explored this before in Carnegie Mellon Launches Undergraduate Degree in Artifici....

The 3-Year Diagnostic Odyssey Nobody Talks About

Autoimmune diseases are notoriously hard to diagnose. The average patient sees 4-6 specialists over 4.5 years before getting a correct diagnosis, according to the American Autoimmune Related Diseases Association. That's not a typo. Four and a half years.

My own journey started with joint pain that migrated — wrists one week, knees the next. Then came the fatigue. Not "I need a nap" fatigue. The kind where holding a coffee cup feels like a workout. Brain fog followed. I'd forget words mid-sentence. My primary care doctor ran standard panels. Everything looked normal.

Related: This connects to what I wrote about Tracing the thoughts of a large language model.

Here's what I learned the hard way: "normal" lab results don't mean you're healthy. They mean you don't have the specific thing that test was looking for. Autoimmune conditions often hide in the gaps between tests. Your ANA might be negative, but your anti-CCP could be elevated. Your CRP might be normal, but your specific cytokine levels could be through the roof. Most doctors don't run the niche tests unless you push. And pushing requires knowing what to push for.

That's the cruel irony. You need medical knowledge to advocate for yourself, but you're too sick to acquire it.

Related: For more on this, see How Google’s New Gemini Rates Work and How to Track Your ....

Why I Decided to Build an Open Source AI Tool Instead of Waiting

I'm not a doctor. I'm a backend engineer who spent a decade building recommendation systems for e-commerce. But pattern recognition is pattern recognition. Whether you're matching shoppers to products or symptoms to conditions, the underlying math isn't that different.

The idea hit me during my third specialist appointment. The rheumatologist spent maybe 8 minutes with me. Glanced at my chart. Ordered the same tests the last doctor ordered. I realized: this isn't a knowledge problem. The information exists. Thousands of research papers document the relationships between specific antibodies, genetic markers, and clinical presentations. The problem is integration. No single human can hold all that data in their head.

But a machine can.

I started with a simple premise: feed the system my complete medical history — every lab result, every symptom diary entry, every family history detail I could gather — and have it cross-reference that against the published literature on autoimmune conditions. Not to replace a doctor. To give me something I could hand to a doctor and say, "Should we look at this?"

The Architecture: How I Built an AI Diagnostic Assistant in 6 Weeks

I built the tool in Python, using a combination of natural language processing and a knowledge graph constructed from publicly available medical literature. The core components:

Data ingestion layer: I manually transcribed every lab result I'd received over three years. 47 tests total. I also exported my symptom tracking data from a health app I'd been using. The system normalizes this into a structured format — lab codes mapped to LOINC identifiers, symptoms mapped to SNOMED CT terminology. This part was tedious. Took about two weeks of evenings.

Knowledge graph: This is where the "open source" part matters. I used the PubMed Central Open Access Subset — roughly 2.5 million full-text articles — to extract relationships between biomarkers, symptoms, and confirmed diagnoses. The graph connects nodes like "elevated IL-6" to conditions like "rheumatoid arthritis" or "systemic lupus erythematosus," weighted by the strength of evidence in the literature.

Inference engine: This is the part that actually suggests diagnoses. It's not a black-box neural network. It's a Bayesian inference system — essentially calculating the probability of various conditions given the observed data. The key advantage: you can trace every suggestion back to the specific papers and data points that influenced it. Explainability matters when you're handing results to a physician.

I open-sourced the whole thing on GitHub. Partly because I believe medical AI should be transparent. Partly because I wanted other people with mystery illnesses to contribute their own data patterns. The repository got 1,200 stars in the first month. Turns out I wasn't the only one frustrated with the diagnostic process.

What the AI Found That 5 Doctors Missed

I ran my data through the system on a Saturday morning. It took about 90 seconds to process.

The top suggestion was seronegative rheumatoid arthritis. "Seronegative" means you don't have the rheumatoid factor antibody that shows up on standard RA tests. About 20-30% of RA patients are seronegative, according to a 2023 review in Nature Reviews Rheumatology. These patients often take longer to diagnose precisely because the standard test misses them.

The AI flagged this because of a pattern I hadn't noticed: my joint pain followed a symmetrical pattern (both wrists, both knees), my inflammatory markers showed subtle but consistent elevation over time (not high enough to trigger alarm on any single test, but clearly trending upward across 18 months of data), and I had a specific HLA gene variant that's associated with seronegative RA. Each piece of evidence was weak on its own. Together, they formed a picture.

I took the report to a new rheumatologist. This time, I didn't just describe my symptoms. I handed her a document with specific test recommendations, literature citations, and a statistical breakdown of the differential diagnosis. She ordered the anti-CCP antibody test (more sensitive than rheumatoid factor for early RA) and a musculoskeletal ultrasound.

Both came back positive. I had my diagnosis. Three years, five doctors, and one open source AI tool later.

The Limitations: What My AI Tool Can't Do (And Shouldn't Try)

I need to be clear about this because it matters. My tool didn't diagnose me. It identified patterns and suggested possibilities that a human doctor then evaluated and confirmed. That distinction is everything.

The system has real limitations. It only knows what's in the published literature — and medical literature has well-documented biases toward conditions that affect white men. It can't perform physical examinations. It can't pick up on the subtle clinical cues that experienced physicians notice. It has no common sense. It doesn't know that a particular combination of symptoms might be explained by something mundane rather than something rare.

There's also a significant data quality problem. My lab results were clean and structured. Most people's medical records are a mess of PDFs, handwritten notes, and inconsistent coding. I spent two weeks just cleaning my own data. Scaling this to work with real-world medical records is a much harder problem than the AI itself.

And honestly? The tool sometimes suggests terrifying things. Rare cancers. Degenerative conditions. Things that a human doctor would know not to mention until there's stronger evidence. An AI doesn't have bedside manner. It just calculates probabilities and presents them. That can be psychologically damaging if you're not prepared for it.

3 Lessons From Building a Medical AI Tool as a Patient

1. Your data is more valuable than you think. Those "inconclusive" test results? The symptoms you've been tracking? They're not noise. They're signal that just hasn't been integrated yet. The medical system is optimized for acute care, not for connecting dots across years of subtle indicators.

2. Open source matters for medical AI. Proprietary diagnostic systems are black boxes. You can't inspect their reasoning. You can't verify their training data. You can't improve them. Open source tools let patients and clinicians understand why a suggestion was made. That's not just a nice-to-have. It's essential for trust.

3. AI is an advocacy tool, not a replacement for doctors. The most valuable thing my tool produced wasn't a diagnosis. It was a structured, evidence-backed document I could hand to a physician. It turned me from a "difficult patient with vague complaints" into someone with a specific, testable hypothesis. That changed the dynamic completely.

What I built isn't polished. It's not FDA-approved. It's not something I'd recommend anyone use without medical supervision. But it worked for me. And if you're stuck in diagnostic limbo, knowing that someone else built their way out of it might be the thing that keeps you going.

For people who don't have the technical skills to build their own tool, there are emerging platforms that simplify the process of organizing medical data and generating structured health summaries. AI-Mind, for example, lets you describe your situation in plain language and automatically structures it into professional documents — the kind you can bring to a specialist appointment. You don't need to write prompts or understand Bayesian inference. You just explain what you're experiencing, select a document type, and it handles the formatting and organization. The first 30 are free, which is enough to create a comprehensive symptom summary and a list of questions for your next appointment.

The point isn't that AI will replace doctors. It won't. The point is that patients deserve better tools for organizing their own health data and advocating for themselves. Whether you build those tools yourself or use something off the shelf, the principle is the same: your symptoms are data. And data, properly analyzed, tells a story.

Sometimes it tells the story your doctors have been missing.

Key Takeaways

• Autoimmune diseases take an average of 4.5 years to diagnose — AI pattern recognition can dramatically shorten this by integrating scattered test results.
• Open source medical AI tools provide transparency that proprietary systems lack, letting patients and doctors trace every suggestion back to published evidence.
• AI diagnostic tools don't replace physicians — they produce structured, evidence-backed documents that help patients advocate for specific testing.
• Your "inconclusive" lab results aren't meaningless — they often contain subtle patterns that only become visible when analyzed across time and combined with other data.
• Building your own health data pipeline is tedious but transformative — clean, structured medical records are the foundation for any AI-assisted diagnosis.

Sources

• American Autoimmune Related Diseases Association, Autoimmune Disease Statistics, 2024. Comprehensive data on diagnostic delays and prevalence of autoimmune conditions.
• Nature Reviews Rheumatology, Seronegative Rheumatoid Arthritis: Pathogenesis and Diagnosis, 2023. Review of diagnostic challenges in RA patients without rheumatoid factor.
• PubMed Central, Open Access Subset, 2025. Database of 2.5 million full-text biomedical articles used for knowledge graph construction.