I Ran My Entire Medical Profile Through an AI Research System. Here’s What It Found That My Doctors Missed.

I’ve always been the kind of patient who asks too many questions. After surviving stage 4 colorectal cancer — beating 11% odds, enduring chemotherapy, surgery, and becoming the first human to receive a specific type of engineered T-cell therapy — I’d learned that nobody cares more about my survival than me.

But there was still data sitting in my medical records that nobody had looked at.

My oncology team ordered comprehensive genomic testing early in my treatment. It’s standard for advanced cancer — they sequence your tumor and your normal DNA to find mutations that might guide therapy. The clinical report came back with about 50 curated variants. My doctors used it to match me to a clinical trial. That’s what it’s designed for, and it worked.

What I didn’t know until recently: the full dataset contained over 3,000 variants. The clinical report was a highlight reel. The rest — the germline variants, the ones baked into my DNA since birth — had been sitting untouched in raw data files for over two years.

So I built a system to analyze all of it.

What I Built (and Why It’s Not What You Think)

I’m not a bioinformatician. I’m a patient who taught himself to read research papers because his life depended on it. But what I built over the past several weeks is closer to a research department than a Google search.

Let me be specific, because “I used AI” conjures images of typing a question into a chatbot and copying the answer. That’s not what this was. Not even close.

Step 1: Building the Data Foundation

I wrote a systematic literature retrieval pipeline — 106 targeted search queries against PubMed’s database of biomedical research, each one designed around a specific aspect of my case: my mutations, my treatment history, my drug interactions, my supplement candidates, my immune profile. This wasn’t “search for colorectal cancer.” This was 106 separate, carefully constructed queries like a research librarian would build for a systematic review.

Simultaneously, I pulled 599 clinical trials from ClinicalTrials.gov across 20 query categories. I downloaded pharmaceutical interaction databases. I extracted my complete genomic data — not the clinical summary, but the raw variant call files, copy number segments, structural variants, and fusion data. Over 3,000 variants versus the 53 in my clinical report.

Step 2: Local Compute Extraction

All of this data went into a dedicated GPU running on my own hardware. No cloud services, no subscriptions, no data leaving my machine. The local system processed the literature in multiple phases:

Intervention extraction: 12,539 potential interventions identified from the literature — pharmaceuticals, supplements, vitamins, metabolic targets, immune modulators, repurposed drugs, dietary compounds. Every one cross-referenced against my specific mutation profile.
Interaction matrix: 350 drug-supplement-drug interactions mapped. When you’re taking 15+ compounds, knowing what conflicts with what isn’t optional — it’s survival.
Self-audit: The system audited its own top 500 candidates for logical consistency, checked dosing against published safety data, and flagged anything that contradicted my clinical picture.

This phase alone processed over 12,500 research papers. It ran for days.

Step 3: Blind Verification Protocol

Here’s where the rigor matters. I took the output from my local system and sent it — along with my complete medical profile — to multiple independent AI analysis models. The critical part: each model worked in isolation. None of them could see what the others produced. This is the same principle behind double-blind peer review in scientific publishing.

I didn’t use one model and trust its answer. I used several, compared their independent conclusions, and built consensus only where multiple models converged without coordination. Where they disagreed, I investigated why. Sometimes the disagreement revealed something more important than agreement would have.

Step 4: Consensus Building

The final output wasn’t a single model’s opinion. It was a scored consensus protocol — interventions ranked by how many independent analyses supported them, with confidence levels, dosing ranges, interaction warnings, and monitoring requirements. Think of it as the equivalent of getting five specialist opinions and mapping where they agree.

Only findings that multiple models independently converged on made the final cut. Anything supported by a single analysis was flagged for further investigation, not treated as actionable.

Step 5: The Human Filter

This is the step most people skip when they talk about AI in healthcare, and it’s the most important one. Every finding went through me — a patient who knows his own body, his treatment history, and his clinical context better than any model ever could. And as you’ll see below, that human filter caught something every single AI model missed.

Then everything went to my doctors. Every finding, every recommendation, every question. This system doesn’t replace medical judgment. It arms you with better questions to bring to the people who have it.

What I Found

I’m going to share three findings because they illustrate different ways this kind of analysis changes your care.

Finding 1: A Genetic Variant That Explains Why I React Differently to Common Medications

Deep in my germline data, multiple models independently flagged a variant in a gene called CYP2D6. This gene controls how your liver metabolizes a wide range of drugs — including common pain medications.

My variant makes me an intermediate metabolizer. In plain English: certain medications don’t work as well for me because my body processes them differently than expected.

I’d already experienced this firsthand. After a prior surgery, the standard dose of a strong pain medication barely touched my pain while making me groggy. I’d chalked it up to a bad day. Turns out it was genetics.

Why this matters now: I have a major surgery scheduled in a few months. Without this information, my anesthesiologist would plan a standard pain management protocol. With it, they can adjust. That’s the difference between waking up from surgery with adequate pain control or spending days in unnecessary suffering because the math was wrong for my body.

This finding was in my data for over two years. Nobody told me.

Finding 2: A Methylation Defect That Changes Everything About Supplements

Another consensus finding: I’m homozygous for a variant in the MTHFR gene — specifically C677T. This means both copies of the gene are affected, reducing enzyme activity to about 30-40% of normal.

MTHFR is involved in methylation, which is fundamental to DNA repair, detoxification, and how your body uses B vitamins. With this variant, my body can’t properly convert regular folic acid into its active form. Every multivitamin I’d ever taken with folic acid in it was mostly wasted on me.

More importantly, for a cancer survivor: the methylation pathway intersects directly with DNA repair. Understanding this variant allowed me to rebuild my entire supplement approach around what my body can actually use — methylated forms of B vitamins instead of standard ones, plus supporting nutrients the models identified to compensate.

My oncologist never mentioned MTHFR. My PCP never tested for it. It showed up in genomic data that was ordered for an entirely different purpose.

Finding 3: A Scary Result That Turned Out to Be Noise — and Why That Matters Most

Here’s where it gets interesting, and honestly, a little scary.

The AI models flagged a homozygous frameshift mutation in a gene called PRKDC. This gene is essential for DNA repair. A true loss-of-function mutation in both copies would be devastating — we’re talking severe combined immunodeficiency. People with this condition struggle to survive childhood, let alone tolerate aggressive cancer treatment.

I was up most of the night after seeing this result. But something didn’t add up.

I’m 50 years old. I was healthy until my cancer diagnosis at 47. I tolerated six months of aggressive chemotherapy without my immune system collapsing — which surprised even my oncologist. I then survived four doses of high-dose IL-2, a treatment so brutal most patients end up in the ICU after just one. My engineered immune cells expanded and functioned exactly as designed, which requires the very DNA repair pathway this mutation supposedly destroyed.

My instinct said: if this were real, I shouldn’t be alive.

So I dug in. The variant appeared in a region of the genome prone to sequencing artifacts — a stretch of repeating DNA letters where the sequencing machine is known to misread. Two different analysis programs had flagged it, but both were reading the same underlying data. It looked like independent confirmation, but it wasn’t — it was the same error counted twice.

The genomic testing company’s own clinical team had apparently caught this too. It wasn’t in their curated report. They’d dismissed it during their quality review, but nobody told me that, because patients don’t typically see the raw data.

This is the most important lesson I can share: AI models treated this data at face value. They said “needs confirmation” but none of them said “this is almost certainly wrong because you are alive and functioning.” That clinical reasoning — the ability to look at data AND the patient standing in front of you — is something only a human can do. I caught it because I know my own body and my own medical history better than any algorithm.

AI is a force multiplier. It is not a replacement for thinking.

What This Means for Other Patients

I want to be honest about the limitations. I had advantages most patients don’t: time, technical aptitude, and the kind of obsessive determination that comes from staring down a terminal diagnosis. Not everyone can or should try to replicate this.

But here’s what I think any cancer patient can take away:

Your genomic data probably contains more than you’ve been shown. Clinical reports are summaries. They’re designed to answer the specific question your oncologist asked — usually “what treatment should we try?” The broader implications for your health, your other medications, your supplements, your future surgeries — those often fall through the cracks. Not because your doctors are bad, but because the system isn’t built to connect those dots.

You have the right to request your full data. Under federal law, your genomic testing company must provide your complete results if you ask. Most people never ask.

Second opinions aren’t just for diagnoses. They’re for data interpretation too. If your treatment plan was based on a 50-variant summary when 3,000 variants exist, there might be actionable information sitting untouched.

Bring everything to your doctors. I cannot stress this enough. I bring every finding, every question, every half-formed theory to my medical team. They’ve been receptive because I show my work and respect their expertise. Self-advocacy doesn’t mean going rogue. It means being the most informed participant in your own care.

The Bigger Picture

I’m now 8 months out from treatment with no evidence of disease. My engineered T cells are still circulating at 15 times normal levels.

Running my full medical profile through an AI research system didn’t save my life. The surgeons, the chemotherapy, the brilliant immunotherapy team at NIH — they saved my life. But this process gave me information that makes my ongoing care smarter, my supplement stack more targeted, and my upcoming surgery safer.

Two years of data was sitting in a file. Nobody was going to look at it. So I did.

If you’re a cancer patient or survivor with comprehensive genomic testing in your records, consider asking what’s in the full dataset versus what made it into the report. The answer might change how you manage your health going forward.

I bring everything to my doctors. You should too.

Aaron M. is a stage 4 colorectal cancer survivor and the first patient to receive TCR-T cell therapy targeting p53 Y220C for a solid tumor. He writes about cancer survivorship, self-advocacy, and the evolving role of technology in patient care at beat-crc.com.