He had a six-pack, a founder's drive, and a terminal diagnosis. Connor Christou didn't just fight cancer - he fed it to an AI. When confronted with the disease, Connor, a previously invincible founder, turned his entire health regime into a dataset and handed it to Claude. This isn't a story about miracle cures. It's about how a new generation of patients are using large language models to become active, data-driven participants in their own care - and what that means for the rest of us.
Connor's regimen was extreme even by founder standards: daily Sauna, cold plunges, a strict paleo diet, and a continuous glucose monitor. His Oura ring tracked sleep and heart rate variability. He logged every meal, every supplement, every blood draw. When cancer hit, he didn't stop. He doubled down, and he began feeding his weekly blood tests - radiology reports, wearable data exports, and free-form journal entries into Claude, asking it to find correlations, flag anomalies, and suggest questions for his oncologist.
This approach represent a radical shift. Most patients hand over control to their doctors, and connor handed it to an APIBut his experiment raises critical questions about reliability, privacy. And the limits of AI in medicine - questions that anyone building or using AI tools for health should consider.
The Quantified Founder: A Case Study in Personal Health Data Overload
Connor Christou isn't a doctor. He's a founder who previously built a fintech platform. But his background gave him a unique tool: the ability to treat his body as a software system. "I had dashboards for my company's KPIs," he told me in a direct message. "Why wouldn't I have one for my body? "
His data stream included six types of inputs: continuous glucose monitor readings (every 5 minutes), Oura ring biometrics (sleep, HRV, temperature), daily step count (Apple Watch), blood biomarker panels (CBC - metabolic panel, inflammatory markers), scan reports (CT, MRI, PET). And subjective mood/energy logs. That's thousands of data points per day.
No single human can synthesize that volume. Traditional oncology relies on discrete snapshots - a scan result, a blood test - but misses the longitudinal dynamics. Connor realized that large language models, with their ability to handle unstructured text and large contexts, could be the missing integrator.
Why Claude? The Technical Rationale for Choosing an LLM Over Specialized Software
Most of the existing digital health tools are siloed. Your CGM data lives in one app, your Oura ring in another, your lab results in a patient portal there's no universal import. Connor needed a platform that could accept arbitrary data formats and extract meaning without a predefined schema. That's where Claude (and similar models like GPT-4) shine.
He used Claude's long-context window - up to 200,000 tokens - to dump entire weeks of raw data into a single prompt. He structured it as a chain-of-thought conversation: "Here are my last 7 days of CGM values. Identify any patterns that correlate with my reported fatigue levels. " Claude could then cross-reference timestamps with his journal entries and flag, for example, that fatigue spiked 4 hours after high-carb dinners.
This isn't something a traditional health monitoring app can do. They rely on pre-built rules, and claude can reason dynamicallyConnor also used the API to build a custom Python pipeline that scraped PDF lab reports using PyMuPDF, extracted tables with Camelot, and passed them to Claude as structured JSON. The result was a near-real-time health status summary.
Data Integration Without an EHR: Building the Patient-as-System Pipeline
One of the biggest challenges Connor faced was getting his data into a machine-readable format. Hospital portals export PDFs - not JSON. His CGM software had a CSV export. But it was formatted for spreadsheets, not APIs. He wrote a series of ETL scripts using Python: pandas for cleaning, regex for scraping unstructured reports, requests to push the cleaned data to Claude's Messages API
His pipeline looked like this:
1. Export raw data from each source (manual download or API).
2Normalize timestamps to UTC and merge into a single CSV.
3. Generate a natural language summary for each day using a lightweight model.
4. Feed the monthly summary to Claude with a custom system prompt specifying his diagnosis, medications, and known side effects.
5. Parse Claude's response into actionable items: "Your CRP is trending up - discuss infection with your oncologist. "
This pipeline is reproducible. Any developer with basic Python skills could adapt it for their own health data. The key insight is that you don't need a medical device integration platform - you need a flexible AI that can understand messy, multimodal input.
The Claude Experiment: What Did the AI Actually Find?
Connor shared a few specific examples. After three weeks of data, Claude identified that his hemoglobin levels dropped by 12% two days after each chemotherapy session. But recovered faster when he consumed beets and iron-rich foods. His oncologist hadn't noticed the correlation because the data was scattered across different reports.
Claude also detected a pattern in his heart rate variability: on days when his sleep duration fell below 6 hours, his HRV dropped by 20 points, and his subjective pain scores increased. The AI suggested timing his pain medication before bed on nights following poor sleep. Connor tried it. And it worked
Perhaps most impressively, Claude helped him formulate better questions for his care team. Instead of asking "Why am I so tired? ", he could ask "Given my latest TSH of 3. 5 and my current dose of levothyroxine, should we consider a dosage adjustment to improve energy levels? " The specificity led to faster interventions.
Hallucination Risks and Mitigation: The Devil in the LLM Details
Let's not pretend this is safe out of the box. Large language models hallucinate. Connor experienced it: Claude once suggested he had a vitamin B12 deficiency based on a misread of his lab values (it confused B12 with folate). Another time it recommended a supplement that could interact with his chemotherapy drug.
To mitigate this, Connor implemented a three-layer safety net:
- Grounding: He included explicit reference ranges in every prompt (e g, and, "Normal WBC: 45-11. 0 x10^9/L"), but
- Confidence scoring: He asked Claude to rate its certainty for each claim (1-10) and ignored anything below 7.
- Human-in-the-loop: All AI-generated suggestions were reviewed by his oncologist before action.
He also used retrieval-augmented generation (RAG) to ground Claude's answers in his actual lab reports, rather than relying solely on the model's training data. By embedding his PDFs into a vector store and fetching relevant chunks before each query, he reduced hallucination rates significantly.
Beyond Cancer: A Framework for Chronic Condition Management
Connor's approach isn't limited to oncology. Anyone managing a chronic condition with multiple data streams - diabetes, hypertension, autoimmune disorders - could benefit. The core idea is to turn the patient-doctor relationship from a series of snapshots into a continuous, data-rich dialogue.
Consider type 1 diabetes: continuous glucose monitors, insulin pump logs, food diaries, activity trackers. No clinician has time to review all that. But an LLM can generate a weekly summary highlighting dangerous patterns, suggesting insulin dose adjustments. And flagging anomalies. A 2024 study in Nature Medicine found that LLM-assisted diabetes management reduced hypoglycemic events by 18%.
The barrier is no longer technical - it's behavioral and regulatory, and patients need to consistently collect dataDoctors need to trust AI outputs. And regulators need to certify these tools as medical devices or exempt them as "wellness" applications. Right now, Connor's system exists in a gray zone. He's not selling it, and he's using it for himselfBut the demand is growing,
Ethical Boundaries: Privacy, Consent. And the Risk of Over-Diagnosis
Connor's pipeline stores his health data in a cloud bucket (encrypted at rest) and sends it to Anthropic's API. He chose Claude partly because of its enterprise-grade privacy policy: Anthropic doesn't train on API data. But many users won't be that careful. Health data is among the most sensitive information a person can share. A breach could lead to discrimination - insurance denial, or worse.
There's also the risk of "over-diagnosis by AI. " Connor noticed that Claude sometimes flagged minor deviations that were clinically insignificant - a 1% drop in albumin, a single elevated liver enzyme. If a patient acts on every suggestion, they can generate unnecessary anxiety and unnecessary medical visits. The AI doesn't have context for statistical noise.
Finally, there's the question of equityBuilding a custom Python pipeline and paying for Claude API credits isn't accessible to everyone. If AI-powered health management remains a privilege of the wealthy, we risk widening health outcome disparities. Open-source alternatives like Ollama with local models could help democratize it. But local hardware requirements are still a barrier.
What It Means for Developers and Engineers
Connor's story is a blueprint. It shows that the APIs we use to build chatbots and code assistants can be repurposed for high-stakes personal analytics. For developers, the takeaway is pragmatic: start thinking about data pipelines, not just prompt engineering. The ability to extract, normalize, and structure real-world data is the bottleneck, not the language model itself.
If you're building a health-related AI tool, consider these design principles:
- Assume messy data. Users won't have clean CSV exports. Build parsers for PDFs, images, and free text.
- Prioritize explainability Show users the source data behind each claim. Use citations,
- Implement kill switches. Let users easily disable any AI-generated recommendation,
- Respect privacy by default Encrypt data in transit and at rest. Allow local processing when possible,
The next wave of health innovation won't come from hospitals. It will come from founders who treat their own bodies as software systems - and from the engineers who build the tools to help them do it safely.
Frequently Asked Questions
- Can I replicate Connor's method without coding.
Not exactlyYou need basic Python skills to set up the data pipeline. However, you can use ChatGPT or Claude directly by uploading PDFs and copying data manually - it's less automated but still useful. - Is it safe to rely on an LLM for medical advice?
No. LLMs should never be the sole source of medical decisions, and always verify with a licensed professionalUse the AI as a thought partner, not a doctor. - Which AI model is best for personal health analytics?
Claude (Sonnet/Opus) and GPT-4 offer large context windows and reasonable factual accuracy. For local processing, try Llama 3 or Mistral with a RAG system. - How do I ensure my health data stays private?
Use APIs that promise not to train on your data (e g., Anthropic, OpenAI enterprise), encrypt your files. And consider running a local model via Ollama if you have a powerful GPU. - What if the AI contradicts my doctor?
Trust your doctor. Use the AI's insights to ask better questions, not to override professional judgment, and hallucinations are real
What do you think?
Do you believe that AI-driven personal analytics will become a standard part of cancer care within the next five years,? Or will regulatory and safety concerns keep it in the hands of a few adventurous patients?
If you had access to your family's complete health records, would you feed them into an LLM? What would you want it to find?
Should platforms like Claude explicitly block health-related queries unless they're reviewed by a medical professional,? Or should personal experimentation be allowed?
.
Need a Custom App Built?
Let's discuss your project and bring your ideas to life.
Contact Me Today →