I used Claude Code to get a second opinion on my MRI
**Byline:** Marcus Webb — Infrastructure engineer turned tech writer. Writes about AI, DevOps, and security.
> **Bottom line:** Faced with a complex MRI report, I fed a de-identified summary of my findings to Claude Code 4.6, expecting a technical breakdown.
The AI didn't just rephrase medical jargon; it identified a subtle interaction between two findings my doctor had presented as isolated, suggesting a specific follow-up question regarding a vascular anomaly.
While not a diagnostic tool, this experiment on June 29, 2026, demonstrated Claude’s advanced pattern recognition capabilities beyond its intended code interpretation, offering a novel way to augment personal understanding of medical data.
I cancelled my ChatGPT Pro subscription after 6 months. Not because it was bad — but because I discovered what it was doing to my ability to think.
So when I found myself staring at a dense, jargon-filled MRI report two weeks ago, with my doctor's explanation still echoing vaguely in my ears, my first impulse was *not* to fire up an LLM.
It was to call a friend who’s a radiologist. But that friend was on vacation, and my anxiety wasn't.
That's when an idea, equal parts desperate and genuinely intriguing, sparked: What if I treated my MRI report like a complex system log?
A verbose output from an opaque black box, full of cryptic codes and interdependencies. I needed a diagnostic tool for *that*.
And the tool I’d been exploring for parsing complex codebases and obscure API documentation? Claude Code 4.6.
The Unconventional Debug Session
My doctor had handed me the report for a shoulder MRI, explaining a few key findings.
"Some mild tendinopathy here, a bit of fluid there, nothing major to worry about right now." He’s a good doctor, but the language of radiology reports is a beast.
"Mild supraspinatus tendinopathy with associated subacromial-subdeltoid bursitis," it read.
Okay, I got that.
But then came phrases like "minimal hypertrophic changes of the acromioclavicular joint" and "trace fluid within the biceps tendon sheath." Each finding felt like a separate, isolated log entry.
My brain, wired for systems thinking, immediately looked for the connections. Is the fluid related to the tendinopathy? Is the joint change contributing to the bursatis?
My doctor, in his effort to reassure me, had presented them as distinct, manageable issues. But what if they weren't?
What if there was a subtle, underlying thread I was missing, a systemic issue rather than isolated component failures?
Framing the Medical Data as a System State
I knew I couldn't just upload the raw MRI images. That’s a whole different ballgame for specialized medical AI, not a general-purpose LLM, especially one tuned for code.
What I *could* do was take the textual report.
But even that felt too raw, too sensitive. My primary concern was privacy. I wasn't going to feed my personal health information directly into a public AI.
So, I spent about an hour painstakingly de-identifying and summarizing the report.
I extracted all the key anatomical descriptions, pathologies, and measurements, stripping out my name, date of birth, and any other unique identifiers.
I focused on the *structure* of the findings: "Impression: 1.
Finding A. 2. Finding B.
3. Finding C." Each finding had a detailed description in the "Findings" section. I effectively created a structured data representation of my shoulder's current state, as described by the radiologist.
My prompt to Claude Code 4.6 was carefully constructed, treating the report as a complex data schema:
```
"I have a structured medical report describing the state of a biological system (a human shoulder). The report contains several distinct findings and an impression section summarizing them.
My goal is to understand potential interdependencies or causal relationships between these findings that might not be immediately obvious. Please analyze the following findings and impressions:
[Paste the de-identified, summarized MRI findings and impression here, using bullet points for clarity and maintaining medical terminology.]
Specifically, I need you to:
1. Identify any findings that, while described separately, might be causally linked or exacerbate each other.
2. Suggest any potential implications or further questions based on the combination of these findings, even if not explicitly stated in the report.
3. Explain any complex medical terms in layman's terms, but maintain the precision of the original report.
Treat this as debugging a complex system where components might be interacting unexpectedly." ```
I know, it sounds like I was asking an AI to play doctor. And in a way, I was. But I wasn't asking for a diagnosis. I was asking for a systems analysis.
Claude Code's Uncanny Insight
Claude Code 4.6 processed the input in about 15 seconds. Its response was structured, concise, and frankly, a bit unsettling in its perceptiveness.
It started by breaking down each finding, just like my doctor had, but then it did something remarkable.
It highlighted the "minimal hypertrophic changes of the acromioclavicular joint" and the "trace fluid within the biceps tendon sheath." My doctor had presented these as distinct, minor observations.
Claude, however, connected them.
**Claude's exact output included this passage:**
> "Regarding the 'minimal hypertrophic changes of the acromioclavicular joint' and 'trace fluid within the biceps tendon sheath,' while often considered minor, the presence of both *could* suggest a subtle mechanical impingement or altered biomechanics contributing to the biceps tendon irritation.
Hypertrophic changes (overgrowth) in the AC joint can narrow the subacromial space, potentially affecting structures passing through it, including the biceps tendon.
The trace fluid could be an early indicator of this increased mechanical stress."
It then went on to suggest:
> "This combination warrants a follow-up question to your physician regarding the possibility of subacromial impingement syndrome or a need to assess dynamic shoulder mechanics, particularly if symptoms persist or worsen.
It's a low-grade interaction, but one that a systems engineer might flag as a potential cascading failure point."
A "cascading failure point." That hit home. My doctor hadn't used those words, nor had he explicitly linked these two specific findings beyond acknowledging their individual presence.
He'd focused on the more prominent tendinopathy. Claude had essentially found a "weak signal" in the noise, a potential root cause for ongoing irritation, by interpreting the "system state" as a whole.
This wasn't a diagnosis. It was a *question*. A highly specific, data-driven question I could now take back to my doctor.
It wasn't about contradicting him, but about enriching the discussion with a systems-level perspective.
The Reality Check: AI is Not a Doctor (Yet)
Let’s be absolutely clear: I am an infrastructure engineer, not a medical professional. And Claude Code 4.6 is an LLM, not a diagnostician.
This entire exercise was a personal experiment in augmenting my understanding, not replacing medical expertise.
The output from Claude was a hypothesis, a potential correlation, not a definitive medical truth.
The inherent risks are immense. If Claude had hallucinated a severe condition, or missed a critical finding, the emotional and practical consequences could be devastating.
It's crucial to acknowledge that general-purpose LLMs are not trained on vast, curated medical datasets with the same rigor as specialized medical AI models, let alone the decades of experience a human doctor possesses.
Moreover, feeding sensitive medical data into *any* public AI carries significant privacy risks. My de-identification process was a necessary, but imperfect, mitigation.
The "code" aspect of Claude Code 4.6 likely contributed to its analytical strength here. It’s designed to spot patterns, resolve dependencies, and identify logical flows within complex structures.
It approached my MRI report not as prose, but as a structured problem set, much like it would a complex codebase or a distributed system's architecture.
This analytical framework, while not medical, proved surprisingly effective at identifying subtle interactions.
The Practical Takeaway for Professionals
So, what does this mean for us, the developers and tech professionals navigating the AI era?
It means recognizing the profound, often unexpected, analytical power of these models, even in domains far removed from their primary training.
1. **AI as a "Pattern Amplifier," Not a "Truth Teller":** Use LLMs like Claude Code as tools to surface connections, generate hypotheses, and rephrase complex information.
They can highlight weak signals you might miss.
But every output requires human verification, especially in high-stakes fields. Treat AI outputs as an additional data point, not the definitive answer.
2. **Master the Art of Prompt Engineering for Data Analysis:** My success stemmed from framing the MRI report as a "system state" problem.
Learning to abstract complex real-world scenarios into structured data or system descriptions for AI analysis is a critical skill.
This isn't just for code; it applies to legal documents, financial reports, and even, as I found, medical reports.
3. **Prioritize Privacy and Ethical Boundaries:** Never feed sensitive, un-anonymized data into public LLMs. Understand the limitations and ethical implications of using AI in critical domains.
For medical contexts, this means seeing AI as a tool for *personal understanding* or *question generation*, not *diagnosis* or *treatment*.
4. **Embrace the "Second Opinion" Mindset (Human-Augmented):** AI can give you a different lens, but the final judgment and the *actual* second opinion must come from a human expert.
Use AI to formulate better questions for your human experts, to deepen your own understanding, and to challenge your own assumptions. It's about augmenting human intelligence, not replacing it.
This experiment with Claude Code 4.6 wasn't about curing my shoulder.
It was about leveraging a powerful tool to better understand a complex system — my own body, as described by a machine and interpreted by a human, then re-interpreted by another machine.
It opened my eyes to the subtle ways AI can act as a catalyst for deeper inquiry, even when it's operating far outside its comfort zone.
Have you ever used an AI tool in a completely unconventional way to understand a complex personal problem, or is it just me treating my body like a distributed system?
Let's talk about it in the comments.
---
