🏉💥 Using a “personal ai” through a severe knee injury

TL;DR: Blew out my knee in a rugby match, fed all my personal info to a set of models ( mostly ChatGPT o3 - privacy? non expected), and let it co-pilot me from field to surgery and recovery plan in under 72 hours. The AI sped up my decisions, helped me evaluate options and pointed me in the right directions; but still needed a human filter.

0. The Hit June 30, remote Canadian rugby tournament. One blind‑side tackle, an audible pop, and my right knee looked … wrong. A medial bulge made me blurt, “That’s dislocated.” Teammates helped me off the field , couldn’t really bear weight but could extend and fold to 90 degrees. While I iced the joint and decided if i needed to go to the ER or not I started a chat with o3.

Prompt (paraphrased): “Visible knee deformity, loud crack, can’t bear weight—fracture, dislocation, or ligament blow‑out? List each with on‑field checks and red‑flags.”

LLM:(paraphrased)  1️⃣ Things that could be wrong based on description and probability (Patella dislocation, Ligaments, Fracture along with a list of tests, descriptions and probabilities (based on symptoms)  2️⃣ Recommended immediate care plan  It warned: no forced relocation, check distal pulses, monitor cap refill, request X‑ray ASAP.

We palpated dorsalis pedis (check artery for blood flow to foot):intact; and, with help, I hobbled to the sideline. Deformity lessened as the leg straightened, hinting “sublux then self‑reduced.” A teammate drove me to the local ER.

1. ER & Suspected Ligament Damage Physical exam screamed ligament trauma: lax Lachman, swelling, but still no neurovascular deficit. ER docs were about discharge and refer to orthopedics when I got home in Seattle. We both agreed that basic imaging should be done to rule out fracture.  X‑ray, however, showed a displaced lateral tibial‑plateau fracture (likely Schatzker II). The Canadian IT system had no way to print or email the X-rays or reports (how much of this was true versus just a small, overworked ER doc is anyone’s guess). So, she simply pulled them up on the screen, and I snapped photos on my phone and fed them straight into the models for the next steps.

Prompt (paraphrased): “Translate report, outline next steps, and give questions for the ER team.”

LLM (paraphrased): “Fracture takes priority. CT is gold‑standard to size depression; ask for sedation protocol if pain is high; confirm no compartment syndrome; splint 0–30°, NWB.”

All that matched the attending plan. CT wasn’t available on site; wait‑list was days. Time to head home to Seattle. Model warned likelihood of surgery: High and time sensitive.

2. Building the Surgical Team

Prompt(paraphrased): “Filter Puget Sound sports medicine and trauma orthopedic  surgeons: tibial‑plateau ORIF expertise, fellowship sports, meniscus publications.”

The shortlist backed a number of clinics and orthopedic groups, based on the literature (and my insurance coverage - this is America 🇺🇲)  selected Dr  M and their team. Models suggested the fastest path to surgeon (request first hour available -> seen by orthopedic ARNP -> escalation to surgeon + all imaging requested in parallel) .

The ai drafted a one‑pager: timeline, imaging, questions to clarify.

3. CT & Reality Check July 2, Seattle CT showed ≈ 1 cm articular depression. Feeding the radiology summary back yielded a decision grid:

If the depression is less than 5 mm, treat without surgery. Between 5–10 mm, use screws. More than 10 mm or if you’re an athlete, do a plate and elevation (ORIF). Always check the meniscus for damage.

That matched Dr M’s evaluation later that day and raised my awareness of potential meniscus damage and shifting recovery timelines based on it - off to surgery.

4. Pre‑Op Countdown (Model‑Assisted) Surgery: Jul 3. Checklist generated from ERAS (enhanced recovery after surgery) + my med stack:

Day | Action (cross‑checked)

–1 | Pause omega‑3, NSAIDs.

–1 | Hibiclens night‑before + morning‑of only (per surgeon).

–1 | Confirm brace 0–90°, crutches, printed med list.

0 | Nothing by mouth: solids 8 h, clear fluids ≤ 4 h pre‑op.

Creatine, collagen, L‑glutamine continued.

At this point , the models had informed me of: what to expect during the surgery, potential complications, things to ask my surgical team. Potential variations of the surgical plan that Dr M might offer (and why).

5. ORIF + Meniscus Repair July 3. In surgery, they found a massive capsular avulsion of the lateral meniscus (roots intact) and a 1 cm depression, which they elevated. I got a lateral plate, screws, and a temporary distal-femur fixator. The fluoroscopy report explained the repair as nearly perfect.

Afterwards, I asked the model for a summary: surgical findings, updated rehab timelines, and validation for when to start exercises and increase range of motion.

I tracked pain and range of motion by voice; the LLM automatically graphed pain scores against oxycodone doses, which helped me taper off the meds within 48 hours.

6. Post‑Op Directives & Early Rehab Core orders (Master File):

• Weight‑bearing: Strict NWB 8 wks.
• Brace: hinged 0–90°, on at all times.
• ROM target: 70–80° passive‑assist Week 1 (hit 0→75° by Day 3).

Feeding those plus home‑gym inventory (dumbbells 90 lb, bench, rack) into the model produced seated upper‑body/core cycles with auto‑volume cuts if pain >3/10.

7. Guidelines for AI

The model flagged some exercises that didn’t make sense for my stage of recovery: easy to spot and skip, but still a reminder to double-check every suggestion. It also hallucinated details in search results and sometimes got the sequence of events wrong. Bottom line: it’s a powerful research assistant, but you need to keep your critical thinking switched on.

Problem 1: Hallucinations Prompt 💬 “Summarize my imaging. X‑ray on Jun 30 shows a lateral tibial‑plateau fracture with 8 mm depression. Cite every fact.”

Model 🤖 (incorrect): “CT rules out meniscus damage.” ❌ (There was no CT!)

Fix Prompt 💬 “Same request, but ONLY use quoted text from the attached radiology report, and add a 🔗 after each sentence you can’t find verbatim.”

Model 🤖 (correct) “X‑ray shows a split‑depressed fracture of the lateral tibial plateau (🔗).”

Problem 2: Unsafe exercises Prompt 💬 “Give me 5 core drills.”

Model 🤖 1. Cat‑cow 🐱🐮 … (kneeling)

Fix Prompt 💬 “Context: 6 days post‑op ORIF, R‑leg NWB; no kneeling, flex ≤ 90°. Give 5 core drills, each with why it’s safe.”

Model 🤖

1. Seated pelvic tilt – no knee load ✅
2. Supine dead‑bug – supine, leg supported ✅
3. Seated band‑resisted trunk rotation ✅

Problem 3: Context drift Prompt 💬 “Load master_context.json and confirm surgery date.” Model 🤖 “Loaded. Surgery date: Jul 3, 2025.”

Master Context File: a single JSON containing bio, timeline, restrictions, meds, and goals. I prepend it to every new chat; drift solved. This is also how to transfer all context to another model or AI. I made a redacted version available in comments below. 

Rule of Thumb 🟡 End every AI answer with “Human clinician review required”.

Turbo‑charged research assistant, yes; clinical oracle, no. 

8. Status 

By post-op Day 6, I was off narcotics, pain was under 2/10, and I’d regained 80° ROM and quad activation. The model projects partial weight-bearing at Week 8, stationary bike by Week 9, and a cautious jog by Month 6. Though full rugby speed may be done, running and lifting are still in play.

Here’s what the experiment proved:

• LLMs compress expertise: They can synthesize imaging, rationale, and surgical decision trees in minutes.
• Agency amplification: AI cuts admin friction, sharpens consults, and makes you a better-informed patient.
• Human gatekeeping is non-negotiable: I had to spot and reject bad exercise suggestions and timeline errors: AI is not infallible.
• Granular data yields granular advice: Tracking pain and ROM let the model personalize rehab and taper meds safely.

9. Rehab plan and moving forward

Looking forward, AI will keep assisting—flagging rehab milestones, surfacing new protocols, and helping me pressure-test advice. But the real value is in the partnership: AI for speed and options, human judgment for the final call.

Weekly Loop

Prompt 💬 (Sunday night) {ROM_deg: 82, QuadTorque_Nm: 45, Pain_VAS: 1, Sleep_Hours: 7.2, Surgeon_Note: "Continue NWB; focus on quad activation.", PT_Note: "Goal: 90° flex by Wk‑2."} “Compare these to protocol milestones and propose next‑week plan. Tag any red‑flags (> 10 % variance).”

Model 🤖

• ✅ ROM ahead (target 80°).
• ⚠️ Quad strength 12 % below curve.
• Plan: add NMES quad sets (3×20) + seated knee‑flexion slides +5 reps/day.
• No red‑flag pain spikes.

Prompt 💬 (auto) “Generate SVG dashboard for surgeon review + 3 follow‑up questions I should ask.”

Model 🤖 (links to chart) Questions:

1. Should we start PWB earlier given ROM progress?
2. Need updated bone‑healing X‑ray schedule?
3. Is NMES intensity safe with hardware in situ?

Escalation Logic

Prompt 💬 (AI trigger) “Pain_VAS jumped from 2→5 for two sessions; swelling + 2 cm. Suggest diff‑dx & imaging.”

Model 🤖 “Possible synovitis vs hardware irritation. Recommend urgent consult; obtain AP/Lateral X‑ray + CRP.”

That prompt‑response handshake drives a metrics → model → clinician → athlete loop until sprint speed matches pre‑injury baseline.

Finish Line Goal 🎯 Top‑speed ± 5 % and symmetrical hop distance by Month 12: or sooner if the data say so.

⚠️ Medical Disclaimer

I’m not a clinician.. Do not rely on this post for personal medical decisions.