AI in the Cancer Journey: How I'm Using AI to Help My Son

Nathan recounts the rapid progression from seemingly normal symptoms to cancer diagnosis, culminating in a pivotal moment where AI models identified a critical red flag (elevated LDH) that human doctors in New Orleans had dismissed. This section covers the exponential nature of cancer growth and the importance of early AI analysis.

• •Burkitt Leukemia has a 24-hour doubling time, making it the fastest-growing cancer but also highly treatable with 80%+ cure rates
• •Multiple AI models (GPT-5 Pro, Claude, Gemini) correctly flagged elevated LDH as concerning when human doctors dismissed it due to sample hemolysis
• •Claude's alarmist tone was initially off-putting but may have been appropriate given the urgency - different presentation of same screenshot data led GPT-4 to make the same mistake as human doctors
• •Key lesson: Use multiple AI models with different data presentations (text vs screenshots) to catch potential diagnostic errors
• •Video documentation of worst symptoms and detailed symptom logs are critical for effective doctor communication

Nathan analyzes the different approaches of Claude and GPT-5 Pro in delivering urgent medical information, revealing important insights about AI persona design and user preferences in high-stakes situations.

• •Claude 4.1 Opus was highly alarmist and emotional, causing a panic attack but potentially necessary to overcome denial
• •GPT-5 Pro provided longer, more clinical, less emotional responses that Nathan preferred for ongoing use
• •The 'virtuous stranger' persona design may have led Claude to be appropriately urgent for someone who tends to downplay concerns
• •Nathan shifted away from Claude despite its accuracy due to tone preferences, highlighting tension between effectiveness and user comfort
• •Gemini 2.5 was consistently good but didn't add significant differentiation in this phase

The critical 48-hour period where Nathan and his wife pushed doctors to start chemotherapy before final diagnosis, demonstrating effective patient advocacy informed by AI analysis and the importance of acting on parental intuition.

• •Doctors initially insisted on waiting for definitive diagnosis before treatment, citing standard protocol
• •Nathan's advocacy, informed by AI understanding of urgency, pushed for treatment to start by Friday to avoid weekend delays
• •Child's statement 'I can't move my legs' became the pivotal moment that convinced oncologist to break protocol
• •Treatment started Wednesday (one day before final diagnosis) likely prevented serious organ failure
• •Key insight: In exponential situations, every 24-hour delay can be critical - advocate forcefully when AI and intuition align

Nathan describes his systematic approach to using GPT-5 Pro as a continuous second opinion system, analyzing lab results every 6 hours and providing confidence in the medical team's decisions while enabling informed advocacy.

• •Created long-running thread with GPT-5 Pro analyzing ~20 lab result updates over 10 days, approaching 100-page prompt length
• •AI confirmed treatment protocol was state-of-the-art and consistent across all major centers, reducing anxiety about needing to transfer
• •Minor differences: AI slightly more aggressive on mineral supplementation (calcium, magnesium) and wanted more confirmatory tests
• •AI provided education enabling informed conversations with doctors rather than blind trust or constant anxiety
• •Without AI, Nathan imagines 'constant Googling, constant anxiety, never feeling like I have enough information' - AI provided empowerment and peace of mind

A powerful example of AI's unique value: synthesizing expertise from completely different domains (oncology and mold remediation) to solve a complex real-world problem that would be nearly impossible to address through traditional means.

• •Basement flood created mold risk during chemotherapy when fungal infections are particularly dangerous
• •AI synthesized knowledge from mold remediation, air quality testing, and pediatric oncology to create actionable plan
• •Coordinating a meeting between oncologist and mold remediation specialist would be practically impossible - AI bridged this gap instantly
• •Solution: 8-10 HEPA filters running continuously will provide cleaner air than outdoor environment during leaf season
• •Demonstrates AI's ability to handle multi-domain problems that fall between traditional expert silos

Nathan uses what he believes is a Gemini 3 checkpoint to plan for potential cancer recurrence, revealing significant improvements in model speed and response quality while preparing for a low-probability but catastrophic outcome.

• •Gemini 3 is ~5-10x faster than GPT-5 Pro with more focused, sophisticated responses that are easier to digest
• •First time Gemini may have a persona advantage - feels like 'talking to the most brilliant person you've ever met' vs GPT-5 Pro's information dumps
• •Used to identify CAR-T therapy and other cutting-edge treatments for relapsed cases, including specific clinical trials and centers
• •Created emergency plan for immediate action if cancer returns (10-15% probability but poor prognosis)
• •Demonstrates AI's value in scenario planning and research that would take months of manual work

Practical lessons on getting medical data into AI systems, including the challenges of poor health data infrastructure and specific techniques for accurate transcription of complex medical documents.

• •Patient portals have terrible data export - often required manual copy-paste or phone photos of paper printouts
• •For crisp website screenshots, direct image input works fine; for photos of documents, transcription quality varies significantly
• •Gemini 2.5 Pro is best at transcribing challenging documents (faxed reports, tables, ad-hoc layouts) but still requires systematic verification
• •Critical medical data requires 100% accuracy - always verify transcriptions, especially for cell markers and genetic information
• •Use 'print to PDF' function from patient portals when available, then have AI transcribe to text for better model performance

Nathan shares techniques for managing complex medical cases with 100+ page prompts, including creating AI-generated handoff documents that distill comprehensive case histories into actionable summaries.

• •Models now handle 100+ page prompts effectively - Amanda Askell's use of multiple 100-page prompts validated this approach
• •Standard prompt includes: symptom history, MRI/CT/PET reports, spinal fluid analysis, bone marrow analysis, pathology reports with cellular markers
• •Key technique: Ask AI to convert long thread into 'handoff document for new attending physician' - creates information-dense 2000-word summary
• •Handoff documents enable starting fresh threads without losing context, useful for trying different models or approaches
• •Demonstrates shift from 'don't overwhelm the context' to 'provide comprehensive information for best analysis'

Using AI to develop a comprehensive infection prevention strategy during immunosuppression, including COVID-style quarantine protocols and risk assessment for the family's specific situation.

• •Chemotherapy will severely compromise immune system - bacterial and fungal infections are major risks (viral less concerning)
• •Family can implement strict social distancing due to work-from-home situation and only one other child in school
• •No travel, speaking engagements, or AI events for 6-month treatment roadmap
• •AI helped assess relative risks and develop practical protocols beyond what doctors discussed
• •Some infections unavoidable (from bacteria already in mouth/skin/gut) but external exposure can be minimized

Nathan uses AI to research cutting-edge treatments for potential relapse scenarios, identifying specific CAR-T therapy trials and medical centers, demonstrating AI's value in proactive worst-case planning.

• •Burkitt Leukemia has bimodal outcome: 85%+ cure rate, but if it relapses, prognosis is poor (most children die)
• •AI identified CAR-T cell therapy as leading-edge treatment for relapsed cases, with specific clinical trials and centers
• •Created comprehensive emergency plan to execute immediately upon any recurrence detection
• •This level of research would traditionally take months and require connections to multiple specialists
• •Still awaiting genetic testing to determine exact subtype, which affects recurrence risk and treatment approach

Nathan synthesizes lessons learned about using AI in medical crises, including specific model recommendations, data presentation strategies, and the transformative potential of AI as a medical advisor accessible to everyone.

• •Use GPT-5 Pro ($200/month) for serious medical situations - regular GPT-4 made same mistakes as human doctors in some cases
• •Run multiple models with different data presentations (text vs screenshots) to catch errors and get comprehensive views
• •Record all appointments, take video of worst symptoms, maintain detailed symptom logs for effective doctor communication
• •AI enables n-of-many second opinions vs limited human doctor access, with analysis quality comparable to human physicians
• •Three key use cases: (1) bedside treatment monitoring, (2) complex multi-domain problem solving, (3) scenario planning for rare outcomes
• •AI provided empowerment, confidence, and peace of mind that would be impossible to achieve through traditional means alone
• •Critical to advocate forcefully when AI analysis and parental intuition align, especially in exponential situations