The Current State of AI in Rheumatology — Virginia Society of Rheumatology 2025 Annual Meeting.

Paul SufkaConferences, Education, Med Tech

Objectives

  • Define the major types of artificial intelligence relevant to healthcare
  • Describe current and emerging applications of AI in rheumatology
  • Evaluate the limitations, ethical concerns, and future directions of AI integration in rheumatologic practice

AI Basics

  • Definitions & hierarchy: Algorithms → Machine Learning → Deep Learning → Transformers → Large Language Models (LLMs)
  • Transformers introduced self-attention (Vaswani et al. 2017)
  • Pretraining concepts: tokenization (OpenAI Tokenizer), embeddings, Q/K/V attention, multi-head attention, transformer blocks (Interactive explainer)
  • Role of GPUs in AI training, FLOPS, and NVIDIA’s dominance

Evolution of Models

  • GPT-3 (2020): 175B parameters, first to capture public attention
  • GPT-3.5 (2022): Refinements, powered first ChatGPT release
  • GPT-4 (2023): 1–1.8T parameters, major leap in reasoning, high exam scores
  • GPT-4o (2024): Multimodal, real-time capabilities
  • GPT-o3 (2024): Optimized for reasoning efficiency
  • GPT-5 (2025): >10T parameters (est.), frontier model, improved efficiency

Leaderboards and benchmarks:

Artificial General Intelligence (AGI)

  • No clear definition; can it perform any intellectual task a human can do?
  • Apple’s perspective: The Illusion of Thinking

Using General AI Tools

  • Different “personalities” across ChatGPT, Grok, Claude, Gemini
  • Prompt engineering: role → task → context → format → extras

Current Uses in Medicine

AI Agents

  • Chatbots answer single questions
  • Agents execute multi-step workflows with oversight
  • Non-healthcare agents: email, browser, coding (Cursor), productivity (Microsoft Copilot, Google Gemini apps)
  • Spreadsheet integrations: Google Sheets =AI(...), Excel =COPILOT(...)
  • EHR agents: Epic, Oracle Health

Risks & Limitations

  • Errors & hallucinations: confident but wrong answers, fabricated references
  • Liability: unclear responsibility, evolving regulatory standards
  • Ethics: bias, transparency, patient trust
  • Why models hallucinate: OpenAI report

Future Directions

  • Deeper integration into everyday tech and EHRs
  • Agentic workflows
  • Multimodal healthcare inputs (history, labs, imaging, pathology, genomics)
  • Regulatory evolution
  • Workforce impact: efficiency ↑, burnout ↓

Practical Tips

  • Create 2+ AI accounts (ChatGPT, Gemini, Claude, Grok)
  • Try an ambient scribe
  • Test decision tools (OpenEvidence, Vera Health)
  • Practice structured prompting
  • Use AI to summarize journal articles

Thanks to everyone who attended!