woke up early and rushed to the conference. i had one goal here, to talk to anyone relevant and ask about OE for my chat tomorrow. and also take some photos of berkeley as a side quest. i did eventually talk to emily who coauthored a paper by OE, and also a few phd students in the CPH program. i liked these smaller conferences, some of the presenters and people in the panel are doing cool work in the AI for healthcare space. i want to devote my life to this field.

some notes i wrote -> cleaned up by gpt 4.5

CHIL highlights

• AI regulations:

  • Pre‑2024 context; HTI‑1.
  • EO 14179 (Jan 2025).
  • OBBB (H.R. 1): House version in May 2025 proposed a 10‑year state/local AI regulation ban; provision was later removed and did not make it into the final law.

Jupyter Health

• What it is: DevOps platform for health data and analytics pipelines.

• Question to leaders: “In your role leading UCSF’s research informatics strategy, how do you evaluate tools like OpenEvidence that bridge research literature and clinical practice? What metrics or frameworks do you use to assess impact on research workflows?”

• EHR realities:

  • Database of patient info and clinical actions (procedures).
  • Also the workflow backbone for care delivery; major investment in Epic.

• SMART on FHIR:

  • Exposes EHR to external apps; enables innovation outside Epic.

• Patient‑generated data (PGD):

  • Standardization is hard (OpenHealth attempts to help).
  • Vendors like Garmin and Oura keep data in proprietary clouds; often undocumented/uncalibrated.
  • Jupyter Health → CommonHealth ingestion to standardize for model training.

• Patient‑facing tooling:

  • Columbia × Jupyter Health: reusable tools for patient outreach.

• CGM data:

  • Many clinicians aren’t trained for continuous streams.
  • Partner with endocrinology on clinical markers & interpretation.
  • Build a starter library of models/algorithms for these use cases.

Panel — ethics, evaluation, and systems (moderator: Emily Alsentzer)

• Themes: publish‑or‑perish; public image of scientists; corporate control; individualist culture; capitalism sets the script.

• "New" gen‑AI ethics issues are old challenges in new clothes; disproportionate impacts on communities of color.

• AI should support clinicians; include philosophers in teams; rethink progress/efficiency; redesign systems, not just models; preserve human agency.

• Rethink paper structure (beyond IMRaD) to keep human decision‑making visible; requires political will.

• Build human‑AI systems that are humble, curious, empathetic; fix perverse incentives that remove joy from learning.

• AI literacy: nurses and patients need it; curricula should revamp yearly; emphasize critical thinking.

• Benchmarks ≠ evaluation:

  • Benchmarks drive procurement/VC, but don’t equal real‑world performance.

  • Principles for eval frameworks:

    1. Hyperlocal: hospital‑ and task‑specific; organize real‑world eval where patients define what matters.
    2. Agile & reflexive: adaptable and effective over time.
    3. Continuous & community‑operated: from the ground up, not just NEURIPS/CSAIL.

  • Always ask: What is this evaluation for? Model selection ≠ clinical deployment.

• People:

  • Jenna C. Lester
  • Leo Anthony Celi
  • Pranav Rajpurkar
  • Deborah Raji

Guillermo Sapiro — wearables & biological age

• Healthspan focus: live well to the last day.
• Measuring biological age: traditional (blood/saliva, imaging) vs. PPG alone.
• PPG: green/IR light (Apple Watch, Oura); light through skin; vessel dilation shapes waveform.
• Rep‑learning setup: ~180k participants; PPG segments; encoder; participant‑level contrastive learning; each person → 512‑dim vector.
• Healthy‑age model: train on healthy subset; average reps; predict age; compare predicted vs chronological.
• Results: healthy cohort ~2.43 MAE (wearables), ~2.5 MAE (blood tests), ~7 MAE (ECG); general cohort ~3.18 MAE.
• Takeaways: heart‑rate waveform often treated as non‑identifying; biological age tracks disease/behavior; larger age gap → higher disease risk; sensitive to pregnancy; learn from healthy distribution to detect OOD.

Research papers & posts

• Audio models: CaReAQA: Cardiac & Respiratory Audio QA for Open‑Ended Diagnostic Reasoning
• VR + rehabilitation: Immersive Virtual Reality and Robotics for Upper Extremity Rehabilitation
• Multimodal genetics: Unlocking rich genetic insights through multimodal AI with M‑REGLE
• Perspective: The Bitter Lesson

Policy & regulation — fact links

• HTI‑1 Final Rule (ONC): Program page · Overview (PDF)

• Executive Order 14179 — Removing Barriers to American Leadership in AI: Presidency (UCSB) · Federal Register (publication)

• OBBB (H.R. 1) — 10‑year AI regulation moratorium (not enacted):

  • House added the 10‑year ban (May 2025)
  • Senate struck the ban (July 2025)
  • Final law did not include the ban