Agentic AI and biostatistics: one-way or two-way street? And how? Sharing the deck used in our lab discussion today.
Context: AI that plans, calls tools, and loops until a task is done is becoming more real. This week alone, Nature published “the anatomy of personal health agent”: a system that writes statistical analysis plans and code over wearable data, evaluated with 7,000+ human annotations.
Prompted by this, I asked my biostatistics group 4 uncomfortable questions:
→ What is one task you would hand to an agent from the next week/month? What is the one you’d never hand over? Is that because of quality, trust, or because “the struggle is where you learn”?
→ Which of the methods we already work on become more valuable in a space where AI agents are used routinely, and which become less?
→ If an agent ran the main analysis in our next paper, what would we need to log, disclose, and be able to re-run before we would put our names on it?
→ Would you let an agent interpret your own sleep and glucose data? Would you let it interpret a study participant’s? Implications for new health information sharing, interpretation, and population studies?
My take: for statisticians, agents are not a threat to the field but a stress test of it. Uncertainty quantification for model outputs, evaluation design, and causal reasoning about sequential decisions just became more valuable, not less. Especially in low-resource settings, where data may be sparse and where these systems are least tested and statistical wisdom may help.
The deck includes a reading list covering foundations, agents-as-scientists, and critiques.
Run it with your own group. I’d love to hear what prompts you’d add.