AI Frontiers: Annie Hill (OpenAI DevDay)

Boston Children’s Hospital is positioning generative AI as a practical tool to reduce healthcare burden, improve learning, and strengthen patient safety—while building guardrails around equity and inclusion. Annie Hill, from the hospital’s Innovation & Digital Health Accelerator, frames the work around three priorities: operational efficiency (especially for staff and patient-facing materials), clinical support that reduces time spent hunting for information, and research acceleration for labor-intensive qualitative analysis.

A central theme is matching real “pain points” gathered from staff to areas where large language models can deliver measurable value. Hill describes collecting challenges across the hospital and sorting them into three buckets: operational, clinical, and research. Operationally, one standout need is patient education—creating materials that work across multiple languages and reading levels. The hospital wants AI to make translation and adaptation more efficient without sacrificing accessibility for a diverse pediatric audience.

In clinical settings, the focus is not on replacing clinical decision-making with AI recommendations. Instead, the goal is to reduce workload and put the right information in front of clinicians when it matters. Hill highlights Swirl, a provider-facing system that aggregates patient data from multiple sources—notes, labs, medications, orders, and forms—into a real-time patient view. The proposed upgrade is an LLM integration that lets providers ask patient-specific questions and receive answers grounded in the aggregated record. Hill also emphasizes traceability: responses would link out to underlying sources for validation, aiming to cut down the time clinicians spend switching between platforms.

Swirl is also a testbed for context-driven error detection and alerting. Traditional alerting can overwhelm clinicians and contribute to alarm fatigue. By using the patient record as context, the system could identify potential errors more accurately—for example, recognizing when a medication previously prescribed becomes unsafe after a new or recent medical event. The intended payoff is fewer clinical mistakes and safer care, not just fewer alerts.

On the research side, Hill points to the heavy lift of working with large qualitative datasets such as interview data. The hospital is exploring using LLMs as a first-pass analysis tool to speed up initial interpretation before deeper review.

For education, the hospital is developing MedTutor, an interactive medical tutor for learners including medical students, residents, and fellows. MedTutor will guide users through medical cases that are currently non-personalized, non-interactive, and hard to access. The plan is to fine-tune a model using hospital education data, start narrowly by generating cases for specific disease areas, and then adapt the case walkthrough based on the learner’s responses—aiming to improve both access to practice and the quality of learning through personalization.

Finally, Hill underscores that equity is not an afterthought. Boston Children’s is building an LLM implementation equity guideline to embed equity, diversity, and inclusion principles into AI research and development, including how models are evaluated and deployed across different patient and learner populations. The overall message: generative AI is being treated as an operational and clinical workflow tool—implemented carefully, grounded in hospital data, and designed to reduce burden while improving safety and education outcomes.