Thomas Frost

I'm a doctor and PhD researcher at University College London, where I work on offline reinforcement learning for intensive care. My thesis explores the real-time optimisation of titratable drug infusions in the ICU using naturally timed data (as featured on TalkRL).

I trained in medicine at Oxford and have spent more than eight years working in the NHS in emergency medicine, first in Scotland and then in London. I still practice medicine and use my clinical experience to inform a lot of my AI-related research. My long-term goal is the real deployment of autonomous decision-making algorithms directly into the patient bedside.

Email · GitHub · LinkedIn · Google Scholar · CV

Research interests

My work sits at the intersection of reinforcement learning and clinical interventions. Clinicians are frequently inconsistent in their decisions, which leads to suboptimal care for patients. Offline reinforcement learning may help us to address this problem. But learning a treatment policy from historical intensive care data means confronting a range of interesting challenges – including delayed rewards, unmeasured confounding, and policy evaluation.

Alongside the thesis, I’ve also served as an expert clinical evaluator for LLM-generated discharge summaries, and worked with Microsoft and UCLH on FlowEHR, an open-source MLOps platform for testing and deploying models inside clinical workflows.

Projects

• Offline Reinforcement Learning for Clinical Data with Natural Timings

  PhD thesis project

  My thesis looks at problems around the manipulation of data timings in retrospective datasets (specifically, data binning), and proposes alternative approaches using realistic, naturally timed data.

  The work covers Insulin4RL (my freely available dataset for offline RL with naturally timed data); real-time mortality prediction in the ICU using temporal-difference learning; and an end-to-end pipeline for insulin infusion optimisation using mortality-based rewards.

• Insulin4RL: Real-time insulin infusions for offline reinforcement learning

  Open-source dataset · PhysioNet, 2026

  My freely available dataset of real-time insulin infusions in intensive care, intended for offline reinforcement learning with naturally timed data.

  Dataset

• FlowEHR

  Contributor · Microsoft & UCLH

  Open-source MLOps platform for secure model development and deployment inside NHS trusted research environments. I acted as a research user and validator, feeding clinical requirements back into the platform design.

  Project site

Selected publications

• Insulin4RL: Real-time insulin management in the intensive care unit for offline reinforcement learning

  Frost, T., & Harris, S. — arXiv preprint arXiv:2606.19481, 2026

  arXiv

• The hidden risks of temporal resampling in clinical reinforcement learning

  Frost, T., Vaidya, H., & Harris, S. — arXiv preprint arXiv:2602.06603, 2026

  arXiv

• Robust real-time mortality prediction in the intensive care unit using temporal difference learning

  Frost, T., Li, K., & Harris, S. — Proceedings of the 4th Machine Learning for Health Symposium, PMLR, 2025

  Paper arXiv Code

• Automated generation of hospital discharge summaries using clinical guidelines and large language models

  Ellershaw, S., Tomlinson, C., Burton, O. E., Frost, T., et al. — AAAI 2024 Spring Symposium on Clinical Foundation Models

  Paper Code

• Cardiovascular disease risk and prevention amongst Syrian refugees: mixed methods study of Médecins Sans Frontières programme in Jordan

  Collins, D. R., Jobanputra, K., Frost, T., et al. — Conflict and Health, 11(1), 14, 2017

  Paper

• Should assisted dying be legalised?

  Frost, T. D., Sinha, D., & Gilbert, B. J. — Philosophy, Ethics, and Humanities in Medicine, 9(1), 3, 2014

  Paper