2026 Critical Care Congress Skip to main content
2026 Critical Care Congress Main banner
Final Program


Advance Program
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.
Session Icons
Thought Leader Session
Thought Leader Session
Guidelines Session
Guidelines Session
Aging Adult Track
Aging Adult Track
Additional Registration Fee
Additional Registration Fee

Research Snapshot Theater: Precision Medicine in Sepsis, Adult and Pediatric

First Evaluation of a Real-Time Reinforcement Learning Decision Support System for Sepsis in the ICU

Sunday, March 22, 2026
2:30 PM - 3:30 PM Central Time
Location: Connections Central - RST 18

  • Matthieu Komorowski, n/a

    GBR

    Disclosure(s): GE Healthcare: Speaker/Honoraria (includes speakers bureau, symposia, and expert witness) (Ongoing); Philips Healthcare: Consulting fees or other remuneration (Ongoing), Speaker/Honoraria (includes speakers bureau, symposia, and expert witness) (Ongoing)

  • LL

    Luchen Li, N/A

    Dr
    Imperial College London

    Disclosure information not submitted.

  • QE

    Qamar El-kotob, N/A

    Ms
    Imperial College London

    Disclosure information not submitted.

  • Paul Festor, N/A (he/him/his)

    England, GBR

    Disclosure(s): STRIVE HEALTH LIMITED: Employee (any industry) (Ongoing), Other Financial or Material Support (Ongoing), Salary (Ongoing)

  • MN

    Myura Nagendran, N/A

    Dr
    Imperial College London, United Kingdom

    Disclosure information not submitted.

  • MN

    Melody Ni, N/A

    Dr
    Imperial College London, United Kingdom

    Disclosure information not submitted.

  • PB

    Peter W Buckle, N/A

    Pr
    Imperial College London, United Kingdom

    Disclosure information not submitted.

  • MM

    Massimo Miccoci, N/A

    Dr
    Imperial College London, United Kingdom

    Disclosure information not submitted.

  • MZ

    Maria Bernarda Zamora Talaya, N/A

    Dr
    Imperial College London, United Kingdom

    Disclosure information not submitted.

  • HK

    Hannah Kettley-Linsell, N/A

    Dr
    Imperial College London

    Disclosure information not submitted.

  • EF

    Elizabeth O Fagbodun, N/A

    Ms
    Imperial College London, United Kingdom

    Disclosure information not submitted.

  • BG

    Ben Glampson, n/a

    Mr
    Imperial College Healthcare NHS Trust

    Disclosure information not submitted.

  • SH

    Steve Harris, N/A

    Dr
    University College London, United Kingdom

    Disclosure information not submitted.

  • DB

    David Brealey, N/A

    Dr
    University College London Hospitals NHS Foundation Trust

    Disclosure information not submitted.

  • AF

    Aldo A Faisal, N/A

    Pr
    Imperial College London

    Disclosure information not submitted.

  • AG

    Anthony Gordon, N/A

    Pr
    Imperial College London

    Disclosure information not submitted.

Introduction: Sepsis remains a major global challenge requiring timely, individualised administration of intravenous fluids and vasopressors. Reinforcement learning (RL) offers a promising approach for optimising sequential treatment decisions by learning from retrospective hospital data. However, RL systems have not yet been deployed in real-time sepsis care. We report the first real-time evaluation of an RL-based clinical decision support system for sepsis resuscitation. The AI Clinician 2 model uses continuous patient states and drug dosing action spaces, with a Bayesian, distributional RL architecture trained on 20,587 ICU sepsis cases from MIMIC-3 database (USA).


Methods: A two-phase study was conducted across 4 UK ICUs in 2 different hospital systems. ICU patients with sepsis were prospectively included (2020-2025). A sample of these patients were enrolled in 2 experiments: XP1 (shadow mode) where AI-generated recommendations were reviewed off-site by ICU clinicians, and XP2 (live mode), where AI suggestions were shown to clinicians during active patient care. We used inverse probability of treatment weighting (IPTW) in the complete patient cohort to estimate the effect of AI recommendations by adjusting for confounders. Propensity scores were estimated on 26 clinical covariates to model the probability of clinician-AI agreement. Stabilised weights were used in weighted logistic regression models to estimate the effect of agreement on in-hospital mortality.


Results: In XP1 (65 patients, 394 decisions), 71% of fluid recommendations and 59% of vasopressor recommendations were rated as “just right” by all 11 clinicians. In XP2 (49 patients, 66 decisions, 25 clinicians), these proportions increased to 87% for fluids and 70% for vasopressors. In the IPTW analysis (860 patients, 58,099 decisions), agreement between clinicians and the AI was associated with lower odds of in-hospital mortality: OR of 0.85 (95%CI: 0.82–0.89; p< 0.001) for fluids and 0.46 (95%CI: 0.44–0.49; p< 0.001) for vasopressors.


Conclusions: This first real-time deployment of an RL system for sepsis care demonstrated high clinician concordance and favourable safety and survival outcome associations. These findings represent a key step in translating AI from retrospective analysis to real-time treatment recommendation in critical care.