Narrative Review | Critical Care & Emergency Medicine
Narrative Review: Reinforcement Learning for Optimising Vasopressin Initiation in Septic Shock
Time to read: 03:08
Time to listen: 05:59
Originally Published: 18 May 2024
Source: JAMA
Type of article: Narrative Review
MedED Catalogue Reference: MIIB017
Category: Critical Care & Emergency Medicine
Cross Reference: Cardiovascular
Keywords: shock, cardiovascular, haemodynamics, reinforced learning models, LLMs
Top
This article is a review of recent studies originally published in the JAMA, 18 May 2025. This article does not represent the original research, nor is it intended to replace the original research. Access the full Disclaimer Information.
Norepinephrine is the first-line vasopressor for patients with septic shock.1
However, when and whether to initiate a second agent, such as vasopressin, remains uncertain and robust trial-based guidance is lacking. This gap presents a significant clinical challenge, particularly given the dynamic and heterogeneous nature of septic shock physiology.5
In the Optimal Vasopressin Initiation in Septic Shock (OVISS) study, published in JAMA on May 18 2025, researchers Kalimouttou, Kennedy, & Feng et al, aimed to derive and validate a reinforcement learning (RL) model to identify an optimal rule for vasopressin initiation in adult patients with septic shock receiving norepinephrine. The primary outcome they were interested in was the impact of each group's decision-making on in-hospital mortality.
Why Use Reinforcement Learning?
Traditional clinical trials are often limited in addressing nuanced, time-sensitive treatment decisions, particularly when patient responses evolve rapidly. Reinforcement learning (RL)— a branch of machine learning—offers a powerful alternative. In RL, a virtual agent learns decision strategies through interaction with data-rich environments, optimising actions (such as drug initiation) based on observed outcomes. This makes RL especially well-suited for septic shock, where treatment effects depend on the timing, sequence, and context of interventions.3
Study Methodology
The study started by developing the model using data from 3,608 patients treated at UCSF between 2012 and 2023. Most of these patients (57%) were men, with a median age of 63. At the time they went into shock, their median SOFA score—a measure of organ failure—was 5.
To see how well the model would work in other settings, the researchers tested it on three large U.S. datasets: MIMIC-IV, eICU-CRD, and UPMC. These included 10,217 patients from 227 hospitals. In this broader group, the median age was slightly higher, at 67, and the SOFA score at shock onset was slightly worse, at 6.
To evaluate the model’s accuracy and usefulness, the team used advanced statistical methods, including weighted importance sampling and logistic regression adjusted for patient differences.
Findings
Compared to real-world clinical decisions, the model recommended vasopressin for a significantly higher proportion of patients (87% vs. 31%), initiated treatment earlier in the course of shock (median 4 vs. 5 hours after onset), and at lower norepinephrine doses (median 0.20 vs. 0.37 µg/kg/min).
Importantly, adherence to the model's recommendations was associated with improved clinical outcomes. Specifically, the model yielded a higher expected reward and was linked to reduced in-hospital mortality. Patients whose treatment aligned with the model’s guidance had a 19% lower adjusted odds of dying in hospital (odds ratio 0.81, 95% CI 0.73–0.91).
These findings were consistent across all external validation datasets, suggesting that the model’s recommendations may offer a more effective and timely approach to vasopressor therapy in septic shock.
Importantly, the model consistently outperformed clinician practice across all datasets.
Implications
The OVISS study provides a compelling demonstration of how reinforcement learning (RL) can be used to guide complex treatment decisions in critical care, suggesting that AI-driven tools may help refine practice in areas where evidence is limited and clinical variation is high.
Current international guidelines recommend vasopressin as a second-line agent when mean arterial pressure remains low despite norepinephrine at moderate doses.1,2 However, these thresholds are loosely defined, and no randomised clinical trials have established the optimal timing, dose, or tapering strategy for vasopressin.5,6 In this context, reinforcement learning offers a data-driven framework capable of exploring a vast range of treatment pathways that conventional trial designs may miss.
Importantly, the OVISS RL rule proposed earlier vasopressin initiation at lower norepinephrine doses—findings that align with subgroup signals from the Vasopressin and Septic Shock Trial (VASST),4,5 which hinted at benefits among patients with less severe shock.6
The rule may outperform clinician judgment by recognising subclinical physiological patterns or weighting variables differently than human decision-makers. From a pathophysiological perspective, early vasopressin may address relative hormone deficiency in septic shock, enhance glomerular filtration pressure, and reduce the need for kidney replacement therapy. Its catecholamine-sparing effect may also mitigate the adverse consequences of high-dose norepinephrine, such as tachyarrhythmias or myocardial ischemia.
Moreover, adherence to the RL rule may serve as a proxy for higher clinical vigilance and earlier intervention, markers consistently associated with improved sepsis outcomes. While residual confounding cannot be excluded, the robustness of findings across diverse external datasets and clinical subgroups strengthens the case for prospective validation.
The study marks a first in applying RL to vasopressor titration and opens a new avenue for integrating machine learning into bedside decision support.
If validated in randomised trials, RL-informed protocols could standardise and personalise care in septic shock, potentially improving survival and organ protection in this high-risk population.
Access the original study
Kalimouttou, A., Kennedy, J. N., Feng, J.,et al (2025). Optimal Vasopressin Initiation in Septic Shock: The OVISS Reinforcement Learning Study. JAMA, 333(19), 1688–1698. https://doi.org/10.1001/jama.2025.3046
Additional References
1. Singer M, Deutschman CS, Seymour CW, et al. The third international consensus definitions for sepsis and septic shock (Sepsis-3). JAMA. 2016;315(8):801-810. doi:10.1001/jama.2016.0287
2. Vincent JL, Moreno R, Takala J, et al. The SOFA (Sepsis-related Organ Failure Assessment) score to describe organ dysfunction/failure. Intensive Care Med. 1996;22(7):707-710. doi:10.1007/BF01709751
3. Khezeli K, Siegel S, Shickel B, Ozrazgat-Baslanti T, Bihorac A, Rashidi P. Reinforcement Learning for Clinical Applications. Clin J Am Soc Nephrol. 2023 Apr 1;18(4):521-523. doi: 10.2215/CJN.0000000000000084. Epub 2023 Feb 8. PMID: 36750034; PMCID: PMC10103233.
4. Saria S. Individualized sepsis treatment using reinforcement learning. Nat Med. 2018;24(11):1641-1642. doi:10.1038/s41591-018-0253-x
5. Russell JA, Walley KR, Singer J, et al; VASST Investigators. Vasopressin versus norepinephrine infusion in patients with septic shock. N Engl J Med. 2008;358(9):877-887. doi:10.1056/NEJMoa067373
6. Nagendran, M., Russell, J. A., Walley, et al. (2019). Vasopressin in septic shock: an individual patient data meta-analysis of randomised controlled trials. Intensive care medicine, 45(6), 844–855. https://doi.org/10.1007/s00134-019-05620-2
Disclaimer
This article is compiled from several resources researched and compiled by the contributor. It is in no way presented as an original work. Every effort has been made to attribute quotes and content correctly. Where possible all information has been independently verified. The Medical Education Network bears no responsibility for any inaccuracies which may occur from the use of third-party sources. If you have any queries regarding this article contact us
Fact-checking Policy
The Medical Education Network makes every effort to review and fact-check the articles used as source material in our summaries and original material. We have strict guidelines in relation to the publications we use as our source data, favouring peer-reviewed research wherever possible. Every effort is made to ensure that the information contained here is an accurate reflection of the original material. Should you find inaccuracies, out of date content or have any additional issues with our articles, please make use of the contact us form to notify us.
