Written by Jason Lesnick
Spoon Feed
This retrospective cohort study derived, validated, and measured the treatment effect of using a reinforcement learning model by reviewing EHR data and found that when patients received care consistent with what the model would have recommended, there was an association with decreased in-hospital mortality.
AI says (A)Initiate vasopressin … now!
The first-line vasopressor for patients with septic shock is norepinephrine, but it is unclear when and even if a second-line agent should be added. This study attempts to give some guidance to clinicians on when to consider starting vasopressin in addition to norepinephrine.
This multicenter retrospective cohort study utilized reinforcement learning applied to EHR data from 14,453 adult ICU patients across four datasets from 2012-2023. The authors wanted to know if reinforcement learning could help identify the optimal timing for vasopressin initiation in septic shock patients receiving norepinephrine with a primary outcome of in-hospital mortality.
The model recommended vasopressin earlier (median 4 vs. 5 hours post-shock), at lower norepinephrine doses (0.20 vs. 0.37 µg/kg/min), and in more patients (87% vs. 31%). Concordance with the model was associated with reduced in-hospital mortality (adjusted OR 0.81, 95% CI 0.73-0.91).
How does this change my practice?
This study is a fascinating preview of how AI could help us as clinicians make important decisions in the future. That said, at this point, this study falls into the category of interesting but not practice-changing for me until more high-quality data is obtained.
Source
Optimal Vasopressin Initiation in Septic Shock: The OVISS Reinforcement Learning Study. JAMA. 2025 May 20;333(19):1688-1698. doi: 10.1001/jama.2025.3046. Erratum in: JAMA. 2025 May 6;333(17):1549. doi: 10.1001/jama.2025.5041. PMID: 40098600
