AI Helps Optimize Therapy Selection and Dosing for Septic Shock

Patients experience dangerously low blood pressure that can rapidly lead to organ failure, requiring urgent treatment with fluids and blood-pressure-raising drugs called vasopressors. While international guidelines outline general treatment sequences, they do not account for how quickly septic shock evolves or how differently patients respond to therapy. Now, new findings show that artificial intelligence (AI) can determine the optimal timing and dosing of vasopressors on an individual basis, improving survival outcomes.

In research led by Johns Hopkins University (Baltimore, MD, USA) in collaboration with academic medical centers, including the University of California, San Francisco (San Francisco, CA, USA), investigators applied reinforcement learning, a form of machine learning in which algorithms learn optimal actions by analyzing outcomes from large datasets. Using electronic medical records, the model continuously assessed patient-specific variables such as blood pressure, organ dysfunction scores, and concurrent medications.

Based on these evolving inputs, the system learned when to initiate vasopressin, a potent vasopressor typically added after norepinephrine if blood pressure remains dangerously low. Reinforcement learning differs from traditional clinical trial approaches by evaluating thousands of treatment scenarios simultaneously rather than testing a single predefined rule. This allowed the researchers to model septic shock as a dynamic condition rather than a static diagnosis.

By learning from real-world clinical data rather than rigid protocols, the system was designed to individualize treatment decisions in a condition known for wide variability across patients, hospitals, and countries. The model was trained using electronic health records from more than 3,500 patients and then validated on unseen data from nearly 11,000 additional patients. Results showed that patients whose care matched the algorithm’s recommendations had significantly lower in-hospital mortality.

Importantly, the model often recommended starting vasopressin earlier than clinicians typically did, but outcomes worsened when the drug was started even earlier than the algorithm suggested. This highlighted a narrow, patient-specific therapeutic window. The findings, published in The Journal of the American Medical Association, demonstrate that timing, not just drug choice, is critical in septic shock management.

An individualized approach to vasopressor initiation could reduce deaths without increasing harmful side effects from overly aggressive treatment. Researchers plan to implement the model in real clinical settings before expanding nationally through a clinical AI platform. The same reinforcement-learning framework could later be applied to other complex treatment decisions in critical care.

"There's no one-size-fits-all rule—in septic shock, there is substantial variability in resuscitation practices between hospitals and in different countries, especially regarding vasopressor support,” said senior author Romain Pirracchio. “Given the diversity of the population included in this study, the results show that an individualized vasopressin initiation rule can improve the outcome of patients with septic shock."

Related Links:
Johns Hopkins University
University of California, San Francisco