The Causal Artificial Intelligence Clinician for early haemodynamic management of septic shock in ICU
- 1University of Applied Sciences and Arts of Southern Switzerland, Dalle Molle Institute for Artificial Intelligence, Lugano, Switzerland
Abstract
RationaleStandardizing fluid and vasopressor resuscitation in septic shock is challenging due to patient heterogeneity.
ObjectivesTo identify optimal intravenous fluid and vasopressor doses in first six hours of ICU admission in septic shock.
MethodsGraphical causal inference models grounded in expert clinical knowledge were applied to estimate heterogeneous treatment effects, limiting bias from spurious correlations. The model was trained on 1,706 MIMIC admissions and externally validated on 1,450 eICU admissions. The primary outcome was in-hospital survival; clinical improvement was the secondary outcome.
Measurements and Main ResultsThe cohort comprised 3,156 participants (age 65 years [IQR 53-75]; 42.8% female). Deviation from vasopressor recommendations was associated with failed clinical improvement (median OR 1.26, 95% CI 1.21-1.32), the most stable association across twelve sensitivity configurations, and with in-hospital mortality (median OR 1.16, 95% CI 1.10-1.24), which was null in five of twelve. Fluid deviations yielded 1.12 (1.08-1.18) and 1.02 (.96-1.07) respectively. These associations are confounded by indication by construction. External validation AUROCs were 0.72 (95% CI 0.67-0.76) for survival and 0.70 (95% CI 0.66-0.73) for improvement, comparable to predictive baselines. Treatment response varied with baseline physiology, fluids showing neutral or positive effects and high-dose norepinephrine worse ones.
ConclusionsAn expert-grounded causal framework produced recommendations that behaved as an informative policy in external validation, using one-third the variables of a conventional predictive baseline. Whether alignment with those recommendations improves outcomes requires prospective evaluation. Code and instructions to reproduce this analysis, including a dashboard to explore our results, are available at https://github.com/IDSIA/causal-ai-clinician.
Impact statementThis study shows that a causal model built from an explicit, clinician-derived graph can generate early fluid and vasopressor dose recommendations for septic shock while using approximately one-third of the variables required by a conventional predictive model. Because the assumptions behind each recommendation are stated in the causal graph rather than embedded in a black-box predictor, they can be inspected, contested and revised by clinicians. The dose recommendations themselves remain hypothesis-generating and require prospective evaluation before any clinical use.
Related
- NOTIFY-DISCUSS: Patient and clinician perspectives on opportunistic screening for coronary artery calcification using artificial intelligence
- Artificial intelligence in medicine: a clinician-oriented guide and operational framework to evaluate new AI devices
- Digital, Mobile and Artificial Intelligence‐Supported Interventions for Pressure Injury Prevention and Management: A Systematic Review of Clinical, Workflow and Implementation Outcomes
- On the variability of generative artificial intelligence methods in conceptual modeling: an experimental evaluation on combining causal maps
- How artificial intelligence could improve the diagnosis and management of chronic obstructive pulmonary disease: a perspective from GOLD
BibTeX
@misc{angelotti2026causal,
title = {{The Causal Artificial Intelligence Clinician for early haemodynamic management of septic shock in ICU}},
author = {G. Angelotti and L. Azzimonti and M. Cecconi and M. Zaffalon},
howpublished = {medRxiv},
year = {2026},
month = jul,
doi = {10.64898/2026.07.06.26357375},
url = {https://www.medrxiv.org/content/10.64898/2026.07.06.26357375v2},
note = {CC BY},
}
RIS
TY - UNPB
TI - The Causal Artificial Intelligence Clinician for early haemodynamic management of septic shock in ICU
AU - G. Angelotti
AU - L. Azzimonti
AU - M. Cecconi
AU - M. Zaffalon
PY - 2026
DA - 2026/07/06
PB - medRxiv
DO - 10.64898/2026.07.06.26357375
UR - https://www.medrxiv.org/content/10.64898/2026.07.06.26357375v2
AB - RationaleStandardizing fluid and vasopressor resuscitation in septic shock is challenging due to patient heterogeneity. ObjectivesTo identify optimal intravenous fluid and vasopressor doses in first six hours of ICU admission in septic shock. MethodsGraphical causal inference models grounded in expert clinical knowledge were applied to estimate heterogeneous treatment effects, limiting bias from spurious correlations. The model was trained on 1,706 MIMIC admissions and externally validated on 1,450 eICU admissions. The primary outcome was in-hospital survival; clinical improvement was the secondary outcome. Measurements and Main ResultsThe cohort comprised 3,156 participants (age 65 years [IQR 53-75]; 42.8% female). Deviation from vasopressor recommendations was associated with failed clinical improvement (median OR 1.26, 95% CI 1.21-1.32), the most stable association across twelve sensitivity configurations, and with in-hospital mortality (median OR 1.16, 95% CI 1.10-1.24), which was null in five of twelve. Fluid deviations yielded 1.12 (1.08-1.18) and 1.02 (.96-1.07) respectively. These associations are confounded by indication by construction. External validation AUROCs were 0.72 (95% CI 0.67-0.76) for survival and 0.70 (95% CI 0.66-0.73) for improvement, comparable to predictive baselines. Treatment response varied with baseline physiology, fluids showing neutral or positive effects and high-dose norepinephrine worse ones. ConclusionsAn expert-grounded causal framework produced recommendations that behaved as an informative policy in external validation, using one-third the variables of a conventional predictive baseline. Whether alignment with those recommendations improves outcomes requires prospective evaluation. Code and instructions to reproduce this analysis, including a dashboard to explore our results, are available at https://github.com/IDSIA/causal-ai-clinician. Impact statementThis study shows that a causal model built from an explicit, clinician-derived graph can generate early fluid and vasopressor dose recommendations for septic shock while using approximately one-third of the variables required by a conventional predictive model. Because the assumptions behind each recommendation are stated in the causal graph rather than embedded in a black-box predictor, they can be inspected, contested and revised by clinicians. The dose recommendations themselves remain hypothesis-generating and require prospective evaluation before any clinical use.
ER -