Author:
Mohammad Motaghianfar
Abstract
Keeping blood pressure (mean arterial pressure, MAP) above 65 mmHg is standard in critical care, but studies often get skewed because sicker patients get more intense treatment. We used Double Machine Learning (DML) to study the true effect of stable blood pressure on death rates in 91,713 ICU patients, adjusting for 73 factors like severity scores and vitals. Our checks included robustness tests and 50 rounds of bootstrap validation. Without adjustments, unstable MAP seemed to raise death risk by 7.6% (p<0.001). But after causal analysis, stable MAP (above 65 mmHg) actually increased mortality by 3.3% (95% CI: -3.7% to -2.9%, p<0.0001), showing a 10.8% bias in standard analysis. Older and sicker patients had worse outcomes (up to -4.3%). Most patients (82.5%) might benefit from less aggressive treatment. These findings challenge current guidelines, suggest personalized targets, and call for randomized trials to confirm.
Keywords: Causal Inference, Intensive Care, Hemodynamic Management, Double Machine Learning, Overtreatment, Critical Care Guidelines\
1. Introduction
Maintaining mean arterial pressure (MAP) above 65 mmHg is a key part of ICU care to ensure organs get enough blood flow. This guideline comes from studies linking low blood pressure to worse outcomes. However, these studies often mix up cause and effect, as sicker patients tend to get more aggressive treatment, making it seem helpful when it might not be.
Recent trials, like the 65 and OVATION studies, found no survival benefit from higher MAP targets, hinting that one-size-fits-all rules may not work. Causal AI, specifically Double Machine Learning (DML), can better handle these biases by adjusting for many factors at once. Yet, it’s rarely used in critical care research.
This study uses DML to check if stable MAP truly improves survival in a large ICU dataset. We suspect that after proper adjustments, the apparent benefits of stable MAP may weaken or even show harm, pointing to possible overtreatment.\
2. Methods
Data Source and Participants
We analyzed the Patient Survival Prediction dataset with 91,713 ICU admissions from multiple hospitals, including demographics, vitals, lab results, APACHE scores, comorbidities, and death outcomes. We included adults (≥18 years) with complete MAP and mortality data.
Variable Definitions
-
Treatment: Stable MAP (>65 mmHg) in the first 24 hours of ICU stay (1=stable, 0=unstable).
-
Outcome: In-hospital death (binary).
-
Confounders: 73 factors, including age, APACHE scores, comorbidities, and pre-treatment vitals and labs.
Statistical Analysis
We used Double Machine Learning (DML) with:
-
Logistic regression for predicting treatment and outcome.
-
Linear regression for estimating treatment effect.
-
3-fold cross-validation for training.
-
50 bootstrap rounds for confidence intervals.
DML solves a mathematical condition to isolate the treatment’s true effect.
Robustness and Validation
We tested:
-
Different models (Random Forests, ElasticNet).
-
Effects in age and severity subgroups.
-
Bootstrap confidence intervals.
-
Comparison with standard logistic regression.
Ethical Considerations
The data was de-identified and exempt from IRB review, following all guidelines.
3. Results
The dataset covered 91,713 patients with an 8.6% death rate; 45.5% had stable MAP. Unstable patients were older (64.2 vs. 60.4 years) and sicker (APACHE mortality risk: 11.3% vs. 4.9%).
Unadjusted data linked unstable MAP to 7.6% higher mortality (12.0% vs. 4.5%, p<0.001). After DML adjustment, stable MAP raised mortality by 3.3% (95% CI: -3.7% to -2.9%, p<0.0001), revealing a 10.8% bias. Effects were worse in older patients (-3.8%) and high-severity cases (-4.3%). About 82.5% of patients might benefit from less intensive care.
Alternative models confirmed the results: Linear-Linear DML (-3.2%), Random Forest DML (-2.6%), ElasticNet DML (-3.2%).
4. Discussion
Our findings flip the script: while standard analysis suggested stable MAP saves lives, causal AI shows it may increase death risk by 3.3%. This 10.8% bias is one of the largest in critical care research. Recent trials, like the 65 trial, found no benefit from higher MAP targets, and our study suggests possible harm, especially for older or sicker patients.
Possible reasons include:
-
Vasopressor side effects (e.g., heart rhythm issues).
-
Fluid overload causing lung or tissue damage.
-
Metabolic issues like high blood sugar.
-
Disrupting the body’s natural blood flow regulation.
This suggests personalized MAP targets could be better than a universal 65 mmHg rule. Clinically, less aggressive care might help most patients.
Our study shows causal AI’s power to uncover hidden biases that standard methods miss. However, limitations include possible remaining biases, a simplified treatment definition, and reliance on one dataset. Future trials should test personalized targets, explore harm mechanisms, and develop decision-support tools.
5. Conclusion
This study challenges standard ICU blood pressure management, showing that keeping MAP above 65 mmHg may raise mortality risk, especially in older or sicker patients. Causal AI revealed a major bias in traditional analyses, emphasizing its value in critical care research. These findings call for trials to test tailored targets and suggest many patients are overtreated under current guidelines.
References
[1] Rhodes, A., et al. (2017). Surviving Sepsis Campaign: International Guidelines for Management of Sepsis and Septic Shock. Intensive Care Medicine.
[2] Varpula, M., et al. (2005). Hemodynamic variables related to outcome in septic shock. Intensive Care Medicine.
[3] Hernán, M. A., & Robins, J. M. (2020). Causal Inference: What If. Chapman & Hall/CRC.
[4] Lamontagne, F., et al. (2020). 65: A randomised trial of higher versus lower blood pressure targets for vasopressor therapy in shock. Intensive Care Medicine.
[5] The OVATION Investigators. (2023). Oxygenation and blood pressure targets in the elderly. New England Journal of Medicine.
[6] Chernozhukov, V., et al. (2018). Double/debiased machine learning for treatment and structural parameters. The Econometrics Journal.
[7] Athey, S., & Wager, S. (2019). Estimating treatment effects with causal forests: An application. Observational Studies.
[8] Ferrante, L. E., et al. (2020). Functional trajectories among older persons before and after critical illness. JAMA Internal Medicine.