Machine learning to predict vasopressin responsiveness in patients with septic shock

Area of Science:

• Critical Care Medicine
• Pharmacology
• Health Informatics

Background:

• Septic shock management often requires vasopressors like norepinephrine and vasopressin.
• Predicting individual patient response to adjunctive vasopressin is crucial for optimizing treatment and improving outcomes.
• Vasopressin nonresponsiveness is common and linked to higher mortality rates.

Purpose of the Study:

• To develop and externally validate machine learning models for predicting vasopressin responsiveness in septic shock patients.
• To assess the performance of gradient-boosting machine (XGBoost) and elastic net (EN) models in predicting vasopressin response.

Main Methods:

• Retrospective analysis of two large septic shock cohorts (BJH and BIDMC).
• Development of two supervised machine learning models (XGBoost, EN) trained on BJH data.
• External validation of models using BIDMC data to predict vasopressin responsiveness.

Main Results:

• Vasopressin responsiveness rates were 45% (BJH) and 39% (BIDMC).
• Mortality was significantly lower in vasopressin responders compared to nonresponders in both cohorts.
• Both models demonstrated modest predictive discrimination in training and validation datasets (AUROC range: 0.59-0.68).

Conclusions:

• Vasopressin nonresponsiveness in septic shock is frequent and associated with increased mortality.
• Current machine learning models have limited performance in predicting vasopressin response, highlighting the complexity of septic shock.
• Further research is needed to develop effective clinical decision support tools for personalized vasopressor therapy.