A Locally Optimized Data-Driven Tool to Predict Sepsis-Associated Vasopressor Use in the ICU

Area of Science:

• Critical Care Medicine
• Machine Learning in Healthcare
• Transfer Learning Applications

Background:

• Sepsis is a life-threatening condition requiring timely intervention, often involving vasopressors.
• Predicting vasopressor need in intensive care unit (ICU) patients with sepsis is crucial for optimizing treatment.
• External validation of predictive models across diverse healthcare systems remains a challenge.

Purpose of the Study:

• To develop a predictive model for vasopressor use in ICU patients with sepsis.
• To optimize the model's performance for external validation across different hospital systems using domain adaptation.
• To leverage transfer learning for improved generalizability of sepsis prediction models.

Main Methods:

• An observational cohort study involving 14,512 ICU patients diagnosed with severe sepsis.
• A neural network Weibull-Cox survival model was trained on 40 features from electronic medical records at a development site.
• Domain adaptation techniques were applied to fine-tune the model for a separate validation site (different healthcare system).

Main Results:

• The initial model predicted vasopressor use 4-24 hours in advance with an AUC of 0.80-0.81 at the development site.
• Domain adaptation significantly improved model performance at the validation site: AUC increased from 0.77 to 0.81 (p < 0.01).
• Specificity and positive predictive value also saw significant improvements after domain adaptation at the validation site.

Conclusions:

• Domain adaptation effectively enhances the performance of predictive models for sepsis-associated vasopressor use during external validation.
• Transfer learning approaches, like domain adaptation, are valuable for generalizing predictive models across different healthcare settings.
• The study demonstrates a successful strategy for improving the external validity of ICU predictive models.