Machine Learning Models to Predict Withdrawal of Life-Sustaining Therapy in Patients With Severe Traumatic Brain...

Area of Science:

• Neurology
• Medical Informatics
• Trauma Surgery

Background:

• Withdrawal of life-sustaining therapy (WLST) accounts for over half of traumatic brain injury (TBI) deaths.
• TBI mortality has improved, yet WLST rates remain unchanged, suggesting potential prognostic misconceptions.
• Decision-making for WLST is complex, influenced by factors beyond clinical prognosis, leading to inter-center variability.

Purpose of the Study:

• To develop a machine learning (ML) model for predicting WLST decisions in severe TBI patients.
• To identify key determinants influencing WLST decisions, hypothesizing facility WLST rate as a significant factor.

Main Methods:

• Observational study using the American College of Surgeons Trauma Quality Improvement Project National Trauma Databank (2017-2021).
• Included severe TBI patients (Abbreviated Injury Scale-Head ≥1, Glasgow Coma Scale <9).
• Developed ML models to predict WLST, optimizing for area under the receiver operating curve (AUROC) and assessing determinant impact using Shapley values.

Main Results:

• Out of 155,639 severe TBI patients, 20.8% underwent WLST.
• ML models achieved high predictive accuracy, with AUROC improving from 0.875 at admission to 0.896 by total length-of-stay.
• Key predictors for WLST included age, highest emergency department GCS, and facility WLST rate.

Conclusions:

• ML models reliably predict WLST decisions in severe TBI.
• Institutional withdrawal culture, indicated by facility WLST rate, is a strong independent predictor.
• Refining prognostic tools is crucial to mitigate biases and prevent premature WLST decisions influenced by institutional practices.