JMIR Publications

ABSTRACT

Background:

Traumatic brain injury is a critically ill disease with a high mortality rate, and clinical treatment is committed to continuously optimizing treatment strategies to improve survival rates.

Objective:

This article aims to establish a reinforcement learning (RL) algorithms to optimize the survival prognosis decision-making scheme for traumatic brain injury (TBI) patients in the intensive care unit (ICU).

Methods:

We included a total of 2745 patients from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database and randomly divided them into a training set and an internal validation set at 8:2; We extracted 34 features for analysis and modeling using a 2-hour time compensation, 2 action features [mean arterial pressure (MAP) and temperature], 1 outcome feature (survival status at 28 days). We used a RL algorithm called Weighted Dueling Double Deep Q-Network with embedded human Expertise (WD3QNE) to maximize cumulative returns and evaluated the model using a doubly robust off-policy evaluation method. Finally, we collected 2463 TBI patients from MIMIC-III as an external validation set to test the model.

Results:

The action features are divided into 6 intervals, and the expected benefits are estimated using doubly robust off-policy evaluation method; The results indicate that the survival rate of artificial intelligence (AI) strategies is higher than that of clinical doctors (88.016% vs 80.422%), with an expected return of (28.978 vs 27.092). Compared to clinical doctors, AI algorithms select normal temperatures more frequently (36.56-36.83℃) and recommend MAP levels of 87.5-95.0mmHg. In external validation, the AI strategy still has a high survival rate of 87.565%, with an expected return of 27.517.

Conclusions:

We established and validated a model using RL to guide the management of important vital signs in patients with TBI in ICU. This model can increase the 28 day survival rate of patients. Clinical Trial: The MIMIC database is an open public database where patient information has been processed and does not require ethical review and trial registration. Our authors Mengyuan Diao and Hongwei Zhang have been assessed and certified for reasonable use of the database.