Latent trajectories of cerebral perfusion pressure and risk prediction models among patients with traumatic brain injury: based on an interpretable artificial neural network.

Abstract

Objective: This study aimed to characterize long-term cerebral perfusion pressure (CPP) trajectory in traumatic brain injury (TBI) patients and construct an interpretable prediction model to assess the risk of unfavorable CPP evolution patterns.

Methods: TBI patients with CPP records were identified from the Medical Information Mart for the Intensive Care (MIMIC)-IV 2.1, eICU Collaborative Research Database (eICU-CRD) 2.0 and HiRID dataset 1.1.1. The research process consisted of two stages. First, group-based trajectory modeling (GBTM) was used to identify different CPP trajectories. Second, different ANN algorithms were employed to predict the trajectories of CPP.

Results: A total of 331 eligible patients' records from MIMIC-IV 2.1 and eICU-CRD 2.0 were used for trajectory analysis and model development. Additionally, 310 patients' data from HiRID were used for external validation. The GBTM identified 5 CPP trajectory groups, group 1 and group 5 were merged into class 1 based on unfavorable in-hospital mortality. The best 6 predictors were invasive systolic blood pressure coefficient of variation (ISBPCV), venous blood chloride ion concentration, PaCO2, PT (Prothrombin Time), CPP coefficient of variation (CPPCV), and mean CPP. Compared with other algorithms, Scaled Conjugate Gradient (SCG) performed relatively better in identifying class 1.

Conclusion: This study identified 2 CPP trajectory groups associated with elevated risk and 3 with reduced risk. PaCO₂ might be a strong predictor for the unfavorable CPP class. The ANN model achieved the primary goal of risk stratification, which is conducive to early intervention and individualized treatment.