# Construction and validation of a prognostic model for clinical outcomes in patients with prolonged disorders of consciousness based on multidimensional indicators: a prospective cohort study

## TL;DR

This study builds a machine learning model to predict outcomes for patients with prolonged disorders of consciousness, using clinical data to improve decision-making.

## Contribution

A novel gradient boosting machine model using multidimensional clinical indicators for accurate prognosis prediction in prolonged disorders of consciousness.

## Key findings

- The GBM model achieved high predictive accuracy with an AUROC of 0.954 in training and 0.922 in testing.
- CRS-R score, age, FOUR score, GCS score, and hospitalization length were key predictors identified via SHAP analysis.
- The model showed substantial net clinical benefit across most threshold probabilities according to decision curve analysis.

## Abstract

To develop and validate a prognostic prediction model for patients with prolonged disorders of consciousness (pDoC) based on multidimensional clinical indicators, aiming to improve prognostic accuracy and provide objective support for clinical decision-making.

A prospective cohort study was conducted involving 304 patients with pDoC admitted to the First Affiliated Hospital of Nanchang University between January 2021 and October 2023. Clinical data were collected, including demographics, etiology, disease duration, behavioral assessment scores (Glasgow Coma Scale [GCS], Full Outline of UnResponsiveness [FOUR], and Coma Recovery Scale-Revised [CRS-R]), laboratory indicators, and other relevant clinical variables. Prognosis was assessed using the Glasgow Outcome Scale-Extended (GOS-E) and dichotomized into good outcome (GOS-E 3–8) and poor outcome (GOS-E 1–2). Feature selection was performed using the Boruta algorithm combined with recursive feature elimination (RFE). Prognostic models were developed using logistic regression, support vector machine (SVM), multilayer perceptron (MLP), XGBoost, and gradient boosting machine (GBM) implemented in scikit-learn. Model performance was evaluated using accuracy, receiver operating characteristic curve area under the curve (ROC-AUC), and decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) were applied to interpret the optimal model.

Among the developed models, the GBM model demonstrated the best predictive performance, with an AUROC of 0.954 (95% CI: 0.924–0.977) in the training set and 0.922 (95% CI: 0.847–0.979) in the test set. Decision curve analysis indicated that the GBM model yielded substantial net clinical benefit across most threshold probabilities. SHAP analysis identified CRS-R score, age, FOUR score, total GCS score, and length of hospitalization as the most influential prognostic predictors.

A robust prognostic prediction model for pDoC patients was developed and validated using multidimensional clinical data and machine learning techniques. The GBM model achieved excellent discriminative performance and clinical utility, providing an objective tool for prognosis estimation and individualized treatment and rehabilitation planning. Further multi-center studies are warranted to optimize the model and confirm its generalizability.