This study is a prospective cohort study that included 304 patients with DoC who were treated in the Department of Rehabilitation Medicine at the First Affiliated Hospital of Nanchang University from January 2021 to October 2023.

Inclusion criteria: ① First diagnosis of DoC; ② UWS or MCS according to the “European Guidelines for Coma and Consciousness Disorder Diagnosis; ③ Duration of DoC ≥28 days.

Exclusion criteria: ① Co-existing severe brain atrophy, congenital neurological disorders, or drug poisoning; ② History of cognitive dysfunction (such as dementia or a history of mental illness).

In this study, we conducted rigorous screening and preprocessing of the data from patients with DoC who were treated at the Department of Rehabilitation Medicine, The First Affiliated Hospital of Nanchang University, between January 2021 and October 2023. The initial cohort included 304 patients, with cases exhibiting missing critical data being excluded. Missing critical data included essential assessment scales, laboratory results, or follow-up data. Ultimately, data from 239 patients with good completeness were included in the final analysis. The data set was divided into a training set (70%) and a testing set (30%) using stratified random sampling, ensuring that the proportion of samples in each category was consistent between the two sets to enhance the model’s generalizability.

This study employed a combination of the Boruta algorithm based on random forests and recursive feature elimination (RFE) for feature selection. The Boruta algorithm constructs random features as a control group and compares their importance with that of the original features, thereby identifying those features that are truly important for the target variable. Recursive feature elimination utilizes feature contribution analysis and model average AUC performance evaluation to quantify the contribution of features to the model output and assess the model’s classification performance, further validating the importance of the features. The combination of these two methods ensures the comprehensiveness and robustness of the feature selection process, while enhancing the predictive power of the feature subset.

This study constructed five machine learning models using Python’s scikit-learn library: Logistic Regression (LR), Support Vector Machine (SVM), XGBoost, Gradient Boosting Machine (GBM), and Multi-Layer Perceptron (MLP). The optimal parameters for each model were determined using Grid Search and 5-fold cross-validation (CV) to minimize overfitting and optimize model performance. Logistic Regression (LR), a classic statistical method, is simple yet performs exceptionally well in medical predictions, offering good interpretability. Support Vector Machine (SVM) constructs a linear decision boundary by mapping input vectors into a high-dimensional feature space, thereby enhancing generalization ability. Multi-Layer Perceptron (MLP), a feedforward neural network model, possesses powerful representation learning capabilities. XGBoost is one of the most powerful boosting algorithms, offering excellent predictive performance via gradient boosting frameworks and regularization techniques. Gradient Boosting Machine (GBM) sequentially integrates multiple weak learners to improve overall performance.

In addition to conventional discrimination metrics such as the area under the receiver operating characteristic curve (AUC), the Net Reclassification Improvement (NRI) was employed as a supplementary performance measure. NRI, originally proposed by Pencina MJ et al., evaluates whether a new prediction model improves individual risk stratification relative to a reference model by quantifying changes in risk category assignment for individuals with and without events (27). Unlike AUC, which summarizes overall discriminative ability, NRI specifically assesses the direction and magnitude of risk reclassification, providing complementary information on model performance.

In this study, the machine learning model demonstrating the best performance based on cross-validation was treated as the new model and was compared with LR, which served as the reference model. The NRI was calculated using predefined clinical risk thresholds to assess differences in patient risk reclassification between the two models.

To enhance clinical applicability and interpretability, SHAP analysis was employed to evaluate the contribution of each feature to the model’s predictions. SHAP, based on the Shapley values from game theory, quantifies each feature’s contribution to the model’s prediction, with theoretical guarantees of completeness, consistency, and local accuracy. By calculating each feature’s marginal contribution across all possible feature combinations, SHAp values reflect the positive or negative influence of features on predictions. In this study, SHAP values were computed for all samples, and features were ranked based on mean absolute SHAP values. SHAP dependence plots were used to explore nonlinear relationships between feature values and their effects. Additionally, SHAP force plots visually demonstrate how features collectively influence individual sample predictions, providing explanations for personalized prognostic predictions.

All data analysis was conducted using Python 3.8 and related libraries. Continuous variables were expressed as mean ± standard deviation or median (interquartile range), while categorical variables were presented as frequencies and percentages. Group comparisons were performed using independent sample t-tests, Mann–Whitney U tests, or chi-square tests, with the appropriate statistical method selected based on data distribution characteristics. For model evaluation, the area under the receiver operating characteristic curve (AUC) was used to assess the classification performance of each model. DCA was used to compare the clinical utility of each model, and calibration curves were employed to assess the consistency between predicted probabilities and actual outcomes. Statistical analyses were carried out using Python libraries, including scikit-learn (version 1.5.1) and statsmodels (version 0.14.4), with visualization generated by matplotlib (version 3.10.0). Data preprocessing relied on pandas (version 2.2.3) and numpy (version 1.26.4), model construction used xgboost (version 2.0.3), and model interpretation was facilitated using shap (version 0.45.1). All statistical tests were two-sided, and a p-value of < 0.05 was considered statistically significant.

A total of 304 patients with Disorders of Consciousness (DoC) who visited the Department of Rehabilitation Medicine at The First Affiliated Hospital of Nanchang University from January 2021 to October 2023 were initially included. After strict screening and excluding cases with missing data, 239 patients were ultimately included in the analysis. Among them, 97 patients had a good prognosis (GOS-E score 3–8), while 142 had a poor prognosis (GOS-E score 1–2). The mean age of the entire cohort was 53.7 ± 15.0 years, with 60.7% of the patients being male. As shown in Table 1, there were statistically significant differences between the good and poor prognosis groups in terms of demographic characteristics, consciousness status, and assessment indicators (p < 0.05). No substantial differences were observed between the training and validation sets. Table 2 reveals that there were no significant differences between the training and validation sets in terms of demographic characteristics, medical history, consciousness status, and physiological indicators, except for the presence of subdural effusion, which showed a statistically significant difference (p < 0.05).

In the feature selection process, we used the Boruta algorithm based on random forests and the Recursive Feature Elimination (RFE) algorithm to identify relevant features from an initial set of 37. The Boruta algorithm accurately identified 8 key features that played a significant role in predicting patient prognosis, including the CRS-R total score, age, length of hospitalization, GCS motor response, GCS total score, and FOUR total score. At the same time, Hb was labeled as “Tentative,” indicating that this feature may have some predictive value, but its statistical significance was somewhat weaker compared to the features that were clearly selected. The remaining features were deemed “Rejected,” suggesting that they contributed minimally to predicting patient prognosis (Figure 1).

To further validate the importance of these features, we employed the Recursive Feature Elimination (RFE) algorithm. Our analysis revealed that features such as age, CRS-R, Hb, and FOUR ranked highly in terms of feature contribution, highlighting their crucial role in shaping the model’s predictive capability and marking them as key factors in predicting patient outcomes. Additionally, the model’s average AUC value gradually stabilized as more features were added, indicating that the selected feature set sufficiently supported the model’s predictive performance. Further addition of features did not significantly improve the model’s performance (Figure 2).

Throughout the feature selection process, we integrated the conclusions from both methods and carefully selected the features that made significant contributions to predicting patient outcomes. These included the CRS-R score, length of hospitalization, FOUR score, and hemoglobin level, among others. These features not only demonstrated high importance within the model but also underwent rigorous validation by the Boruta algorithm, ensuring the scientific rigor and robustness of the feature selection process. Based on this procedure, we meticulously constructed a highly efficient and accurate feature set for subsequent model development. This not only significantly enhanced the model’s predictive capability but also greatly increased its value and reliability in practical applications.

This study utilized five machine learning models to predict the prognosis of patients with chronic consciousness disorders, specifically LR, MLP, SVM, XGBoost, and GBM. Figure 3 displays the ROC curves for each model in the training set, with the GBM model showing the best performance (AUC = 0.954, 95% CI: 0.924–0.977). The MLP model followed, with an AUC of 0.896 (95% CI, 0.842–0.940), followed by GBM (AUC = 0.946, 95% CI: 0.912–0.973), LR (AUC = 0.830, 95% CI: 0.761–0.891), and SVM (AUC = 0.859, 95% CI: 0.800–0.911).

On the test set (Figure 3), the performance of all models slightly decreased, yet they still maintained good predictive power. The GBM model continued to perform the best (AUC = 0.922, 95% CI: 0.847–0.979), followed by XGBoost (AUC = 0.907, 95% CI: 0.817–0.976), MLP (AUC = 0.903, 95% CI: 0.814–0.962), SVM (AUC = 0.898, 95% CI: 0.812–0.963), and LR (AUC = 0.895, 95% CI: 0.806–0.963).

Compared with the LR, the GBM demonstrated a significantly higher threshold-based net reclassification improvement (Thr-NRI = 0.403, 95% CI: 0.069–0.714, p = 0.019), indicating improved patient risk reclassification at predefined clinical risk thresholds.

This study evaluated the classification performance of each model using confusion matrices on both the training and test sets. On the test set, the GBM model exhibited the highest specificity (0.977), accurately identifying patients with poor prognosis. However, its sensitivity was relatively low (0.621), resulting in the misclassification of some patients with good prognosis. The XGBoost model had a specificity of 0.930 and sensitivity of 0.690, showing a more balanced performance. The SVM and MLP models had sensitivities of 0.690 and 0.62, respectively, but their specificities were lower. The Logistic Regression model showed a sensitivity of 0.654 and specificity of 0.883, with overall lower accuracy (Figure 4).

On the training set, all models demonstrated high accuracy and low error rates. XGBoost exhibited the best sensitivity (0.750), while the GBM model excelled in specificity (0.909; Figure 5).

The model calibration performance evaluation (Figure 6) revealed that the GBM model had the smallest expected calibration error on the training set. On the test set, all models showed similar calibration performance, with expected calibration errors ranging from 0.076 to 0.167.

Decision curve analysis (Figure 7) further confirmed that, across most clinically relevant decision threshold ranges, all machine learning models offered higher net benefits compared to the “treat-all” or “treat-none” strategies, with the GBM model performing best across most threshold intervals.

Precision-recall curve analysis (Figure 8) indicated that the GBM model achieved an average precision (AP) of 0.87 on the test set, significantly higher than the baseline level (0.40). On the training set, the average precision (AP) was 0.94, substantially higher than the baseline level (0.41), confirming that the model maintained excellent classification performance across different thresholds.

Based on feature importance analysis and the SHAP value explanation framework, this study systematically identified the key factors influencing prognosis prediction in patients with chronic consciousness disorders. Through the visualization of SHAP summary plots, the impact of core features, including the CRS-R total score, age, GCS total score, and FOUR total score, was quantitatively analyzed. The SHAP feature importance plot, arranged in descending order of the absolute mean SHAP values, clearly demonstrated that these four indicators dominate the prediction model, with their cumulative contribution to the explained variance significantly higher than that of other variables. This confirmed the central role of these features in prognosis prediction.

Further analysis revealed that higher scores on the CRS-R, FOUR, and GCS scales were significantly positively correlated with favorable prognosis, highlighting the critical role of neurological function assessments in identifying recovery potential. In contrast, the age variable showed a significant negative correlation, suggesting that older age independently exerts a negative effect on prognosis. Additionally, while the weights of hospital stay duration, hemoglobin levels, and D-dimer concentrations were relatively lower, fluctuations in their SHAP values still significantly influenced the prediction outcomes, suggesting that these clinical parameters play an essential, though auxiliary, role in prognosis evaluation.

Case-specific analysis using SHAP force plots provided individualized insights into the model’s predictions. For example, in a typical case, features such as a FOUR score of 14.0, GCS total score of 10.0, CRS-R total score of 8.0, and a hospital stay of 33.0 days positively shifted the SHAP values, significantly increasing the likelihood of predicting a favorable prognosis. In contrast, indicators such as D-dimer concentration (1.18), hemoglobin level (97.0 g/L), age (63.0 years), and serum albumin level (32.0 g/L) contributed negative SHAP values, thereby decreasing the predicted probability of a favorable outcome. This combined approach of group feature analysis and individual case validation strengthens the statistical robustness of the model and enhances its clinical interpretability and practical utility (Figure 9).

Based on the above findings, the GBM model was selected as the optimal model. It not only demonstrated excellent predictive performance but also provides intuitive decision support for clinicians, enabling them to better understand and apply the model’s predictions, thereby optimizing treatment plans and resource allocation for patients with prolonged disorders of consciousness pDOC.

This study is limited by its single-center design, which may introduce selection bias and reduce generalizability. Although multiple routine clinical predictors were included, unmeasured factors may still influence prognosis, and the lack of neurophysiological and advanced neuroimaging data may constrain accuracy and interpretability. The model also shows high specificity but moderate sensitivity for Good vs. Poor GOS-E, potentially underestimating recovery in some patients; thus, predictions should be used cautiously with clinically appropriate thresholds. Multicenter, larger-sample validation is warranted.