Explainable and Externally Validated Resting-State fMRI Machine Learning Reveals Network Mechanisms Supporting Preserved Consciousness: A Cross-Sectional Study Comparing Minimally Conscious State and Unresponsive Wakefulness Syndrome

Explainable and Externally Validated Resting-State fMRI Machine Learning Reveals Network Mechanisms Supporting Preserved Consciousness: A Cross-Sectional Study Comparing Minimally Conscious State and Unresponsive Wakefulness Syndrome | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Explainable and Externally Validated Resting-State fMRI Machine Learning Reveals Network Mechanisms Supporting Preserved Consciousness: A Cross-Sectional Study Comparing Minimally Conscious State and Unresponsive Wakefulness Syndrome Linghui Dong, Hui Li, Kaiyue Han, Zhiqing Tang, Xingxing Liao, and 4 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8959330/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 7 You are reading this latest preprint version Abstract Background Accurate differentiation between minimally conscious state (MCS) and unresponsive wakefulness syndrome (UWS) is a significant clinical challenge because behavioral assessments are often constrained by patients' motor impairments and fluctuating arousal. Beyond diagnostic classification, clarifying the specific brain network mechanisms that sustain residual consciousness in MCS remains a clinical priority for developing targeted therapeutic interventions. Methods This multicenter study acquired rs-fMRI data from 100 patients across two independent clinical centers. Six participants were excluded due to suboptimal data quality related to excessive head motion, resulting in a final analysis of 94 patients (discovery cohort: n = 48; external validation cohort: n = 46). A diagnostic framework was developed using regional and network markers, including the amplitude of low-frequency fluctuations, regional homogeneity, degree centrality, and functional connectivity. Nine machine learning classifiers were optimized, with the best-performing model tested on the external validation cohort. SHAP analysis quantified circuit contributions to elucidate neurobiological mechanisms. Results L1-regularized logistic regression selected 10 core features, dominated by default mode network (DMN) interactions with cerebellar and subcortical nodes and salience-related local synchrony. The support vector machine emerged as the leading model, achieving an AUC of 0.859 (accuracy: 82.6%; sensitivity: 90.5%) in the external validation cohort. SHAP attribution identified a core neurobiological signature dominated by DMN–cerebellar coupling (specifically left angular gyrus to right cerebellum), alongside DMN–subcortical (thalamus and putamen) pathways and insular synchrony. Conclusions An externally validated rs-fMRI framework differentiated MCS from UWS and localized residual consciousness in MCS to preserved DMN-centered corticocerebellar and cortico–subcortical circuits. prolonged disorders of consciousness minimally conscious state unresponsive wakefulness syndrome resting-state fMRI default mode network machine learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 1. Background After severe brain injury, a meaningful number of survivors enter a prolonged disorder of consciousness (pDoC), creating a sustained medical and societal burden that extends far beyond the acute ICU phase. Even in expert care, diagnostic separation between minimally conscious state (MCS) and unresponsive wakefulness syndrome (UWS) is vulnerable to fluctuations in arousal and to motor impairments that prevent patients from expressing awareness. A meta-analysis of diagnostic procedures reported substantial misclassification between these conditions, emphasizing the limits of behavior alone for identifying residual consciousness [ 1 ]. At the same time, converging evidence from multimodal paradigms has shown that some unresponsive patients can demonstrate preserved higher-order cognitive functions and brain network activities consistent with cognitive motor dissociation, reinforcing the need to detect retained mechanisms of awareness and underlying brain activity states, rather than relying only on outward expression [ 2 ]. These realities make pDoC a medical problem of both diagnosis and mechanism: early and accurate recognition of MCS can shape treatment intensity and rehabilitation planning, while clarifying which brain networks remain functionally preserved in MCS may reveal why its prognosis is typically better than UWS and may point toward targets for therapeutic escalation. Clinical classification rests on careful behavioral examination. MCS is defined by reproducible but inconsistent evidence of awareness, whereas UWS denotes wakefulness without behavioral signs of awareness [ 3 , 4 ]. The Coma Recovery Scale–Revised (CRS-R) is widely used because it samples multiple response domains and improves diagnostic reliability when administered systematically [ 5 ]. Yet behavior remains an indirect readout. A patient may understand commands but fail to produce movement, or may show only intermittent signs because arousal is unstable. This mismatch between cognition and action has driven the development of “command-following independent” assessments. Task-based functional MRI (fMRI) studies demonstrated that some patients diagnosed as vegetative could volitionally modulate brain activity during mental imagery, which provided a proof-of-concept that awareness can exist without outward response [ 6 ]. More recent large multicenter work extends this principle with combined fMRI and EEG paradigms, linking covert cognition to clinically meaningful outcomes [ 2 ]. Together, these findings reinforce a clinical priority: when behavior is limited, physiology must be used to reduce the risk of missed awareness. Resting-state fMRI (rs-fMRI) offers a practical approach in this setting because it does not require active participation. Spontaneous blood oxygenation level–dependent (BOLD) fluctuations reveal intrinsic functional organization, including long-range interactions among distributed systems that support cognition and state regulation. Contemporary models emphasize that consciousness depends on coordinated activity across large-scale networks and thalamo-cortical circuits rather than isolated regional activation. The mesocircuit hypothesis, for example, links impaired cortico-striato-thalamo-cortical dynamics to reduced capacity to sustain globally integrated processing after severe injury [ 7 ]. Network-oriented perspectives also highlight interactions among the default mode network, salience network, and frontoparietal control systems as core components of higher-order brain function, with relevance to both normal cognition and clinical dysfunction [ 8 ]. In practice, rs-fMRI can be summarized using complementary markers that capture different aspects of functional integrity. The amplitude of low-frequency fluctuations (ALFF) reflects regional signal power and baseline activity patterns [ 9 ]. Regional homogeneity (ReHo) indexes local synchronization among neighboring voxels [ 10 ]. Graph-based measures such as degree centrality (DC) quantify how strongly a region participates as a functional hub within whole-brain coupling [ 11 ]. Together with pairwise functional connectivity (FC), these measures provide a multi-scale description of preserved and disrupted network function that is well aligned with mechanistic questions in pDoC. Despite strong rationale and accumulating evidence, two translation barriers remain. First, many rs-fMRI studies report group differences but do not deliver clinically reliable, patient-level tools that generalize across hospitals, scanners, and etiologies. Overfitting and site effects are well-recognized challenges in medical neuroimaging, and models that perform well in a single cohort may degrade when applied externally. Beyond classification accuracy, the clinical utility of neuroimaging in pDoC hinges on interpretability. For clinicians and families facing end-of-life or long-term rehabilitation decisions, a 'black-box' model offers little guidance. Understanding the specific functional circuits—rather than just the diagnostic label—is essential for identifying therapeutic targets and predicting neuroplastic potential. This interpretability gap also limits the ability to translate predictive features into neurophysiological hypotheses about why MCS patients maintain consciousness and recover better than UWS patients. Explainable machine learning (EML) has therefore become increasingly relevant in medicine, not as an engineering add-on, but as a clinical requirement for transparency and trust [ 12 ]. In particular, Shapley-value–based approaches provide a principled way to quantify how each feature contributes to a model’s output in an individual patient and across a cohort [ 13 ]. Guided by these needs, this study integrates multicenter rs-fMRI with clinically grounded interpretation to address two aims. First, we develop and externally validate an imaging-based diagnostic model to identify MCS among pDoC patients using a harmonized acquisition protocol across two independent clinical centers. We derive rs-fMRI markers that quantify regional activity and whole-brain network organization across a validated functional atlas, enabling patient-level classification that is designed to be robust across sites [ 14 ]. This diagnostic focus is clinically motivated by early recognition of covert or fluctuating awareness, which can influence rehabilitation engagement, communication attempts, and therapeutic intensity. Second, and central to the medical interpretation of pDoC, we quantify the contribution of specific networks and brain regions to MCS classification using an explainability framework. This allows us to move from “which label fits” to “which preserved circuit functions matter,” with the goal of clarifying neural mechanisms that help patients remain in MCS rather than decline to UWS. By aligning predictive modeling with interpretable neurobiology, the study aims to support earlier diagnosis while also generating clinically legible hypotheses about network-level substrates of consciousness maintenance after severe brain injury. 2. Methods 2.1 Study design and Participants This multicenter, cross-sectional study was approved by the Ethics Committee of the China Rehabilitation Research Center (No. 2023-012-01), with ethical oversight from the Affiliated Hospital of Qingdao University, and conducted in accordance with the Declaration of Helsinki. A total of 100 patients with pDoC were enrolled from two independent clinical centers. The discovery cohort included 50 patients (25 MCS and 25 UWS) recruited from the China Rehabilitation Research Center between June 2023 and June 2025. The external validation cohort consisted of 50 patients (25 MCS and 25 UWS) from the Affiliated Hospital of Qingdao University. Inclusion criteria: (1) a diagnosis of UWS or MCS based on behavioral assessments with the CRS-R; (2) a disease duration ≥ 28 days with a stable clinical condition; (3) the first onset of a prolonged disorder of consciousness; (4) age between 18 and 75 years ; and (5) written informed consent obtained from the family or legal representative. Exclusion criteria: (1) severe medical comorbidities, such as heart failure, renal failure, acute lung injury, or acute pulmonary infection; (2) a history of neurological or psychiatric diseases, including previous stroke, traumatic brain injury, or depression; (3) a history of alcohol or drug abuse; (4) intracranial metal foreign bodies or other contraindications to MRI; (5) neuromodulation therapy within the past three months; (6) persistent involuntary shaking or movements that precluded successful fMRI data acquisition. 2.2 Clinical Assessment and Outcome Definition Consciousness levels were rigorously evaluated using the CRS-R, which comprises six subscales: auditory, visual, motor, oromotor/verbal, communication, and arousal. To account for potential fluctuations in arousal and minimize diagnostic error, two trained physicians jointly performed three separate assessments within a single day. These evaluations were conducted during the morning, afternoon, and evening. The final behavioral diagnosis and consciousness level were determined by the highest total score achieved across the three assessment sessions. 2.3 MRI Data Acquisition To ensure the consistency and comparability of neuroimaging data across the multicenter study, all participants from both the discovery and external validation cohorts were scanned using identical 3.0T scanner models (Philips Ingenia, Philips Healthcare, The Netherlands) and strictly synchronized imaging protocols. Resting-state BOLD parameters at both centers were set as follows: repetition time = 2000 ms, echo time = 30 ms, flip angle = 90°, field of view = 224 mm, slice thickness = 3.5 mm, inter-slice gap = 0.85 mm, 32 slices, and a matrix size of 64 × 64. For each participant at each time point, three consecutive resting-state runs were acquired, with each run lasting 8 minutes. Standardized precautions were implemented across both sites to minimize data artifacts, including the use of secure head restraints to reduce motion and earplugs or headphones to attenuate acoustic noise. 2.4 Data Preprocessing Preprocessing was conducted in MATLAB 2023a using the RestPlus toolbox. Data were converted from DICOM to NIfTI format and the first 10 volumes were discarded. Images then underwent slice-timing correction and rigid-body realignment for head-motion correction. Participants were excluded if their head motion exceeded 3 mm of displacement or 3° of rotation in any direction. Structural T1 images were segmented using DARTEL and functional images were normalized to MNI space with resampling to 3×3×3 mm³. Linear detrending and temporal band-pass filtering (0.01–0.08 Hz) were applied. Nuisance regression included Friston-24 motion parameters and signals from white matter and cerebrospinal fluid. To reduce spatial blurring across neighboring regions and inflation of inter-regional correlations, spatial smoothing was applied only for ALFF computation (Gaussian kernel, 6-mm FWHM). In contrast, DC, FC, and ReHo were derived from unsmoothed data. 2.5 Computation of Brain Network Metrics We utilized a refined version of the Power-264 functional atlas to define the network nodes. The atlas has since been widely adopted and validated, supporting its reproducibility across studies [ 14 ]. The refined atlas identified 236 key nodes distributed across 13 functional networks, spanning multiple systems such as the Cingulo-opercular (CON), Default Mode (DMN), Frontoparietal (FPCN), Subcortical (SCN), Cerebellar (CN), and Dorsal Attention (DAN) networks. To enhance robustness against residual registration errors and local noise, ROI-level measures were extracted from a 27-voxel neighborhood centered on each atlas coordinate. This neighborhood comprised the central coordinate voxel and its 26 adjacent voxels. All network metrics were computed using the RestPlus toolbox. ALFF: Calculated on smoothed data by transforming the BOLD time series to the frequency domain and averaging the square root of the power spectrum (0.01–0.08 Hz); ReHo: Computed on unsmoothed data using Kendall’s coefficient of concordance to capture local synchronization between a voxel and its 26 functional neighbors; DC: Quantified on unsmoothed data as the sum of suprathreshold Pearson correlations (r > 0.25) between a voxel and the rest of the brain; FC: Pairwise Pearson correlations between representative ROI time series were computed and Fisher z-transformed for statistical normality. 2.6 Dimensionality Reduction and Predictive Modeling In the discovery cohort, candidate neuroimaging features were first identified by group-level comparisons between the MCS and UWS groups. Independent-samples t-tests were performed with age, sex, disease duration, etiology, and mean head motion parameters entered as covariates. Multiple comparisons were controlled using the false discovery rate procedure. This initial screening reduced the feature space from several tens of thousands of measures to 20 variables. To further limit overfitting, sparse feature selection was then conducted using L1-regularized logistic regression (Logistic LASSO). The regularization parameter was tuned by stratified five-fold cross-validation within the discovery set using a prespecified grid of inverse penalty values (C). Model selection was based on the mean area under the receiver operating characteristic curve across folds. To prevent information leakage, Z-score normalization was performed within the cross-validation framework, with scaling parameters estimated from the training split of each fold and applied to the corresponding held-out split. After selecting the optimal parameter, the model was refit on the full discovery set, and variables with non-zero coefficients were retained as the final feature set for subsequent analyses. The retained features were then used to construct nine distinct machine learning classifiers: Artificial Neural Network, Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree, LightGBM, Gradient Boosting, Extra Trees, Random Forest, and XGBoost. During the training phase, hyperparameter optimization for each model was conducted via grid search with five-fold cross-validation on the discovery set. The Area Under the Receiver Operating Characteristic Curve (AUC) was utilized as the primary performance indicator to evaluate and compare the diagnostic capability of the candidate models. 2.7 Model Validation and Explainability The final models were applied to the independent external validation cohort from Qingdao University to assess generalizability. Performance was evaluated using the AUC and secondary metrics, including accuracy, sensitivity, specificity, F1-score, and Cohen’s kappa. To elucidate the neurobiological mechanisms distinguishing MCS from UWS, we applied SHAP (SHapley Additive exPlanations) to the best-performing model to quantify the contribution of individual network features to the diagnostic prediction. 2.8 Statistical Analysis Statistical analyses were performed using Python 3.12. Continuous variables are expressed as mean ± standard deviation (SD) and compared between MCS and UWS groups within each center using independent-samples t-tests. Categorical variables (e.g., sex) were compared using Chi-square tests. All tests were two-sided, and p < 0.05 was considered statistically significant. 2.9 Use of generative AI tools As non-native English speakers, we used Google Gemini to improve language clarity and grammar during manuscript preparation. Gemini was also used to assist in drafting the graphical abstract and preparing the study flowchart. All figures presenting results from data analyses are original, unaltered outputs generated by the analysis pipeline. All content generated or revised with AI assistance was reviewed and verified by the authors, who take full responsibility for the final manuscript. 3. Results 3.1 Demographic and Clinical Characteristics In the discovery cohort, 50 participants were recruited. Two were excluded from the final analysis due to suboptimal fMRI data quality, leaving 48 patients, including 24 with MCS and 24 with UWS. In the external validation cohort, 50 participants were recruited; four were excluded due to excessive head motion, resulting in a final sample of 46 patients (21 with MCS and 25 with UWS). Statistical analysis revealed no significant differences between the MCS and UWS groups regarding age, sex, disease duration, or etiology in either cohort (P>0.05, Tables 1 and 2 ). As expected, baseline CRS-R scores were significantly higher in the MCS group than in the UWS group for both the discovery and validation cohorts (P<0.05, Tables 1 and 2 ). The schematic representation of the integrated multicenter study design and the machine learning-based diagnostic pipeline is illustrated in Fig. 1 . Table 1 Demographic and clinical characteristics of the discovery cohort. Discovery cohort MCS (n = 24) UWS (n = 24) Statistic P value Age, years 52.33 ± 10.21 54.62 ± 7.57 t = -0.88 0.382 Male, n (%) 18 (75.0%) 16 (66.7%) χ² = 0.1 0.751 Traumatic brain injury, n (%) 10 (41.7%) 12 (50.0%) χ² = 0.08 0.772 Time since injury, days 121.83 ± 67.23 98.92 ± 41.67 t = -1.419 0.163 CRS-R total score 11.96 ± 2.05 4.29 ± 1.73 t = 13.98 < 0.001 Notes: Continuous variables are presented as mean ± standard deviation. Categorical variables are expressed as counts and percentages. P-values were calculated using independent-samples t-tests or Chi-square tests. Abbreviations: MCS, minimally conscious state; UWS, unresponsive wakefulness syndrome; CRS-R, Coma Recovery Scale–Revised Table 2 Demographic and clinical characteristics of the validation cohort. Validation Cohort MCS (n = 21) UWS (n = 25) Statistic P value Age, years 51.76 ± 10.03 53.08 ± 8.49 t = -0.48 0.637 Male, n (%) 16 (76.2%) 17 (68.0%) χ² = 0.08 0.775 Traumatic brain injury, n (%) 9 (42.9%) 14 (56.0%) χ² = 0.35 0.554 Time since injury, days 116.24 ± 67.8 112.48 ± 58.18 t = 0.2 0.841 CRS-R total score 12.14 ± 1.98 4.28 ± 1.7 t = 14.5 < 0.001 Notes: Continuous variables are presented as mean ± standard deviation. Categorical variables are expressed as counts and percentages. P-values were calculated using independent-samples t-tests or Chi-square tests. Participants were recruited into a discovery cohort (Center 1) and an independent validation cohort (Center 2) (each n = 50; 25 MCS and 25 UWS). Diagnosis was established using repeated CRS-R as-sessments, followed by 3.0-T rs-fMRI acquired with identical protocols. After standardized pre-processing and motion-based quality control (total excluded n = 6), brain network features were ex-tracted using the Power-264 atlas (ALFF, ReHo, DC, and FC). In the discovery cohort, features were screened (FDR-corrected) and further selected using L1-regularized logistic regression with five-fold cross-validation, and nine classifiers were trained and optimized. The leading model was then applied to the external cohort for performance evaluation and interpreted using SHAP. Ab-breviations: MCS, minimally conscious state; UWS, unresponsive wakefulness syndrome; CRS-R, Coma Recovery Scale–Revised; rs-fMRI, resting-state functional magnetic resonance imaging; ALFF, amplitude of low-frequency fluctuations; ReHo, regional homogeneity; DC, degree centrality; FC, functional connectivity; FDR, false discovery rate; SHAP, SHapley Additive exPlanations. 3.2 Differences in Resting-state fMRI 3.2.1 ALFF Analysis Patients in the MCS group exhibited significantly higher ALFF than those in the UWS group across multiple functional networks. These increases were identified in default mode network nodes, including DMN.86 (left angular gyrus; t = 5.273; P<0.05; Fig. 2 ), DMN.117 (left middle temporal gyrus; t = 6.812; P<0.05; Fig. 2 ), and DMN.130 (right angular gyrus; t = 5.835). Higher ALFF was also detected in the subcortical node SCN.223 (left thalamus; t = 6.132; P<0.05; Fig. 2 ) and the cerebellar node CN.245 (right cerebellum; t = 5.722; P<0.05; Fig. 2 ). Regional signal power was significantly increased in the MCS group (red spheres) within default mode network (DMN) nodes, specifically the left angular gyrus (DMN.86), left middle temporal gyrus (DMN.117), and right angular gyrus (DMN.130). Additional increases were localized to the left thalamus (SCN.223) and right cerebellum (CN.245). Color bar represents t-values (FDR corrected, P < 0.05). Abbreviations: MCS, minimally conscious state; UWS, unresponsive wakefulness syndrome; DMN, default mode network; SCN, subcortical network; CN, cerebellar network; FDR, false discovery rate. 3.2.2 ReHo Analysis Regarding ReHo, the MCS group showed significantly higher local synchronization in DMN.86 (left angular gyrus; t = 6.452; P<0.05; Fig. 3 ) and DMN.92 (right posterior cingulate gyrus; t = 5.573; P<0.05; Fig. 3 ). Additional increases in ReHo were observed in FPCN.176 (left inferior frontal gyrus, opercular part; t = 5.879; P<0.05; Fig. 3 ) and SN.208 (left insula; t = 6.204; P<0.05; Fig. 3 ). The MCS group demonstrated significantly higher local synchronization (red spheres) in the left angular gyrus (DMN.86), right posterior cingulate gyrus (DMN.92), left inferior frontal gyrus (FPCN.176), and left insula (SN.208). Color bar represents t-values (FDR corrected, P < 0.05). Abbreviations: FPCN, frontoparietal control network; SN, salience network. 3.2.3 DC Analysis DC was significantly elevated in the MCS group compared with the UWS group. This effect involved several key nodes: DMN.78 (left superior frontal gyrus, orbital part; t = 5.586; P<0.05; Fig. 4 ), MRN.136 (right precuneus; t = 6.139; P<0.05; Fig. 4 ), SCN.223 (left thalamus; t = 6.081; P<0.05; Fig. 4 ), and DAN.251 (right precuneus; t = 5.882; P<0.05; Fig. 4 ). Key functional hubs identified in MCS included the left superior frontal gyrus (DMN.78), right precuneus (MRN.136 and DAN.251), and left thalamus (SCN.223). Red spheres indicate regions with significant global connectivity increases compared to UWS (FDR corrected, P < 0.05). Abbreviations: MRN, memory retrieval network; DAN, dorsal attention network. 3.2.4 FC Analysis Between-group comparisons revealed significantly higher FC in patients with MCS compared to those with UWS. The most extensive increases were observed in DMN related connections, including DMN.108–VAN.239 (right medial superior frontal gyrus–right middle temporal gyrus, t = 6.601; P<0.05; Fig. 5 ), DMN.117–SCN.231 (left middle temporal gyrus–right putamen, t = 6.121; P<0.05; Fig. 5 ), DMN.124–MRN.136 (left parahippocampal gyrus–right precuneus, t = 5.892; P<0.05; Fig. 5 ), DMN.79–FPCN.201 (left middle temporal gyrus–left inferior frontal gyrus, t = 5.481; P<0.05; Fig. 5 ), and DMN.90–SCN.223 (left precuneus–left thalamus, t = 5.284; P<0.05; Fig. 5 ). Furthermore, increased FC was identified in cerebellar-related pathways, including CN.245–DMN.86 (right cerebellum–left angular gyrus, t = 6.308; P<0.05; Fig. 5 ) and CN.245–SCN.223 (right cerebellum–left thalamus, t = 5.572; P<0.05; Fig. 5 ). Finally, enhanced functional coupling was observed for FPCN.186–VAN.239 (right precentral gyrus–right middle temporal gyrus, t = 5.933; P<0.05; Fig. 5 ). Connectome analysis revealed enhanced functional coupling in MCS (red lines), predominantly involving DMN-subcortical, DMN-cerebellar, and DMN-frontoparietal circuits. Blue spheres represent network nodes defined by the Power-264 atlas (FDR corrected, P < 0.05). Abbreviations: VAN, ventral attention network 3.3 Feature Selection by Logistic L1 Regularization Initial statistical screening across the discovery cohort identified 21 neuroimaging features with significant group differences. These features were entered into a L1-regularized logistic regression model for secondary screening. The inverse regularization strength (C) was tuned using stratified five-fold cross-validation, with the mean cross-validated AUC as the optimization criterion. The optimal value was C = 0.412, yielding a mean cross-validated AUC of 0.818. At this parameter setting, 10 features had non-zero coefficients and were therefore retained in the final feature set (Fig. 6 A). As C increased beyond the optimum, additional features entered the model, but the cross-validated AUC showed no further improvement. This procedure yielded 10 core features for diagnostic modeling: ALFF-DMN.117, DMN.86-CN.245 FC, DMN.117-SCN.231 FC, ALFF-DMN.130, DMN.90-SCN.223 FC, SCN.223-CN.245 FC, DMN.108-VAN.239 FC, DMN.124-MRN.136 FC, FPCN.186-VAN.239 FC, and ReHo-SN.208 (Fig. 6 B). (A) Mean cross-validated AUC (left y-axis) and the number of retained features with non-zero coefficients (right y-axis) are shown as a function of the inverse regularization parameter C in the discovery cohort using stratified five-fold cross-validation. The red dashed line marks the optimal value (C = 0.412), which yielded a mean cross-validated AUC of 0.818 and retained 10 features. (B) Coefficients of the 10 selected neuroimaging features from the Logistic L1 model used for downstream diagnostic modeling. 3.4 Diagnostic Performance and External Validation Nine machine learning classifiers were evaluated for their ability to distinguish MCS from UWS. While several models achieved near-perfect performance in the training set, the models were primarily assessed based on their generalizability to the external validation cohort. The SVM was identified as the most robust model, achieving an AUC of 0.859 in the external validation set (Fig. 7 A). The SVM model demonstrated an accuracy of 0.826, a sensitivity of 0.905, and a specificity of 0.760 (Fig. 7 C). Other classifiers, such as XGBoost and KNN, also yielded strong performance with AUCs of 0.850 (Fig. 7 A). Detailed performance metrics for all classifiers are provided in the Fig. 7 . (A-B) ROC curves for nine classifiers in the discovery and external validation cohorts. SVM achieved the highest AUC of 0.859 in the external set. (C-D) Comparison of accuracy, sensitivity, specificity, and other performance metrics across models. Abbreviations: ROC, receiver operating characteristic; AUC, area under the curve; SVM, support vector machine. 3.5 Model Interpretability by SHAP SHAP values were calculated for the optimized SVM model to identify the key neurobiological signatures of MCS. The analysis indicated that FC between DMN.86 and CN.245 was the most influential feature for the diagnostic prediction (Fig. 8 A). Other critical contributors included DMN.117-SCN.231 FC, ReHo-SN.208, ALFF-DMN.130, and SCN.223-CN.245 FC (Fig. 8 A). Patients with higher values in these network preservation were significantly more likely to be classified as MCS rather than UWS (Fig. 8 B). (A) Global feature importance ranked by the mean absolute SHAP value, identifying DMN.86–CN.245 FC as the primary predictor. (B) SHAP summary plot showing that higher values of network preservation (pink) increase the probability of an MCS diagnosis. 4. Discussion In this study, we show that patients with MCS exhibit a clear preservation of large-scale network function compared with UWS, with the most robust signal concentrated in the DMN, CN, and key cortico–subcortical circuits. Consistent classification performance across model training and external testing indicates that these preserved interactions provide clinically informative features for differentiating MCS from UWS. Importantly, the SHAP-based attribution analysis did not point to a diffuse connectome effect, but instead converged on a small set of biologically coherent contributors, with the cross-network coupling DMN.86–CN.245 emerging as the most characteristic preserved feature in MCS. This pattern supports a clinically grounded interpretation that consciousness maintenance in MCS relies on sustained communication between cortical integrative systems and subcortical enabling pathways, rather than on isolated regional activity. A major finding was that the most informative features were not diffuse across the connectome, but converged on hubs within cortico–subcortical circuits and the DMN. This is clinically meaningful because behavioral assessment alone is vulnerable to under-detection of residual awareness when motor output is impaired, arousal fluctuates, or bedside conditions are suboptimal, all of which are common in pDoC [ 15 ]. In that context, a model that extracts stable diagnostic signal from resting-state networks addresses a key translational gap: shifting the diagnostic emphasis from observed behavior to preserved integrative physiology. Meta-analytic evidence supports that fMRI-derived markers show moderate diagnostic value for differentiating MCS from UWS/VS, reinforcing the plausibility of our approach and the need for clinically practical, interpretable implementations [ 16 ]. The interpretability analysis further suggested that preservation within posterior associative cortices and their subcortical coupling is central to differentiating MCS from UWS. Nodes with strong positive contributions included the left angular gyrus and right middle temporal gyrus, regions that participate in the temporoparietal junction and lateral temporal components of the DMN and semantic-associative processing streams. In mechanistic terms, these regions are positioned to support multimodal integration and internal model updating, functions that are repeatedly implicated in conscious access and the maintenance of a coherent internal milieu. Modern accounts increasingly view consciousness as depending on large-scale integration rather than isolated regional activity, with the DMN acting as a key convergence system that is especially vulnerable across pharmacological and pathological perturbations of consciousness [ 17 ]. Importantly, in pDoC, reduced DMN integrity and disrupted interaction between DMN and subcortical arousal systems are among the most reproducible network abnormalities, and are closely tied to diagnostic category and recovery potential [ 18 ]. Our results extend this by localizing clinically useful diagnostic information to specific DMN-associated cortical nodes rather than treating the DMN as a monolithic entity. Subcortical nodes with high contribution, particularly the left thalamus and right putamen, support a complementary interpretation rooted in thalamo-striato-cortical gating. The thalamus is not merely a relay, but a central regulator of cortical effective connectivity and state transitions. Circuit-level work demonstrates layer-specific thalamic control over cortex that tracks changes in consciousness level, providing a mechanistic substrate for why thalamic integrity can be disproportionately informative in consciousness disorders [ 19 ]. In clinical pDoC cohorts, abnormalities in cortico–striato–pallido–thalamo–cortical loops have been repeatedly observed and are associated with level of consciousness, aligning with mesocircuit models that emphasize thalamic underactivation and impaired cortico-subcortical facilitation [ 20 ]. The putamen’s contribution in our model may reflect preserved basal ganglia participation in action selection, salience-weighting, and cortical enabling conditions, which can influence the probability that residual cognition manifests as reproducible, command-related behavior. This interpretation is consistent with broader pDoC connectivity syntheses highlighting subcortical–cortical disruptions as a core neurophysiologic mechanism rather than an epiphenomenon [ 21 ]. Our findings challenge the traditional cortico-centric view of consciousness by highlighting the pivotal role of DMN–cerebellar coupling in sustaining MCS. Cerebellar findings are often under-discussed in pDoC, yet the cerebellum is increasingly recognized as embedded within multiple intrinsic networks, including DMN, SN, and executive systems, through cerebello-thalamo-cortical loops [ 22 ]. In our data, cerebellum involvement may reflect preservation of timing, error monitoring, and autonomic-affective integration that indirectly stabilizes large-scale cortical dynamics necessary for conscious processing. While the cerebellum is unlikely to “generate” conscious content in isolation, it can modulate the consistency and coordination of distributed processing, thereby supporting the network conditions under which consciousness is sustained. This view aligns with contemporary connectivity frameworks that emphasize distributed control and the vulnerability of network coordination in pDoC [ 23 ]. Comparison with existing literature highlights both consistency and novelty. Our emphasis on DMN–subcortical circuitry matches a large body of work identifying DMN disruption and thalamocortical decoupling as key signatures of impaired consciousness, and it aligns with clinical calls to incorporate neuroimaging into routine pDoC assessment to reduce diagnostic error and identify covert awareness [ 15 ]. At the same time, our study contributes by pairing diagnostic modeling with node-level interpretability, enabling a clinically legible statement: MCS is characterized less by “global connectivity” and more by selective preservation of integrative hubs and their subcortical enabling circuits. This is compatible with recent approaches using interpretable machine learning in multicentre pDoC datasets, which also found meaningful interactions between modality and anatomical structures and emphasized that interpretability can bridge the gap between prediction and mechanism [ 24 ]. Moreover, deep learning approaches have demonstrated strong discrimination between MCS and UWS using rs-fMRI, supporting the feasibility of automated detection of residual awareness; however, many such models are criticized for limited transparency, which can impede clinical adoption [ 25 ]. Our work addresses this translational barrier by keeping mechanistic interpretability central rather than ancillary. The clinical significance of this work lies in two domains. First, earlier and more accurate identification of MCS has immediate downstream consequences for treatment intensity, rehabilitation planning, ethical decision-making, and family counseling. Contemporary guidance emphasizes that functional neuroimaging can identify covert awareness in a substantial minority of behaviorally unresponsive patients, and that implementation barriers are increasingly logistical rather than conceptual [ 15 ]. Second, by identifying which brain regions and networks carry the strongest diagnostic signal, the study advances a mechanistic hypothesis with therapeutic implications: interventions that strengthen DMN hubs (PCC/precuneus and lateral parietotemporal nodes) and thalamo-striato-cortical coupling, potentially including cerebello-thalamic modulation, may be rational targets for future therapies. Such targeting can also inform trial stratification, because network preservation may predict responsiveness to neuromodulatory approaches aimed at restoring large-scale integration [ 24 ]. Several limitations should be considered when interpreting these findings. The model was trained on resting-state connectivity, which is sensitive to acquisition differences, head motion, physiological noise, and variations in vigilance. Although external testing reduces the risk of overfitting to a single-site distribution, multicentre harmonization and prospective evaluation remain essential before clinical deployment. Additionally, diagnostic labels in pDoC are imperfect because bedside behavior can fluctuate and can be confounded by motor impairment; thus, any supervised model trained on behavioral diagnosis inherits some label noise. This issue is widely acknowledged in the field and is a central argument for integrating imaging with repeated standardized assessment rather than replacing it [ 26 ]. Finally, while interpretability helps link predictive features to plausible neurobiology, attribution does not prove causality. The identified nodes should therefore be treated as candidate biomarkers of preserved network function rather than definitive mechanistic drivers. Future studies combining rs-fMRI with task-based paradigms, electrophysiology, and longitudinal outcomes will be important to separate markers of current state from predictors of recovery [ 27 ]. 6. Conclusions In summary, this study supports the clinical utility of connectome-based modeling for differentiating MCS from UWS and provides anatomically grounded, interpretable evidence that preservation of DMN-associated parietotemporal hubs, thalamic and striatal circuitry, and cerebello-thalamo-cortical contributions are key network features of MCS. Future work should prioritize prospective multicentre validation, longitudinal outcome prediction, and multimodal integration to determine whether strengthening these preserved networks can be translated into targeted therapies and improved patient trajectories. Abbreviations ALFF Amplitude of low-frequency fluctuations AUC Area under the curve CN Cerebellar network CON Cingulo-opercular network CRS-R Coma Recovery Scale–Revised DAN Dorsal attention network DC Degree centrality DMN Default mode network FC Functional connectivity FDR False discovery rate FPCN Frontoparietal control network LASSO Least absolute shrinkage and selection operator MCS Minimally conscious state pDoC Prolonged disorders of consciousness ReHo Regional homogeneity RF Random forest ROC Receiver operating characteristic ROI Region of interest rs-fMRI Resting-state functional MRI SCN Subcortical network SD Standard deviation SHAP SHapley additive explanations SN Salience network UWS Unresponsive wakefulness syndrome Declarations Ethics approval and consent to participate: The study was conducted in accordance with the Declaration of Helsinki, and approved by the Ethics Committees of the China Rehabilitation Research Center and the Affiliated Hospital of Qingdao University (protocol code 2023–012-01). Written informed consent was obtained from each participant or, when the participant lacked decision-making capacity, from a legally authorized representative. Consent for publication: Not applicable Availability of data and materials: The data that support the findings of this study are not publicly available due to ethical and privacy restrictions. Deidentified data may be made available from the corresponding author upon reasonable request and with approval from the relevant ethics committee. Competing interests: The authors declare that they have no competing interests Funding: This study was self-funded and supported by personal funds from Zhanghao (grant number: 2022HZ-06-01). The article processing charge (APC) was also covered by Zhanghao’s personal funds (grant number: 2022HZ-06-01). The funder (Zhanghao, the corresponding author) was involved in the study as an author; no additional roles beyond the authors’ contributions were imposed by any external funding body. The funder had no independent role, separate from the authorship role, in the study design, data collection, data analysis, interpretation, the decision to publish, or the preparation of the manuscript. Authors' contributions: HZ: Conceptualization, project administration. LD, HL, KH: Methodology. HL: Formal analysis. HL, ZT, JL, XL: Investigation. LD, HZ: Writing—original draft. KH, XL, TL, HZ: Writing—review and editing. XL, HZ: Supervision. All authors read and approved the final manuscript. Acknowledgements: We gratefully acknowledge all assessors at the China Rehabilitation Research Center and the Affiliated Hospital of Qingdao University who participated in the clinical assess-ments of the enrolled patients. References Bender A, Jox RJ, Grill E, Straube A, Lulé D. Persistent vegetative state and minimally conscious state: a sys-tematic review and meta-analysis of diagnostic procedures. Dtsch Arztebl Int. 2015;112(14):235–42. 10.3238/arztebl.2015.0235 . Bodien YG, Allanson J, Cardone P, et al. Cognitive Motor Dissociation in Disorders of Consciousness. N Engl J Med. 2024;391(7):598–608. 10.1056/NEJMoa2400645 . Giacino JT, Ashwal S, Childs N, et al. The minimally conscious state: definition and diagnostic criteria. Neurology. 2002;58(3):349–53. 10.1212/wnl.58.3.349 . Laureys S, Celesia GG, Cohadon F, et al. Unresponsive wakefulness syndrome: a new name for the vegetative state or apallic syndrome. BMC Med. 2010;8:68. 10.1186/1741-7015-8-68 . Published 2010 Nov 1. Giacino JT, Kalmar K, Whyte J. The JFK Coma Recovery Scale-Revised: measurement characteristics and diag-nostic utility. Arch Phys Med Rehabil. 2004;85(12):2020–9. 10.1016/j.apmr.2004.02.033 . Owen AM, Coleman MR, Boly M, Davis MH, Laureys S, Pickard JD. Detecting awareness in the vegetative state. Science. 2006;313(5792):1402. 10.1126/science.1130197 . Schiff ND. Recovery of consciousness after brain injury: a mesocircuit hypothesis. Trends Neurosci. 2010;33(1):1–9. 10.1016/j.tins.2009.11.002 . Menon V. Large-scale brain networks and psychopathology: a unifying triple network model. Trends Cogn Sci. 2011;15(10):483–506. 10.1016/j.tics.2011.08.003 . Zang YF, He Y, Zhu CZ, et al. Altered baseline brain activity in children with ADHD revealed by resting-state functional MRI. Brain Dev. 2007;29(2):83–91. 10.1016/j.braindev.2006.07.002 . Zang Y, Jiang T, Lu Y, He Y, Tian L. Regional homogeneity approach to fMRI data analysis. NeuroImage. 2004;22(1):394–400. 10.1016/j.neuroimage.2003.12.030 . Zuo XN, Ehmke R, Mennes M, et al. Network centrality in the human functional connectome. Cereb Cortex. 2012;22(8):1862–75. 10.1093/cercor/bhr269 . Pahud de Mortanges A, Luo H, Shu SZ et al. Orchestrating explainable artificial intelligence for multimodal and longitudinal data in medical imaging. NPJ Digit Med. 2024;7(1):195. Published 2024 Jul 22. 10.1038/s41746-024-01190-w Lundberg SM, Erion G, Chen H, et al. From Local Explanations to Global Understanding with Explainable AI for Trees. Nat Mach Intell. 2020;2(1):56–67. 10.1038/s42256-019-0138-9 . Power JD, Cohen AL, Nelson SM, et al. Functional network organization of the human brain. Neuron. 2011;72(4):665–78. 10.1016/j.neuron.2011.09.006 . Kazazian K, Monti MM, Owen AM. Functional neuroimaging in disorders of consciousness: towards clinical implementation. Brain. 2025;148(7):2283–98. 10.1093/brain/awaf075 . Zheng H, Tian L, Cai J. Meta-analysis of the diagnostic value of functional magnetic resonance imaging for dis-tinguishing unresponsive wakefulness syndrome/vegetative state and minimally conscious state. Front Neurosci. 2024;18:1395639. Published 2024 Sep 9. 10.3389/fnins.2024.1395639 Heine L, Soddu A, Gómez F et al. Resting state networks and consciousness: alterations of multiple resting state network connectivity in physiological, pharmacological, and pathological consciousness States. Front Psychol. 2012;3:295. Published 2012 Aug 27. 10.3389/fpsyg.2012.00295 Vanhaudenhuyse A, Noirhomme Q, Tshibanda LJ, et al. Default network connectivity reflects the level of con-sciousness in non-communicative brain-damaged patients. Brain. 2010;133(Pt 1):161–71. 10.1093/brain/awp313 . Redinbaugh MJ, Phillips JM, Kambi NA, et al. Thalamus Modulates Consciousness via Layer-Specific Control of Cortex. Neuron. 2020;106(1):66–e7512. 10.1016/j.neuron.2020.01.005 . Li H, Dong L, Liu J, Zhang X, Zhang H. Abnormal characteristics in disorders of consciousness: A resting-state functional magnetic resonance imaging study. Brain Res. 2025;1850:149401. 10.1016/j.brainres.2024.149401 . Plosnić G, Raguž M, Deletis V, Chudy D. Dysfunctional connectivity as a neurophysiologic mechanism of dis-orders of consciousness: a systematic review. Front Neurosci. 2023;17:1166187. 10.3389/fnins.2023.1166187 . Published 2023 Jul 19. Habas C. Functional Connectivity of the Cognitive Cerebellum. Front Syst Neurosci. 2021;15:642225. 10.3389/fnsys.2021.642225 . Published 2021 Apr 8. Naro A, Bramanti A, Leo A, et al. Shedding new light on disorders of consciousness diagnosis: The dynamic functional connectivity. Cortex. 2018;103:316–28. 10.1016/j.cortex.2018.03.029 . Manasova D, Belloli LML, Rosenfelder MJ, et al. Multimodal multicentre investigation of diagnostic and prog-nostic markers in disorders of consciousness. Brain Published online January. 2026;7. 10.1093/brain/awaf412 . Yang H, Wu H, Kong L, et al. Precise detection of awareness in disorders of consciousness using deep learning framework. NeuroImage. 2024;290:120580. 10.1016/j.neuroimage.2024.120580 . Schnakers C, Vanhaudenhuyse A, Giacino J, et al. Diagnostic accuracy of the vegetative and minimally conscious state: clinical consensus versus standardized neurobehavioral assessment. BMC Neurol. 2009;9:35. 10.1186/1471-2377-9-35 . Published 2009 Jul 21. Wu S, Zhu B, Ye Z et al. Functional connectivity in whole-brain and network analysis differentiates minimally conscious from unresponsive patients: a resting-state fNIRS study. J Transl Med. 2025;23(1):1093. Published 2025 Oct 14. 10.1186/s12967-025-07181-z Additional Declarations No competing interests reported. Supplementary Files GraphicalAbstract.tif Cite Share Download PDF Status: Under Review Version 1 posted Reviewers agreed at journal 30 Mar, 2026 Reviewers agreed at journal 26 Mar, 2026 Reviewers invited by journal 26 Mar, 2026 Editor invited by journal 02 Mar, 2026 Editor assigned by journal 27 Feb, 2026 Submission checks completed at journal 27 Feb, 2026 First submitted to journal 24 Feb, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8959330","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":612709751,"identity":"450e07f9-7233-4665-a586-94d02c817411","order_by":0,"name":"Linghui Dong","email":"","orcid":"","institution":"Qilu Hospital of Shandong University","correspondingAuthor":false,"prefix":"","firstName":"Linghui","middleName":"","lastName":"Dong","suffix":""},{"id":612709752,"identity":"70632b87-fa76-4939-aa20-c9a9dcc91349","order_by":1,"name":"Hui Li","email":"","orcid":"","institution":"Capital Medical University Xuanwu Hospital","correspondingAuthor":false,"prefix":"","firstName":"Hui","middleName":"","lastName":"Li","suffix":""},{"id":612709753,"identity":"61c47890-a965-4446-ac29-7750b9316fd0","order_by":2,"name":"Kaiyue Han","email":"","orcid":"","institution":"China Rehabilitation Research Center","correspondingAuthor":false,"prefix":"","firstName":"Kaiyue","middleName":"","lastName":"Han","suffix":""},{"id":612709754,"identity":"1f11b06e-b1fa-454b-92a0-69cb9d09a17c","order_by":3,"name":"Zhiqing Tang","email":"","orcid":"","institution":"China Rehabilitation Research Center","correspondingAuthor":false,"prefix":"","firstName":"Zhiqing","middleName":"","lastName":"Tang","suffix":""},{"id":612709755,"identity":"99487180-5327-4896-8495-880f3f2c15f3","order_by":4,"name":"Xingxing Liao","email":"","orcid":"","institution":"China Rehabilitation Research Center","correspondingAuthor":false,"prefix":"","firstName":"Xingxing","middleName":"","lastName":"Liao","suffix":""},{"id":612709756,"identity":"0c8a3a51-5f25-45f6-9832-f1f2e6e27c35","order_by":5,"name":"Junzi Long","email":"","orcid":"","institution":"China Rehabilitation Research Center","correspondingAuthor":false,"prefix":"","firstName":"Junzi","middleName":"","lastName":"Long","suffix":""},{"id":612709757,"identity":"617dc2c5-7838-40de-b9fe-4d1eb42fe8c4","order_by":6,"name":"Xinyue Liu","email":"","orcid":"","institution":"Qilu Hospital of Shandong University","correspondingAuthor":false,"prefix":"","firstName":"Xinyue","middleName":"","lastName":"Liu","suffix":""},{"id":612709758,"identity":"1b933ff1-ab79-4e9f-8b00-dad0a4ffca1b","order_by":7,"name":"Tao Lin","email":"","orcid":"","institution":"China Rehabilitation Research Center","correspondingAuthor":false,"prefix":"","firstName":"Tao","middleName":"","lastName":"Lin","suffix":""},{"id":612709759,"identity":"864aa86f-3166-4952-8bdb-3924d3909f03","order_by":8,"name":"Hao Zhang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAyklEQVRIiWNgGAWjYLACxgYbHn5m5gMHPvwgXkuanGR7W+LBmT3EazlsbHDmjPFhDjYiVPPPPvvw488dzIkNN3I+HGbgYZDnFzuAX4vEuXRjCckzbImNM3I3HC6wYDCcOTsBvxYDHjY2BsM2nsRmCaCWGTwMCQa3idGS2CYBRDkPDvOwEavlYJuBMQ/PGQbitEicYWOWbGxLkJNgbzMABrIEYb/w97AxfvzZ9p/H/jDz4w8fftjI80sT0IJhK2nKR8EoGAWjYBRgBwDt80Im6VRRwgAAAABJRU5ErkJggg==","orcid":"","institution":"Qilu Hospital of Shandong University","correspondingAuthor":true,"prefix":"","firstName":"Hao","middleName":"","lastName":"Zhang","suffix":""}],"badges":[],"createdAt":"2026-02-24 15:54:41","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8959330/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8959330/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":105843898,"identity":"9792c029-9efb-4a35-8a3b-4613152e4352","added_by":"auto","created_at":"2026-03-31 17:28:48","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":738238,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eStudy workflow and explainable machine-learning pipeline.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8959330/v1/89659adc6cf3f96a81106ed1.png"},{"id":105904852,"identity":"b37e1839-e6c5-49dd-a393-89f8746783fb","added_by":"auto","created_at":"2026-04-01 10:10:49","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":247873,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eBrain regions exhibiting significantly higher ALFF in MCS compared to UWS.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8959330/v1/b705aa5353d398c91a7ab8bf.png"},{"id":105843900,"identity":"49b839b0-d2db-4690-8cad-03d1c085ef97","added_by":"auto","created_at":"2026-03-31 17:28:48","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":246792,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eRegional homogeneity (ReHo) differences between MCS and UWS groups.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-8959330/v1/177915398839e2c89e9592e8.png"},{"id":105843901,"identity":"a2d4249b-7856-41c0-a74d-79289ead19b8","added_by":"auto","created_at":"2026-03-31 17:28:48","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":249783,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDistribution of nodes with significantly higher degree centrality (DC) in MCS.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-8959330/v1/225343be5f3ed6c7bcf24327.png"},{"id":105843903,"identity":"4d4745f8-0e32-404c-8e22-01de3b442c5a","added_by":"auto","created_at":"2026-03-31 17:28:48","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":278018,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDifferential functional connectivity (FC) patterns between MCS and UWS.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-8959330/v1/18423cd34bd3eebf28a8e10d.png"},{"id":105904696,"identity":"818613d7-766e-4aec-8e88-f3173c86111e","added_by":"auto","created_at":"2026-04-01 10:10:17","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":73632,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eLogistic L1 feature selection and cross-validated performance.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e(A) Mean cross-validated AUC (left y-axis) and the number of retained features with non-zero coefficients (right y-axis) are shown as a function of the inverse regularization parameter C in the discovery cohort using stratified five-fold cross-validation. The red dashed line marks the optimal value (C = 0.412), which yielded a mean cross-validated AUC of 0.818 and retained 10 features. (B) Coefficients of the 10 selected neuroimaging features from the Logistic L1 model used for downstream diagnostic modeling.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-8959330/v1/dbb145abe54ec9fba421b4df.png"},{"id":105904293,"identity":"c7ca6534-b922-40b7-ac8c-cc2188b8962c","added_by":"auto","created_at":"2026-04-01 10:07:12","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":147994,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDiagnostic performance and external validation of machine learning models.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e(A-B) ROC curves for nine classifiers in the discovery and external validation cohorts. SVM achieved the highest AUC of 0.859 in the external set. (C-D) Comparison of accuracy, sensitivity, specificity, and other performance metrics across models. Abbreviations: ROC, receiver operating characteristic; AUC, area under the curve; SVM, support vector machine.\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-8959330/v1/eecdc89bf68c17b19a8c7de7.png"},{"id":105843906,"identity":"318f84c3-a351-4df0-95a4-03d99ec80644","added_by":"auto","created_at":"2026-03-31 17:28:48","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":109798,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eNeurobiological interpretation of the SVM model using SHAP analysis.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e(A) Global feature importance ranked by the mean absolute SHAP value, identifying DMN.86–CN.245 FC as the primary predictor. (B) SHAP summary plot showing that higher values of network preservation (pink) increase the probability of an MCS diagnosis.\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-8959330/v1/32baba8d499a54cba6357aa7.png"},{"id":105906532,"identity":"524c689b-9eef-44b2-9b80-fcbe8df85cf5","added_by":"auto","created_at":"2026-04-01 10:22:51","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3130147,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8959330/v1/43e46213-2605-4776-92e4-86346e4e54b9.pdf"},{"id":105904674,"identity":"c8c25522-8bdc-4f4e-b453-f452efb3b293","added_by":"auto","created_at":"2026-04-01 10:10:12","extension":"tif","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":1223670,"visible":true,"origin":"","legend":"","description":"","filename":"GraphicalAbstract.tif","url":"https://assets-eu.researchsquare.com/files/rs-8959330/v1/14ed26c3c2cd009e25a3190d.tif"}],"financialInterests":"No competing interests reported.","formattedTitle":"Explainable and Externally Validated Resting-State fMRI Machine Learning Reveals Network Mechanisms Supporting Preserved Consciousness: A Cross-Sectional Study Comparing Minimally Conscious State and Unresponsive Wakefulness Syndrome","fulltext":[{"header":"1. Background","content":"\u003cp\u003eAfter severe brain injury, a meaningful number of survivors enter a prolonged disorder of consciousness (pDoC), creating a sustained medical and societal burden that extends far beyond the acute ICU phase. Even in expert care, diagnostic separation between minimally conscious state (MCS) and unresponsive wakefulness syndrome (UWS) is vulnerable to fluctuations in arousal and to motor impairments that prevent patients from expressing awareness. A meta-analysis of diagnostic procedures reported substantial misclassification between these conditions, emphasizing the limits of behavior alone for identifying residual consciousness [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. At the same time, converging evidence from multimodal paradigms has shown that some unresponsive patients can demonstrate preserved higher-order cognitive functions and brain network activities consistent with cognitive motor dissociation, reinforcing the need to detect retained mechanisms of awareness and underlying brain activity states, rather than relying only on outward expression [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. These realities make pDoC a medical problem of both diagnosis and mechanism: early and accurate recognition of MCS can shape treatment intensity and rehabilitation planning, while clarifying which brain networks remain functionally preserved in MCS may reveal why its prognosis is typically better than UWS and may point toward targets for therapeutic escalation.\u003c/p\u003e \u003cp\u003eClinical classification rests on careful behavioral examination. MCS is defined by reproducible but inconsistent evidence of awareness, whereas UWS denotes wakefulness without behavioral signs of awareness [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. The Coma Recovery Scale\u0026ndash;Revised (CRS-R) is widely used because it samples multiple response domains and improves diagnostic reliability when administered systematically [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Yet behavior remains an indirect readout. A patient may understand commands but fail to produce movement, or may show only intermittent signs because arousal is unstable. This mismatch between cognition and action has driven the development of \u0026ldquo;command-following independent\u0026rdquo; assessments. Task-based functional MRI (fMRI) studies demonstrated that some patients diagnosed as vegetative could volitionally modulate brain activity during mental imagery, which provided a proof-of-concept that awareness can exist without outward response [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. More recent large multicenter work extends this principle with combined fMRI and EEG paradigms, linking covert cognition to clinically meaningful outcomes [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Together, these findings reinforce a clinical priority: when behavior is limited, physiology must be used to reduce the risk of missed awareness.\u003c/p\u003e \u003cp\u003eResting-state fMRI (rs-fMRI) offers a practical approach in this setting because it does not require active participation. Spontaneous blood oxygenation level\u0026ndash;dependent (BOLD) fluctuations reveal intrinsic functional organization, including long-range interactions among distributed systems that support cognition and state regulation. Contemporary models emphasize that consciousness depends on coordinated activity across large-scale networks and thalamo-cortical circuits rather than isolated regional activation. The mesocircuit hypothesis, for example, links impaired cortico-striato-thalamo-cortical dynamics to reduced capacity to sustain globally integrated processing after severe injury [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Network-oriented perspectives also highlight interactions among the default mode network, salience network, and frontoparietal control systems as core components of higher-order brain function, with relevance to both normal cognition and clinical dysfunction [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. In practice, rs-fMRI can be summarized using complementary markers that capture different aspects of functional integrity. The amplitude of low-frequency fluctuations (ALFF) reflects regional signal power and baseline activity patterns [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Regional homogeneity (ReHo) indexes local synchronization among neighboring voxels [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Graph-based measures such as degree centrality (DC) quantify how strongly a region participates as a functional hub within whole-brain coupling [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Together with pairwise functional connectivity (FC), these measures provide a multi-scale description of preserved and disrupted network function that is well aligned with mechanistic questions in pDoC.\u003c/p\u003e \u003cp\u003eDespite strong rationale and accumulating evidence, two translation barriers remain. First, many rs-fMRI studies report group differences but do not deliver clinically reliable, patient-level tools that generalize across hospitals, scanners, and etiologies. Overfitting and site effects are well-recognized challenges in medical neuroimaging, and models that perform well in a single cohort may degrade when applied externally. Beyond classification accuracy, the clinical utility of neuroimaging in pDoC hinges on interpretability. For clinicians and families facing end-of-life or long-term rehabilitation decisions, a 'black-box' model offers little guidance. Understanding the specific functional circuits\u0026mdash;rather than just the diagnostic label\u0026mdash;is essential for identifying therapeutic targets and predicting neuroplastic potential. This interpretability gap also limits the ability to translate predictive features into neurophysiological hypotheses about why MCS patients maintain consciousness and recover better than UWS patients. Explainable machine learning (EML) has therefore become increasingly relevant in medicine, not as an engineering add-on, but as a clinical requirement for transparency and trust [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. In particular, Shapley-value\u0026ndash;based approaches provide a principled way to quantify how each feature contributes to a model\u0026rsquo;s output in an individual patient and across a cohort [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eGuided by these needs, this study integrates multicenter rs-fMRI with clinically grounded interpretation to address two aims. First, we develop and externally validate an imaging-based diagnostic model to identify MCS among pDoC patients using a harmonized acquisition protocol across two independent clinical centers. We derive rs-fMRI markers that quantify regional activity and whole-brain network organization across a validated functional atlas, enabling patient-level classification that is designed to be robust across sites [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. This diagnostic focus is clinically motivated by early recognition of covert or fluctuating awareness, which can influence rehabilitation engagement, communication attempts, and therapeutic intensity. Second, and central to the medical interpretation of pDoC, we quantify the contribution of specific networks and brain regions to MCS classification using an explainability framework. This allows us to move from \u0026ldquo;which label fits\u0026rdquo; to \u0026ldquo;which preserved circuit functions matter,\u0026rdquo; with the goal of clarifying neural mechanisms that help patients remain in MCS rather than decline to UWS. By aligning predictive modeling with interpretable neurobiology, the study aims to support earlier diagnosis while also generating clinically legible hypotheses about network-level substrates of consciousness maintenance after severe brain injury.\u003c/p\u003e"},{"header":"2. Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Study design and Participants\u003c/h2\u003e \u003cp\u003eThis multicenter, cross-sectional study was approved by the Ethics Committee of the China Rehabilitation Research Center (No. 2023-012-01), with ethical oversight from the Affiliated Hospital of Qingdao University, and conducted in accordance with the Declaration of Helsinki. A total of 100 patients with pDoC were enrolled from two independent clinical centers. The discovery cohort included 50 patients (25 MCS and 25 UWS) recruited from the China Rehabilitation Research Center between June 2023 and June 2025. The external validation cohort consisted of 50 patients (25 MCS and 25 UWS) from the Affiliated Hospital of Qingdao University.\u003c/p\u003e \u003cp\u003eInclusion criteria: (1) a diagnosis of UWS or MCS based on behavioral assessments with the CRS-R; (2) a disease duration\u0026thinsp;\u0026ge;\u0026thinsp;28 days with a stable clinical condition; (3) the first onset of a prolonged disorder of consciousness; (4) age between 18 and 75 years ; and (5) written informed consent obtained from the family or legal representative.\u003c/p\u003e \u003cp\u003eExclusion criteria: (1) severe medical comorbidities, such as heart failure, renal failure, acute lung injury, or acute pulmonary infection; (2) a history of neurological or psychiatric diseases, including previous stroke, traumatic brain injury, or depression; (3) a history of alcohol or drug abuse; (4) intracranial metal foreign bodies or other contraindications to MRI; (5) neuromodulation therapy within the past three months; (6) persistent involuntary shaking or movements that precluded successful fMRI data acquisition.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Clinical Assessment and Outcome Definition\u003c/h2\u003e \u003cp\u003eConsciousness levels were rigorously evaluated using the CRS-R, which comprises six subscales: auditory, visual, motor, oromotor/verbal, communication, and arousal. To account for potential fluctuations in arousal and minimize diagnostic error, two trained physicians jointly performed three separate assessments within a single day. These evaluations were conducted during the morning, afternoon, and evening. The final behavioral diagnosis and consciousness level were determined by the highest total score achieved across the three assessment sessions.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 MRI Data Acquisition\u003c/h2\u003e \u003cp\u003eTo ensure the consistency and comparability of neuroimaging data across the multicenter study, all participants from both the discovery and external validation cohorts were scanned using identical 3.0T scanner models (Philips Ingenia, Philips Healthcare, The Netherlands) and strictly synchronized imaging protocols. Resting-state BOLD parameters at both centers were set as follows: repetition time\u0026thinsp;=\u0026thinsp;2000 ms, echo time\u0026thinsp;=\u0026thinsp;30 ms, flip angle\u0026thinsp;=\u0026thinsp;90\u0026deg;, field of view\u0026thinsp;=\u0026thinsp;224 mm, slice thickness\u0026thinsp;=\u0026thinsp;3.5 mm, inter-slice gap\u0026thinsp;=\u0026thinsp;0.85 mm, 32 slices, and a matrix size of 64 \u0026times; 64. For each participant at each time point, three consecutive resting-state runs were acquired, with each run lasting 8 minutes. Standardized precautions were implemented across both sites to minimize data artifacts, including the use of secure head restraints to reduce motion and earplugs or headphones to attenuate acoustic noise.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Data Preprocessing\u003c/h2\u003e \u003cp\u003ePreprocessing was conducted in MATLAB 2023a using the RestPlus toolbox. Data were converted from DICOM to NIfTI format and the first 10 volumes were discarded. Images then underwent slice-timing correction and rigid-body realignment for head-motion correction. Participants were excluded if their head motion exceeded 3 mm of displacement or 3\u0026deg; of rotation in any direction. Structural T1 images were segmented using DARTEL and functional images were normalized to MNI space with resampling to 3\u0026times;3\u0026times;3 mm\u0026sup3;. Linear detrending and temporal band-pass filtering (0.01\u0026ndash;0.08 Hz) were applied. Nuisance regression included Friston-24 motion parameters and signals from white matter and cerebrospinal fluid. To reduce spatial blurring across neighboring regions and inflation of inter-regional correlations, spatial smoothing was applied only for ALFF computation (Gaussian kernel, 6-mm FWHM). In contrast, DC, FC, and ReHo were derived from unsmoothed data.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Computation of Brain Network Metrics\u003c/h2\u003e \u003cp\u003eWe utilized a refined version of the Power-264 functional atlas to define the network nodes. The atlas has since been widely adopted and validated, supporting its reproducibility across studies [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. The refined atlas identified 236 key nodes distributed across 13 functional networks, spanning multiple systems such as the Cingulo-opercular (CON), Default Mode (DMN), Frontoparietal (FPCN), Subcortical (SCN), Cerebellar (CN), and Dorsal Attention (DAN) networks.\u003c/p\u003e \u003cp\u003eTo enhance robustness against residual registration errors and local noise, ROI-level measures were extracted from a 27-voxel neighborhood centered on each atlas coordinate. This neighborhood comprised the central coordinate voxel and its 26 adjacent voxels. All network metrics were computed using the RestPlus toolbox. ALFF: Calculated on smoothed data by transforming the BOLD time series to the frequency domain and averaging the square root of the power spectrum (0.01\u0026ndash;0.08 Hz); ReHo: Computed on unsmoothed data using Kendall\u0026rsquo;s coefficient of concordance to capture local synchronization between a voxel and its 26 functional neighbors; DC: Quantified on unsmoothed data as the sum of suprathreshold Pearson correlations (r\u0026thinsp;\u0026gt;\u0026thinsp;0.25) between a voxel and the rest of the brain; FC: Pairwise Pearson correlations between representative ROI time series were computed and Fisher z-transformed for statistical normality.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.6 Dimensionality Reduction and Predictive Modeling\u003c/h2\u003e \u003cp\u003eIn the discovery cohort, candidate neuroimaging features were first identified by group-level comparisons between the MCS and UWS groups. Independent-samples t-tests were performed with age, sex, disease duration, etiology, and mean head motion parameters entered as covariates. Multiple comparisons were controlled using the false discovery rate procedure. This initial screening reduced the feature space from several tens of thousands of measures to 20 variables. To further limit overfitting, sparse feature selection was then conducted using L1-regularized logistic regression (Logistic LASSO). The regularization parameter was tuned by stratified five-fold cross-validation within the discovery set using a prespecified grid of inverse penalty values (C). Model selection was based on the mean area under the receiver operating characteristic curve across folds. To prevent information leakage, Z-score normalization was performed within the cross-validation framework, with scaling parameters estimated from the training split of each fold and applied to the corresponding held-out split. After selecting the optimal parameter, the model was refit on the full discovery set, and variables with non-zero coefficients were retained as the final feature set for subsequent analyses.\u003c/p\u003e \u003cp\u003eThe retained features were then used to construct nine distinct machine learning classifiers: Artificial Neural Network, Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree, LightGBM, Gradient Boosting, Extra Trees, Random Forest, and XGBoost. During the training phase, hyperparameter optimization for each model was conducted via grid search with five-fold cross-validation on the discovery set. The Area Under the Receiver Operating Characteristic Curve (AUC) was utilized as the primary performance indicator to evaluate and compare the diagnostic capability of the candidate models.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e2.7 Model Validation and Explainability\u003c/h2\u003e \u003cp\u003eThe final models were applied to the independent external validation cohort from Qingdao University to assess generalizability. Performance was evaluated using the AUC and secondary metrics, including accuracy, sensitivity, specificity, F1-score, and Cohen\u0026rsquo;s kappa. To elucidate the neurobiological mechanisms distinguishing MCS from UWS, we applied SHAP (SHapley Additive exPlanations) to the best-performing model to quantify the contribution of individual network features to the diagnostic prediction.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e2.8 Statistical Analysis\u003c/h2\u003e \u003cp\u003eStatistical analyses were performed using Python 3.12. Continuous variables are expressed as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation (SD) and compared between MCS and UWS groups within each center using independent-samples t-tests. Categorical variables (e.g., sex) were compared using Chi-square tests. All tests were two-sided, and p\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e2.9 Use of generative AI tools\u003c/h2\u003e \u003cp\u003eAs non-native English speakers, we used Google Gemini to improve language clarity and grammar during manuscript preparation. Gemini was also used to assist in drafting the graphical abstract and preparing the study flowchart. All figures presenting results from data analyses are original, unaltered outputs generated by the analysis pipeline. All content generated or revised with AI assistance was reviewed and verified by the authors, who take full responsibility for the final manuscript.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Demographic and Clinical Characteristics\u003c/h2\u003e \u003cp\u003eIn the discovery cohort, 50 participants were recruited. Two were excluded from the final analysis due to suboptimal fMRI data quality, leaving 48 patients, including 24 with MCS and 24 with UWS. In the external validation cohort, 50 participants were recruited; four were excluded due to excessive head motion, resulting in a final sample of 46 patients (21 with MCS and 25 with UWS). Statistical analysis revealed no significant differences between the MCS and UWS groups regarding age, sex, disease duration, or etiology in either cohort (P\u0026gt;0.05, Tables\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e and \u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). As expected, baseline CRS-R scores were significantly higher in the MCS group than in the UWS group for both the discovery and validation cohorts (P\u0026lt;0.05, Tables\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e and \u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). The schematic representation of the integrated multicenter study design and the machine learning-based diagnostic pipeline is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDemographic and clinical characteristics of the discovery cohort.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiscovery cohort\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMCS (n\u0026thinsp;=\u0026thinsp;24)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eUWS (n\u0026thinsp;=\u0026thinsp;24)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eStatistic\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eP value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge, years\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e52.33\u0026thinsp;\u0026plusmn;\u0026thinsp;10.21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e54.62\u0026thinsp;\u0026plusmn;\u0026thinsp;7.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003et = -0.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.382\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMale, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e18 (75.0%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e16 (66.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eχ\u0026sup2; = 0.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.751\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTraumatic brain injury, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10 (41.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12 (50.0%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eχ\u0026sup2; = 0.08\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.772\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTime since injury, days\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e121.83\u0026thinsp;\u0026plusmn;\u0026thinsp;67.23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e98.92\u0026thinsp;\u0026plusmn;\u0026thinsp;41.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003et = -1.419\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.163\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCRS-R total score\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e11.96\u0026thinsp;\u0026plusmn;\u0026thinsp;2.05\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.29\u0026thinsp;\u0026plusmn;\u0026thinsp;1.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003et\u0026thinsp;=\u0026thinsp;13.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eNotes: Continuous variables are presented as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation. Categorical variables are expressed as counts and percentages. P-values were calculated using independent-samples t-tests or Chi-square tests. Abbreviations: MCS, minimally conscious state; UWS, unresponsive wakefulness syndrome; CRS-R, Coma Recovery Scale\u0026ndash;Revised\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDemographic and clinical characteristics of the validation cohort.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eValidation Cohort\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMCS (n\u0026thinsp;=\u0026thinsp;21)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eUWS (n\u0026thinsp;=\u0026thinsp;25)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eStatistic\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eP value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge, years\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e51.76\u0026thinsp;\u0026plusmn;\u0026thinsp;10.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e53.08\u0026thinsp;\u0026plusmn;\u0026thinsp;8.49\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003et = -0.48\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.637\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMale, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e16 (76.2%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e17 (68.0%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eχ\u0026sup2; = 0.08\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.775\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTraumatic brain injury, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9 (42.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e14 (56.0%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eχ\u0026sup2; = 0.35\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.554\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTime since injury, days\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e116.24\u0026thinsp;\u0026plusmn;\u0026thinsp;67.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e112.48\u0026thinsp;\u0026plusmn;\u0026thinsp;58.18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003et\u0026thinsp;=\u0026thinsp;0.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.841\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCRS-R total score\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e12.14\u0026thinsp;\u0026plusmn;\u0026thinsp;1.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.28\u0026thinsp;\u0026plusmn;\u0026thinsp;1.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003et\u0026thinsp;=\u0026thinsp;14.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003eNotes: Continuous variables are presented as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation. Categorical variables are expressed as counts and percentages. P-values were calculated using independent-samples t-tests or Chi-square tests.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eParticipants were recruited into a discovery cohort (Center 1) and an independent validation cohort (Center 2) (each n\u0026thinsp;=\u0026thinsp;50; 25 MCS and 25 UWS). Diagnosis was established using repeated CRS-R as-sessments, followed by 3.0-T rs-fMRI acquired with identical protocols. After standardized pre-processing and motion-based quality control (total excluded n\u0026thinsp;=\u0026thinsp;6), brain network features were ex-tracted using the Power-264 atlas (ALFF, ReHo, DC, and FC). In the discovery cohort, features were screened (FDR-corrected) and further selected using L1-regularized logistic regression with five-fold cross-validation, and nine classifiers were trained and optimized. The leading model was then applied to the external cohort for performance evaluation and interpreted using SHAP. Ab-breviations: MCS, minimally conscious state; UWS, unresponsive wakefulness syndrome; CRS-R, Coma Recovery Scale\u0026ndash;Revised; rs-fMRI, resting-state functional magnetic resonance imaging; ALFF, amplitude of low-frequency fluctuations; ReHo, regional homogeneity; DC, degree centrality; FC, functional connectivity; FDR, false discovery rate; SHAP, SHapley Additive exPlanations.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Differences in Resting-state fMRI\u003c/h2\u003e \u003cdiv id=\"Sec15\" class=\"Section3\"\u003e \u003ch2\u003e3.2.1 ALFF Analysis\u003c/h2\u003e \u003cp\u003ePatients in the MCS group exhibited significantly higher ALFF than those in the UWS group across multiple functional networks. These increases were identified in default mode network nodes, including DMN.86 (left angular gyrus; t\u0026thinsp;=\u0026thinsp;5.273; P\u0026lt;0.05; Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e), DMN.117 (left middle temporal gyrus; t\u0026thinsp;=\u0026thinsp;6.812; P\u0026lt;0.05; Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e), and DMN.130 (right angular gyrus; t\u0026thinsp;=\u0026thinsp;5.835). Higher ALFF was also detected in the subcortical node SCN.223 (left thalamus; t\u0026thinsp;=\u0026thinsp;6.132; P\u0026lt;0.05; Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) and the cerebellar node CN.245 (right cerebellum; t\u0026thinsp;=\u0026thinsp;5.722; P\u0026lt;0.05; Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eRegional signal power was significantly increased in the MCS group (red spheres) within default mode network (DMN) nodes, specifically the left angular gyrus (DMN.86), left middle temporal gyrus (DMN.117), and right angular gyrus (DMN.130). Additional increases were localized to the left thalamus (SCN.223) and right cerebellum (CN.245). Color bar represents t-values (FDR corrected, P\u0026thinsp;\u0026lt;\u0026thinsp;0.05). Abbreviations: MCS, minimally conscious state; UWS, unresponsive wakefulness syndrome; DMN, default mode network; SCN, subcortical network; CN, cerebellar network; FDR, false discovery rate.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section3\"\u003e \u003ch2\u003e3.2.2 ReHo Analysis\u003c/h2\u003e \u003cp\u003eRegarding ReHo, the MCS group showed significantly higher local synchronization in DMN.86 (left angular gyrus; t\u0026thinsp;=\u0026thinsp;6.452; P\u0026lt;0.05; Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e) and DMN.92 (right posterior cingulate gyrus; t\u0026thinsp;=\u0026thinsp;5.573; P\u0026lt;0.05; Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). Additional increases in ReHo were observed in FPCN.176 (left inferior frontal gyrus, opercular part; t\u0026thinsp;=\u0026thinsp;5.879; P\u0026lt;0.05; Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e) and SN.208 (left insula; t\u0026thinsp;=\u0026thinsp;6.204; P\u0026lt;0.05; Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe MCS group demonstrated significantly higher local synchronization (red spheres) in the left angular gyrus (DMN.86), right posterior cingulate gyrus (DMN.92), left inferior frontal gyrus (FPCN.176), and left insula (SN.208). Color bar represents t-values (FDR corrected, P\u0026thinsp;\u0026lt;\u0026thinsp;0.05). Abbreviations: FPCN, frontoparietal control network; SN, salience network.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section3\"\u003e \u003ch2\u003e3.2.3 DC Analysis\u003c/h2\u003e \u003cp\u003eDC was significantly elevated in the MCS group compared with the UWS group. This effect involved several key nodes: DMN.78 (left superior frontal gyrus, orbital part; t\u0026thinsp;=\u0026thinsp;5.586; P\u0026lt;0.05; Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e), MRN.136 (right precuneus; t\u0026thinsp;=\u0026thinsp;6.139; P\u0026lt;0.05; Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e), SCN.223 (left thalamus; t\u0026thinsp;=\u0026thinsp;6.081; P\u0026lt;0.05; Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e), and DAN.251 (right precuneus; t\u0026thinsp;=\u0026thinsp;5.882; P\u0026lt;0.05; Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eKey functional hubs identified in MCS included the left superior frontal gyrus (DMN.78), right precuneus (MRN.136 and DAN.251), and left thalamus (SCN.223). Red spheres indicate regions with significant global connectivity increases compared to UWS (FDR corrected, P\u0026thinsp;\u0026lt;\u0026thinsp;0.05). Abbreviations: MRN, memory retrieval network; DAN, dorsal attention network.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section3\"\u003e \u003ch2\u003e3.2.4 FC Analysis\u003c/h2\u003e \u003cp\u003eBetween-group comparisons revealed significantly higher FC in patients with MCS compared to those with UWS. The most extensive increases were observed in DMN related connections, including DMN.108\u0026ndash;VAN.239 (right medial superior frontal gyrus\u0026ndash;right middle temporal gyrus, t\u0026thinsp;=\u0026thinsp;6.601; P\u0026lt;0.05; Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e), DMN.117\u0026ndash;SCN.231 (left middle temporal gyrus\u0026ndash;right putamen, t\u0026thinsp;=\u0026thinsp;6.121; P\u0026lt;0.05; Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e), DMN.124\u0026ndash;MRN.136 (left parahippocampal gyrus\u0026ndash;right precuneus, t\u0026thinsp;=\u0026thinsp;5.892; P\u0026lt;0.05; Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e), DMN.79\u0026ndash;FPCN.201 (left middle temporal gyrus\u0026ndash;left inferior frontal gyrus, t\u0026thinsp;=\u0026thinsp;5.481; P\u0026lt;0.05; Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e), and DMN.90\u0026ndash;SCN.223 (left precuneus\u0026ndash;left thalamus, t\u0026thinsp;=\u0026thinsp;5.284; P\u0026lt;0.05; Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e). Furthermore, increased FC was identified in cerebellar-related pathways, including CN.245\u0026ndash;DMN.86 (right cerebellum\u0026ndash;left angular gyrus, t\u0026thinsp;=\u0026thinsp;6.308; P\u0026lt;0.05; Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e) and CN.245\u0026ndash;SCN.223 (right cerebellum\u0026ndash;left thalamus, t\u0026thinsp;=\u0026thinsp;5.572; P\u0026lt;0.05; Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e). Finally, enhanced functional coupling was observed for FPCN.186\u0026ndash;VAN.239 (right precentral gyrus\u0026ndash;right middle temporal gyrus, t\u0026thinsp;=\u0026thinsp;5.933; P\u0026lt;0.05; Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eConnectome analysis revealed enhanced functional coupling in MCS (red lines), predominantly involving DMN-subcortical, DMN-cerebellar, and DMN-frontoparietal circuits. Blue spheres represent network nodes defined by the Power-264 atlas (FDR corrected, P\u0026thinsp;\u0026lt;\u0026thinsp;0.05). Abbreviations: VAN, ventral attention network\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Feature Selection by Logistic L1 Regularization\u003c/h2\u003e \u003cp\u003eInitial statistical screening across the discovery cohort identified 21 neuroimaging features with significant group differences. These features were entered into a L1-regularized logistic regression model for secondary screening. The inverse regularization strength (C) was tuned using stratified five-fold cross-validation, with the mean cross-validated AUC as the optimization criterion. The optimal value was C\u0026thinsp;=\u0026thinsp;0.412, yielding a mean cross-validated AUC of 0.818. At this parameter setting, 10 features had non-zero coefficients and were therefore retained in the final feature set (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eA). As C increased beyond the optimum, additional features entered the model, but the cross-validated AUC showed no further improvement. This procedure yielded 10 core features for diagnostic modeling: ALFF-DMN.117, DMN.86-CN.245 FC, DMN.117-SCN.231 FC, ALFF-DMN.130, DMN.90-SCN.223 FC, SCN.223-CN.245 FC, DMN.108-VAN.239 FC, DMN.124-MRN.136 FC, FPCN.186-VAN.239 FC, and ReHo-SN.208 (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eB).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e(A) Mean cross-validated AUC (left y-axis) and the number of retained features with non-zero coefficients (right y-axis) are shown as a function of the inverse regularization parameter C in the discovery cohort using stratified five-fold cross-validation. The red dashed line marks the optimal value (C\u0026thinsp;=\u0026thinsp;0.412), which yielded a mean cross-validated AUC of 0.818 and retained 10 features. (B) Coefficients of the 10 selected neuroimaging features from the Logistic L1 model used for downstream diagnostic modeling.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Diagnostic Performance and External Validation\u003c/h2\u003e \u003cp\u003eNine machine learning classifiers were evaluated for their ability to distinguish MCS from UWS. While several models achieved near-perfect performance in the training set, the models were primarily assessed based on their generalizability to the external validation cohort.\u003c/p\u003e \u003cp\u003eThe SVM was identified as the most robust model, achieving an AUC of 0.859 in the external validation set (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003eA). The SVM model demonstrated an accuracy of 0.826, a sensitivity of 0.905, and a specificity of 0.760 (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003eC). Other classifiers, such as XGBoost and KNN, also yielded strong performance with AUCs of 0.850 (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003eA). Detailed performance metrics for all classifiers are provided in the Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e(A-B) ROC curves for nine classifiers in the discovery and external validation cohorts. SVM achieved the highest AUC of 0.859 in the external set. (C-D) Comparison of accuracy, sensitivity, specificity, and other performance metrics across models. Abbreviations: ROC, receiver operating characteristic; AUC, area under the curve; SVM, support vector machine.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003e3.5 Model Interpretability by SHAP\u003c/h2\u003e \u003cp\u003eSHAP values were calculated for the optimized SVM model to identify the key neurobiological signatures of MCS. The analysis indicated that FC between DMN.86 and CN.245 was the most influential feature for the diagnostic prediction (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003eA). Other critical contributors included DMN.117-SCN.231 FC, ReHo-SN.208, ALFF-DMN.130, and SCN.223-CN.245 FC (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003eA). Patients with higher values in these network preservation were significantly more likely to be classified as MCS rather than UWS (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003eB).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e(A) Global feature importance ranked by the mean absolute SHAP value, identifying DMN.86\u0026ndash;CN.245 FC as the primary predictor. (B) SHAP summary plot showing that higher values of network preservation (pink) increase the probability of an MCS diagnosis.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eIn this study, we show that patients with MCS exhibit a clear preservation of large-scale network function compared with UWS, with the most robust signal concentrated in the DMN, CN, and key cortico\u0026ndash;subcortical circuits. Consistent classification performance across model training and external testing indicates that these preserved interactions provide clinically informative features for differentiating MCS from UWS. Importantly, the SHAP-based attribution analysis did not point to a diffuse connectome effect, but instead converged on a small set of biologically coherent contributors, with the cross-network coupling DMN.86\u0026ndash;CN.245 emerging as the most characteristic preserved feature in MCS. This pattern supports a clinically grounded interpretation that consciousness maintenance in MCS relies on sustained communication between cortical integrative systems and subcortical enabling pathways, rather than on isolated regional activity.\u003c/p\u003e \u003cp\u003eA major finding was that the most informative features were not diffuse across the connectome, but converged on hubs within cortico\u0026ndash;subcortical circuits and the DMN. This is clinically meaningful because behavioral assessment alone is vulnerable to under-detection of residual awareness when motor output is impaired, arousal fluctuates, or bedside conditions are suboptimal, all of which are common in pDoC [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. In that context, a model that extracts stable diagnostic signal from resting-state networks addresses a key translational gap: shifting the diagnostic emphasis from observed behavior to preserved integrative physiology. Meta-analytic evidence supports that fMRI-derived markers show moderate diagnostic value for differentiating MCS from UWS/VS, reinforcing the plausibility of our approach and the need for clinically practical, interpretable implementations [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe interpretability analysis further suggested that preservation within posterior associative cortices and their subcortical coupling is central to differentiating MCS from UWS. Nodes with strong positive contributions included the left angular gyrus and right middle temporal gyrus, regions that participate in the temporoparietal junction and lateral temporal components of the DMN and semantic-associative processing streams. In mechanistic terms, these regions are positioned to support multimodal integration and internal model updating, functions that are repeatedly implicated in conscious access and the maintenance of a coherent internal milieu. Modern accounts increasingly view consciousness as depending on large-scale integration rather than isolated regional activity, with the DMN acting as a key convergence system that is especially vulnerable across pharmacological and pathological perturbations of consciousness [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Importantly, in pDoC, reduced DMN integrity and disrupted interaction between DMN and subcortical arousal systems are among the most reproducible network abnormalities, and are closely tied to diagnostic category and recovery potential [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Our results extend this by localizing clinically useful diagnostic information to specific DMN-associated cortical nodes rather than treating the DMN as a monolithic entity.\u003c/p\u003e \u003cp\u003eSubcortical nodes with high contribution, particularly the left thalamus and right putamen, support a complementary interpretation rooted in thalamo-striato-cortical gating. The thalamus is not merely a relay, but a central regulator of cortical effective connectivity and state transitions. Circuit-level work demonstrates layer-specific thalamic control over cortex that tracks changes in consciousness level, providing a mechanistic substrate for why thalamic integrity can be disproportionately informative in consciousness disorders [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. In clinical pDoC cohorts, abnormalities in cortico\u0026ndash;striato\u0026ndash;pallido\u0026ndash;thalamo\u0026ndash;cortical loops have been repeatedly observed and are associated with level of consciousness, aligning with mesocircuit models that emphasize thalamic underactivation and impaired cortico-subcortical facilitation [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. The putamen\u0026rsquo;s contribution in our model may reflect preserved basal ganglia participation in action selection, salience-weighting, and cortical enabling conditions, which can influence the probability that residual cognition manifests as reproducible, command-related behavior. This interpretation is consistent with broader pDoC connectivity syntheses highlighting subcortical\u0026ndash;cortical disruptions as a core neurophysiologic mechanism rather than an epiphenomenon [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eOur findings challenge the traditional cortico-centric view of consciousness by highlighting the pivotal role of DMN\u0026ndash;cerebellar coupling in sustaining MCS. Cerebellar findings are often under-discussed in pDoC, yet the cerebellum is increasingly recognized as embedded within multiple intrinsic networks, including DMN, SN, and executive systems, through cerebello-thalamo-cortical loops [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. In our data, cerebellum involvement may reflect preservation of timing, error monitoring, and autonomic-affective integration that indirectly stabilizes large-scale cortical dynamics necessary for conscious processing. While the cerebellum is unlikely to \u0026ldquo;generate\u0026rdquo; conscious content in isolation, it can modulate the consistency and coordination of distributed processing, thereby supporting the network conditions under which consciousness is sustained. This view aligns with contemporary connectivity frameworks that emphasize distributed control and the vulnerability of network coordination in pDoC [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eComparison with existing literature highlights both consistency and novelty. Our emphasis on DMN\u0026ndash;subcortical circuitry matches a large body of work identifying DMN disruption and thalamocortical decoupling as key signatures of impaired consciousness, and it aligns with clinical calls to incorporate neuroimaging into routine pDoC assessment to reduce diagnostic error and identify covert awareness [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. At the same time, our study contributes by pairing diagnostic modeling with node-level interpretability, enabling a clinically legible statement: MCS is characterized less by \u0026ldquo;global connectivity\u0026rdquo; and more by selective preservation of integrative hubs and their subcortical enabling circuits. This is compatible with recent approaches using interpretable machine learning in multicentre pDoC datasets, which also found meaningful interactions between modality and anatomical structures and emphasized that interpretability can bridge the gap between prediction and mechanism [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. Moreover, deep learning approaches have demonstrated strong discrimination between MCS and UWS using rs-fMRI, supporting the feasibility of automated detection of residual awareness; however, many such models are criticized for limited transparency, which can impede clinical adoption [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. Our work addresses this translational barrier by keeping mechanistic interpretability central rather than ancillary.\u003c/p\u003e \u003cp\u003eThe clinical significance of this work lies in two domains. First, earlier and more accurate identification of MCS has immediate downstream consequences for treatment intensity, rehabilitation planning, ethical decision-making, and family counseling. Contemporary guidance emphasizes that functional neuroimaging can identify covert awareness in a substantial minority of behaviorally unresponsive patients, and that implementation barriers are increasingly logistical rather than conceptual [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Second, by identifying which brain regions and networks carry the strongest diagnostic signal, the study advances a mechanistic hypothesis with therapeutic implications: interventions that strengthen DMN hubs (PCC/precuneus and lateral parietotemporal nodes) and thalamo-striato-cortical coupling, potentially including cerebello-thalamic modulation, may be rational targets for future therapies. Such targeting can also inform trial stratification, because network preservation may predict responsiveness to neuromodulatory approaches aimed at restoring large-scale integration [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eSeveral limitations should be considered when interpreting these findings. The model was trained on resting-state connectivity, which is sensitive to acquisition differences, head motion, physiological noise, and variations in vigilance. Although external testing reduces the risk of overfitting to a single-site distribution, multicentre harmonization and prospective evaluation remain essential before clinical deployment. Additionally, diagnostic labels in pDoC are imperfect because bedside behavior can fluctuate and can be confounded by motor impairment; thus, any supervised model trained on behavioral diagnosis inherits some label noise. This issue is widely acknowledged in the field and is a central argument for integrating imaging with repeated standardized assessment rather than replacing it [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. Finally, while interpretability helps link predictive features to plausible neurobiology, attribution does not prove causality. The identified nodes should therefore be treated as candidate biomarkers of preserved network function rather than definitive mechanistic drivers. Future studies combining rs-fMRI with task-based paradigms, electrophysiology, and longitudinal outcomes will be important to separate markers of current state from predictors of recovery [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e].\u003c/p\u003e"},{"header":"6. Conclusions","content":"\u003cp\u003eIn summary, this study supports the clinical utility of connectome-based modeling for differentiating MCS from UWS and provides anatomically grounded, interpretable evidence that preservation of DMN-associated parietotemporal hubs, thalamic and striatal circuitry, and cerebello-thalamo-cortical contributions are key network features of MCS. Future work should prioritize prospective multicentre validation, longitudinal outcome prediction, and multimodal integration to determine whether strengthening these preserved networks can be translated into targeted therapies and improved patient trajectories.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eALFF\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eAmplitude of low-frequency fluctuations\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eAUC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eArea under the curve\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCN\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eCerebellar network\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCON\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eCingulo-opercular network\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCRS-R\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eComa Recovery Scale\u0026ndash;Revised\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eDAN\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eDorsal attention network\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eDC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eDegree centrality\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eDMN\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eDefault mode network\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eFC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eFunctional connectivity\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eFDR\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eFalse discovery rate\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eFPCN\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eFrontoparietal control network\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLASSO\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eLeast absolute shrinkage and selection operator\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMCS\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMinimally conscious state\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003epDoC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eProlonged disorders of consciousness\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eReHo\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eRegional homogeneity\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eRF\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eRandom forest\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eROC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eReceiver operating characteristic\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eROI\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eRegion of interest\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003ers-fMRI\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eResting-state functional MRI\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSCN\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSubcortical network\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSD\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eStandard deviation\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSHAP\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSHapley additive explanations\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSN\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSalience network\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eUWS\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eUnresponsive wakefulness syndrome\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003eEthics approval and consent to participate: The study was conducted in accordance with the Declaration of Helsinki, and approved by the Ethics Committees of the China Rehabilitation Research Center and the Affiliated Hospital of Qingdao University (protocol code 2023–012-01).\u0026nbsp;Written informed consent was obtained from each participant or, when the participant lacked decision-making capacity, from a legally authorized representative.\u003c/p\u003e\n\u003cp\u003eConsent for publication: Not applicable\u003c/p\u003e\n\u003cp\u003eAvailability of data and materials: The data that support the findings of this study are not publicly available due to ethical and privacy restrictions. Deidentified data may be made available from the corresponding author upon reasonable request and with approval from the relevant ethics committee.\u003c/p\u003e\n\u003cp\u003eCompeting interests: The authors declare that they have no competing interests\u003c/p\u003e\n\u003cp\u003eFunding: This study was self-funded and supported by personal funds from Zhanghao (grant number: 2022HZ-06-01). The article processing charge (APC) was also covered by Zhanghao’s personal funds (grant number: 2022HZ-06-01). The funder (Zhanghao, the corresponding author) was involved in the study as an author; no additional roles beyond the authors’\u0026nbsp;contributions were imposed by any external funding body. The funder had no independent role, separate from the authorship role, in the study design, data collection, data analysis, interpretation, the decision to publish, or the preparation of the manuscript.\u003c/p\u003e\n\u003cp\u003eAuthors' contributions: HZ: Conceptualization, project administration. LD, HL, KH: Methodology. HL: Formal analysis. HL, ZT, JL, XL: Investigation. LD, HZ: Writing—original draft. KH, XL, TL, HZ: Writing—review and editing. XL, HZ: Supervision. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003eAcknowledgements: We gratefully acknowledge all assessors at the China Rehabilitation Research Center and the Affiliated Hospital of Qingdao University who participated in the clinical assess-ments of the enrolled patients.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eBender A, Jox RJ, Grill E, Straube A, Lul\u0026eacute; D. Persistent vegetative state and minimally conscious state: a sys-tematic review and meta-analysis of diagnostic procedures. Dtsch Arztebl Int. 2015;112(14):235\u0026ndash;42. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3238/arztebl.2015.0235\u003c/span\u003e\u003cspan address=\"10.3238/arztebl.2015.0235\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBodien YG, Allanson J, Cardone P, et al. Cognitive Motor Dissociation in Disorders of Consciousness. N Engl J Med. 2024;391(7):598\u0026ndash;608. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1056/NEJMoa2400645\u003c/span\u003e\u003cspan address=\"10.1056/NEJMoa2400645\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGiacino JT, Ashwal S, Childs N, et al. The minimally conscious state: definition and diagnostic criteria. Neurology. 2002;58(3):349\u0026ndash;53. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1212/wnl.58.3.349\u003c/span\u003e\u003cspan address=\"10.1212/wnl.58.3.349\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLaureys S, Celesia GG, Cohadon F, et al. Unresponsive wakefulness syndrome: a new name for the vegetative state or apallic syndrome. BMC Med. 2010;8:68. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/1741-7015-8-68\u003c/span\u003e\u003cspan address=\"10.1186/1741-7015-8-68\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Published 2010 Nov 1.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGiacino JT, Kalmar K, Whyte J. The JFK Coma Recovery Scale-Revised: measurement characteristics and diag-nostic utility. Arch Phys Med Rehabil. 2004;85(12):2020\u0026ndash;9. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.apmr.2004.02.033\u003c/span\u003e\u003cspan address=\"10.1016/j.apmr.2004.02.033\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOwen AM, Coleman MR, Boly M, Davis MH, Laureys S, Pickard JD. Detecting awareness in the vegetative state. Science. 2006;313(5792):1402. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1126/science.1130197\u003c/span\u003e\u003cspan address=\"10.1126/science.1130197\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchiff ND. Recovery of consciousness after brain injury: a mesocircuit hypothesis. Trends Neurosci. 2010;33(1):1\u0026ndash;9. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.tins.2009.11.002\u003c/span\u003e\u003cspan address=\"10.1016/j.tins.2009.11.002\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMenon V. Large-scale brain networks and psychopathology: a unifying triple network model. Trends Cogn Sci. 2011;15(10):483\u0026ndash;506. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.tics.2011.08.003\u003c/span\u003e\u003cspan address=\"10.1016/j.tics.2011.08.003\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZang YF, He Y, Zhu CZ, et al. Altered baseline brain activity in children with ADHD revealed by resting-state functional MRI. Brain Dev. 2007;29(2):83\u0026ndash;91. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.braindev.2006.07.002\u003c/span\u003e\u003cspan address=\"10.1016/j.braindev.2006.07.002\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZang Y, Jiang T, Lu Y, He Y, Tian L. Regional homogeneity approach to fMRI data analysis. NeuroImage. 2004;22(1):394\u0026ndash;400. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.neuroimage.2003.12.030\u003c/span\u003e\u003cspan address=\"10.1016/j.neuroimage.2003.12.030\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZuo XN, Ehmke R, Mennes M, et al. Network centrality in the human functional connectome. Cereb Cortex. 2012;22(8):1862\u0026ndash;75. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/cercor/bhr269\u003c/span\u003e\u003cspan address=\"10.1093/cercor/bhr269\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePahud de Mortanges A, Luo H, Shu SZ et al. Orchestrating explainable artificial intelligence for multimodal and longitudinal data in medical imaging. NPJ Digit Med. 2024;7(1):195. Published 2024 Jul 22. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41746-024-01190-w\u003c/span\u003e\u003cspan address=\"10.1038/s41746-024-01190-w\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLundberg SM, Erion G, Chen H, et al. From Local Explanations to Global Understanding with Explainable AI for Trees. Nat Mach Intell. 2020;2(1):56\u0026ndash;67. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s42256-019-0138-9\u003c/span\u003e\u003cspan address=\"10.1038/s42256-019-0138-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePower JD, Cohen AL, Nelson SM, et al. Functional network organization of the human brain. Neuron. 2011;72(4):665\u0026ndash;78. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.neuron.2011.09.006\u003c/span\u003e\u003cspan address=\"10.1016/j.neuron.2011.09.006\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKazazian K, Monti MM, Owen AM. Functional neuroimaging in disorders of consciousness: towards clinical implementation. Brain. 2025;148(7):2283\u0026ndash;98. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/brain/awaf075\u003c/span\u003e\u003cspan address=\"10.1093/brain/awaf075\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZheng H, Tian L, Cai J. Meta-analysis of the diagnostic value of functional magnetic resonance imaging for dis-tinguishing unresponsive wakefulness syndrome/vegetative state and minimally conscious state. Front Neurosci. 2024;18:1395639. Published 2024 Sep 9. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fnins.2024.1395639\u003c/span\u003e\u003cspan address=\"10.3389/fnins.2024.1395639\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHeine L, Soddu A, G\u0026oacute;mez F et al. Resting state networks and consciousness: alterations of multiple resting state network connectivity in physiological, pharmacological, and pathological consciousness States. Front Psychol. 2012;3:295. Published 2012 Aug 27. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fpsyg.2012.00295\u003c/span\u003e\u003cspan address=\"10.3389/fpsyg.2012.00295\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVanhaudenhuyse A, Noirhomme Q, Tshibanda LJ, et al. Default network connectivity reflects the level of con-sciousness in non-communicative brain-damaged patients. Brain. 2010;133(Pt 1):161\u0026ndash;71. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/brain/awp313\u003c/span\u003e\u003cspan address=\"10.1093/brain/awp313\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRedinbaugh MJ, Phillips JM, Kambi NA, et al. Thalamus Modulates Consciousness via Layer-Specific Control of Cortex. Neuron. 2020;106(1):66\u0026ndash;e7512. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.neuron.2020.01.005\u003c/span\u003e\u003cspan address=\"10.1016/j.neuron.2020.01.005\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi H, Dong L, Liu J, Zhang X, Zhang H. Abnormal characteristics in disorders of consciousness: A resting-state functional magnetic resonance imaging study. Brain Res. 2025;1850:149401. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.brainres.2024.149401\u003c/span\u003e\u003cspan address=\"10.1016/j.brainres.2024.149401\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePlosnić G, Raguž M, Deletis V, Chudy D. Dysfunctional connectivity as a neurophysiologic mechanism of dis-orders of consciousness: a systematic review. Front Neurosci. 2023;17:1166187. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fnins.2023.1166187\u003c/span\u003e\u003cspan address=\"10.3389/fnins.2023.1166187\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Published 2023 Jul 19.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHabas C. Functional Connectivity of the Cognitive Cerebellum. Front Syst Neurosci. 2021;15:642225. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fnsys.2021.642225\u003c/span\u003e\u003cspan address=\"10.3389/fnsys.2021.642225\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Published 2021 Apr 8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNaro A, Bramanti A, Leo A, et al. Shedding new light on disorders of consciousness diagnosis: The dynamic functional connectivity. Cortex. 2018;103:316\u0026ndash;28. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.cortex.2018.03.029\u003c/span\u003e\u003cspan address=\"10.1016/j.cortex.2018.03.029\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eManasova D, Belloli LML, Rosenfelder MJ, et al. Multimodal multicentre investigation of diagnostic and prog-nostic markers in disorders of consciousness. Brain Published online January. 2026;7. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/brain/awaf412\u003c/span\u003e\u003cspan address=\"10.1093/brain/awaf412\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang H, Wu H, Kong L, et al. Precise detection of awareness in disorders of consciousness using deep learning framework. NeuroImage. 2024;290:120580. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.neuroimage.2024.120580\u003c/span\u003e\u003cspan address=\"10.1016/j.neuroimage.2024.120580\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchnakers C, Vanhaudenhuyse A, Giacino J, et al. Diagnostic accuracy of the vegetative and minimally conscious state: clinical consensus versus standardized neurobehavioral assessment. BMC Neurol. 2009;9:35. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/1471-2377-9-35\u003c/span\u003e\u003cspan address=\"10.1186/1471-2377-9-35\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Published 2009 Jul 21.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu S, Zhu B, Ye Z et al. Functional connectivity in whole-brain and network analysis differentiates minimally conscious from unresponsive patients: a resting-state fNIRS study. J Transl Med. 2025;23(1):1093. Published 2025 Oct 14. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12967-025-07181-z\u003c/span\u003e\u003cspan address=\"10.1186/s12967-025-07181-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"prolonged disorders of consciousness, minimally conscious state, unresponsive wakefulness syndrome, resting-state fMRI, default mode network, machine learning","lastPublishedDoi":"10.21203/rs.3.rs-8959330/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8959330/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cb\u003eBackground\u003c/b\u003e\u003c/p\u003e \u003cp\u003eAccurate differentiation between minimally conscious state (MCS) and unresponsive wakefulness syndrome (UWS) is a significant clinical challenge because behavioral assessments are often constrained by patients' motor impairments and fluctuating arousal. Beyond diagnostic classification, clarifying the specific brain network mechanisms that sustain residual consciousness in MCS remains a clinical priority for developing targeted therapeutic interventions.\u003c/p\u003e\u003cp\u003e\u003cb\u003eMethods\u003c/b\u003e\u003c/p\u003e \u003cp\u003eThis multicenter study acquired rs-fMRI data from 100 patients across two independent clinical centers. Six participants were excluded due to suboptimal data quality related to excessive head motion, resulting in a final analysis of 94 patients (discovery cohort: n\u0026thinsp;=\u0026thinsp;48; external validation cohort: n\u0026thinsp;=\u0026thinsp;46). A diagnostic framework was developed using regional and network markers, including the amplitude of low-frequency fluctuations, regional homogeneity, degree centrality, and functional connectivity. Nine machine learning classifiers were optimized, with the best-performing model tested on the external validation cohort. SHAP analysis quantified circuit contributions to elucidate neurobiological mechanisms.\u003c/p\u003e\u003cp\u003e\u003cb\u003eResults\u003c/b\u003e\u003c/p\u003e \u003cp\u003eL1-regularized logistic regression selected 10 core features, dominated by default mode network (DMN) interactions with cerebellar and subcortical nodes and salience-related local synchrony. The support vector machine emerged as the leading model, achieving an AUC of 0.859 (accuracy: 82.6%; sensitivity: 90.5%) in the external validation cohort. SHAP attribution identified a core neurobiological signature dominated by DMN\u0026ndash;cerebellar coupling (specifically left angular gyrus to right cerebellum), alongside DMN\u0026ndash;subcortical (thalamus and putamen) pathways and insular synchrony.\u003c/p\u003e\u003cp\u003e\u003cb\u003eConclusions\u003c/b\u003e\u003c/p\u003e \u003cp\u003eAn externally validated rs-fMRI framework differentiated MCS from UWS and localized residual consciousness in MCS to preserved DMN-centered corticocerebellar and cortico\u0026ndash;subcortical circuits.\u003c/p\u003e","manuscriptTitle":"Explainable and Externally Validated Resting-State fMRI Machine Learning Reveals Network Mechanisms Supporting Preserved Consciousness: A Cross-Sectional Study Comparing Minimally Conscious State and Unresponsive Wakefulness Syndrome","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-31 17:28:43","doi":"10.21203/rs.3.rs-8959330/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"233838552326390145675076598131594064270","date":"2026-03-30T06:43:53+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"301200879430923112779539048823014375392","date":"2026-03-26T13:26:55+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-03-26T13:10:34+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-03-03T01:09:36+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-02-27T10:58:53+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-02-27T10:57:31+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Imaging","date":"2026-02-24T15:45:45+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"bed9617e-11d0-43e8-9838-59c0e4cbe440","owner":[],"postedDate":"March 31st, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-03-31T17:28:43+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-31 17:28:43","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8959330","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8959330","identity":"rs-8959330","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}