Can artificial intelligence improve the diagnosis and prognosis of disorders of consciousness? A scoping review

Disorders of consciousness (DoC) encompass neurological impairments after severe acquired brain injury (e.g., ischemic/hemorrhagic stroke, traumatic brain injury) and include unresponsive wakefulness syndrome (UWS) and minimally conscious state (MCS, with MCS+ and MCS− subtypes). Diagnostic challenges include detecting covert consciousness (cognitive‑motor dissociation) using neurophysiological/neuroimaging modalities, and precisely differentiating DoC subgroups. Prognostically, limited therapeutic options and heterogeneous etiologies complicate personalized rehabilitation planning. AI (ML/DL) has potential to assist clinical decision‑making by integrating multimodal data for earlier, more accurate diagnosis and prognosis, and tailoring interventions. This scoping review aims to map AI applications for DoC diagnosis and prognosis, identify commonly used technologies, and explore how AI can enhance patient care.

Prior AI applications in related neurological contexts suggest clinical value: AI decision support in dementia care improves early diagnosis and management; AI with non‑invasive neuroimaging aids early autism diagnosis and severity stratification. In DoC, rs‑fMRI and ML have been used to identify neuroimaging biomarkers differentiating consciousness levels, though subgroup differentiation (UWS vs. MCS, MCS+ vs. MCS−) remains difficult. Reviews (e.g., Lee and Laureys, 2024) indicate most AI work in DoC focuses on diagnostic differentiation rather than prognosis, underscoring the need for multimodal integration and interpretability.

Design: Scoping review following SR‑PRISMA guidelines; protocol registered on OSF (DOI: 10.17605/OSF.IO/6D89Z). Databases and dates: PubMed, Embase, Scopus, and Cochrane Library searched starting October 3, 2024, for peer‑reviewed studies published 2000–2024. Search strategy: Common query combining terms for artificial intelligence, machine learning, deep learning with DoC-related diagnoses (coma, vegetative state/UWS, minimally conscious state, post‑traumatic confusional state). Eligibility: Adults (>18 years) with DoC of any etiology; AI/ML/DL applied to diagnostic/prognostic assessment; English language; peer‑reviewed. Exclusions: Theoretical/methodological papers without patient data, animal studies, conference proceedings, reviews, pediatric studies, and case reports. Screening and selection: Two blinded independent reviewers used Rayyan for title/abstract screening and full‑text assessment; disagreements resolved by consensus. Data extraction: Study metadata; sample characteristics; intervention/evaluation details; predictors (neurophysiological, neuroimaging, autonomic, clinical); AI/ML/DL algorithms; performance metrics (accuracy, sensitivity, specificity, AUC); and interpretability techniques.

Search and selection: 49,417 records identified (PubMed, Embase, Web of Science, Cochrane); 17,184 duplicates; 77 non‑English; 31,543 excluded at title/abstract; 613 full‑texts assessed; 21 studies included. Samples: 14,683 patients and 180 healthy controls (in 9 studies); 11,256 DoC unspecified; 367 UWS; 329 MCS; 2,731 coma. Predictors: - Neurophysiological: EEG (PSD bands delta/theta/alpha/beta, burst suppression ratio), qEEG, permutation entropy, dominant frequency, ERPs (MMN), TMS‑EEG; combined EEG/ECG; EEG connectivity measures. - Neuroimaging: rs‑fMRI connectivity; diffusion tractography thalamo‑cortical connectivity; CT features (hemorrhage volume, ventricular involvement). - Autonomic: HRV to emotional stimuli; polysomnography sleep staging (ML classification). - Clinical: CRS‑R total/subscores, CDI index, weighted CRS+, GCS/LCF, CPC, DRS, ERBI, CIRS, medical complications, vital signs, IoT‑based data pipelines. Diagnostic performance highlights: - DeepDOC (3D EfficientNet‑B3) on rs‑fMRI distinguished MCS vs. UWS (AUC 0.927; accuracy 0.861) and detected covert motor dysfunction (AUC 1.0; accuracy 0.909); Grad‑CAM indicated posterior/visual cortex importance. - Ensemble‑of‑SVM with novel EEG connectivity feature CPSDD: accuracy 98.21%, sensitivity 100%, specificity 95.79% for DoC vs. awake. - Sleep staging in DoC via supervised ML: F1 score 0.87 aligning with eye closure patterns; unsupervised clustering revealed organized sleep in MCS vs. fragmented in UWS. - qEEG‑based LDA classified improved vs. non‑improved patients; tailored by etiology (TBI vs. non‑TBI). Prognostic performance highlights: - CNNs on early post‑arrest EEG predicted outcomes: at 12 h, poor outcomes predicted with 58% sensitivity and 0% FPR; good outcomes with 48% sensitivity at 5% FPR; temporal EEG evolution informative. - EEGNet (CNN) predicted survival (test AUC ~0.70±0.04; PPV 0.83±0.03; NPV 0.57±0.04); in clinical gray zone, PPV 0.86; interpretability linked higher neural synchrony/lower complexity to survival. - Bi‑LSTM on temporal EEG features achieved peak AUC‑ROC 0.88 at 66 h; outperformed CNN/RF; good calibration. - XGBoost and RF on multimodal features for post‑anoxic coma achieved cross‑validation scores ~0.34–0.381; convolutional baselines suffered high FPR. - CRS‑R-derived CDI via unsupervised clustering plus LR improved sensitivity/specificity for long‑term recovery (6, 12, 24 months); motor, visual, auditory subscores most predictive; enhanced prognostication in younger/TBI/early assessments. - Clinical ML (Elastic‑Net, OMP, KNN, SVR) integrating medical complexity improved outcome prediction accuracy from 88.6% to 92.6% after adding medical complications (SMOTE, 5‑fold CV). Overall: AI enables multimodal integration for improved diagnostic differentiation (UWS vs. MCS), detection of covert consciousness, and individualized prognostication; interpretability methods (LRP, Grad‑CAM) highlight physiologically plausible regions/features.

Findings address the review questions by demonstrating that AI (ML/DL) can meaningfully assist both diagnosis and prognosis in DoC. Diagnostic gains include robust differentiation of UWS vs. MCS and detection of covert consciousness using rs‑fMRI and EEG, while prognostic models leverage temporal EEG trends, clinical scales (CRS‑R, CPC), and medical complications to predict recovery and survival. Diverse AI technologies are employed: DL excels with high‑dimensional neuroimaging and time‑series EEG, whereas traditional ML models (SVM, RF, LR) perform well on smaller, engineered clinical feature sets and offer better interpretability. Clinically, AI supports multimodal data fusion, objective monitoring (e.g., sleep staging, autonomic markers), and personalized rehabilitation planning by identifying key prognostic features (e.g., CRS‑R subscores, hemorrhage volume, EEG dynamics). However, heterogeneity in etiologies, variability in data acquisition protocols, limited availability of advanced imaging, and inconsistent feature standards impede generalizability and clinical translation. Interpretability remains critical; studies employing LRP/Grad‑CAM and physiologically grounded features improve clinician trust and potential workflow integration. Ethical considerations (consent, privacy, bias) are central for DoC patients who cannot provide consent directly, necessitating transparent, accountable AI systems.

AI applications show strong potential to enhance diagnosis and prognosis in DoC by integrating multimodal clinical, neurophysiological, autonomic, and neuroimaging data, enabling better differentiation of consciousness states (e.g., UWS vs. MCS) and tailoring rehabilitation strategies based on key recovery factors. To advance clinical translation, future work should: establish standardized acquisition/preprocessing protocols; ensure demographic and etiological diversity; prioritize interpretable models; perform prospective, multicenter validation; and address neuroethical issues (consent, privacy, fairness). Integrating AI into clinical workflows (decision support, monitoring, alerts) can personalize care, optimize resource allocation, and improve patient outcomes.

Study-level limitations include small sample sizes, retrospective designs, and heterogeneous etiologies reducing generalizability; variability in EEG configurations and neuroimaging protocols; limited access to advanced imaging (e.g., fMRI); and frequent lack of model interpretability. Review-level limitations include exclusion of non‑English studies and absence of quantitative meta‑analysis, potentially omitting relevant evidence and precluding pooled performance estimates.