Multimodal multicentre investigation of diagnostic and prognostic markers in disorders of consciousness.

Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Brain . 2026 Jan 7;149(4):1381–1395. doi: 10.1093/brain/awaf412 Search in PMC Search in PubMed View in NLM Catalog Add to search Multimodal multicentre investigation of diagnostic and prognostic markers in disorders of consciousness Dragana Manasova Dragana Manasova 1 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France 2 Université Paris Cité, Paris 75006, France Find articles by Dragana Manasova 1, 2, ✉ , Laouen Mayal Louan Belloli Laouen Mayal Louan Belloli 3 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France 4 Laboratorio de Inteligencia Artificial Aplicada, Instituto de Ciencias de la Computación, Universidad de Buenos Aires, Buenos Aires C1053, Argentina 5 Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET), Ministry of Science, Technology and Innovation, Buenos Aires C1053, Argentina Find articles by Laouen Mayal Louan Belloli 3, 4, 5 , Martin Justinus Rosenfelder Martin Justinus Rosenfelder 6 Department of Neurology, University Hospital of the Ludwig-Maximilians-Universität München, Munich 82152, Germany 7 Therapiezentrum Burgau, Hospital for Neurological Rehabilitation, Burgau 89331, Germany 8 Clinical and Biological Psychology, Institute of Psychology and Education, Ulm University, Ulm 89081, Germany Find articles by Martin Justinus Rosenfelder 6, 7, 8 , Lina Willacker Lina Willacker 9 Department of Neurology, University Hospital of the Ludwig-Maximilians-Universität München, Munich 82152, Germany Find articles by Lina Willacker 9 , Emilia Fló Rama Emilia Fló Rama 10 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France Find articles by Emilia Fló Rama 10 , Chiara Valota Chiara Valota 11 Department of Biomedical and Clinical Sciences, University of Milano, Milan 20157, Italy 12 IRCCS Fondazione Don Carlo Gnocchi ONLUS, Milan 20148, Italy Find articles by Chiara Valota 11, 12 , Bertrand Hermann Bertrand Hermann 13 Inserm 1266, Institute of Psychiatry and Neurosciences of Paris, Université Paris Cité, Paris F-75014, France 14 Medical Intensive Care Unit, HEGP Hôpital, Assistance Publique - Hôpitaux de Paris-Centre (APHP-Centre), Paris 75014, France Find articles by Bertrand Hermann 13, 14 , Brigitte Charlotte Kaufmann Brigitte Charlotte Kaufmann 15 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France Find articles by Brigitte Charlotte Kaufmann 15 , Alice Pirastru Alice Pirastru 16 IRCCS Fondazione Don Carlo Gnocchi ONLUS, Milan 20148, Italy Find articles by Alice Pirastru 16 , Chiara Camilla Derchi Chiara Camilla Derchi 17 IRCCS Fondazione Don Carlo Gnocchi ONLUS, Milan 20148, Italy Find articles by Chiara Camilla Derchi 17 , Theresa Raiser Theresa Raiser 18 Department of Neurology, University Hospital of the Ludwig-Maximilians-Universität München, Munich 82152, Germany Find articles by Theresa Raiser 18 , Melanie Valente Melanie Valente 19 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France Find articles by Melanie Valente 19 , Aude Sangare Aude Sangare 20 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France Find articles by Aude Sangare 20 , Başak Türker Başak Türker 21 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France Find articles by Başak Türker 21 , Nadya Pyatigorskaya Nadya Pyatigorskaya 22 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France Find articles by Nadya Pyatigorskaya 22 , Benoît Béranger Benoît Béranger 23 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France Find articles by Benoît Béranger 23 , Michele Colombo Michele Colombo 24 Department of Biomedical and Clinical Sciences, University of Milano, Milan 20157, Italy Find articles by Michele Colombo 24 , Esteban Munoz-Musat Esteban Munoz-Musat 25 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France 26 Centre Mémoire de Ressources et de Recherche, Paris Nord/Université Paris-Cité, Paris 75006, France Find articles by Esteban Munoz-Musat 25, 26 , Anira Escrichs Anira Escrichs 27 Center for Brain and Cognition, Computational Neuroscience Group, Universitat Pompeu Fabra, Barcelona 08005, Spain Find articles by Anira Escrichs 27 , Tiziana Atzori Tiziana Atzori 28 IRCCS Fondazione Don Carlo Gnocchi ONLUS, Milan 20148, Italy Find articles by Tiziana Atzori 28 , Francesca Baglio Francesca Baglio 29 IRCCS Fondazione Don Carlo Gnocchi ONLUS, Milan 20148, Italy Find articles by Francesca Baglio 29 , Constantin Lapa Constantin Lapa 30 Nuclear Medicine, Faculty of Medicine, University of Augsburg, Augsburg 86156, Germany Find articles by Constantin Lapa 30 , Ansgar Berlis Ansgar Berlis 31 Diagnostic and Interventional Neuroradiology, Faculty of Medicine, University of Augsburg, Augsburg 86156, Germany Find articles by Ansgar Berlis 31 , Kristina Krüger Kristina Krüger 32 Diagnostic and Interventional Neuroradiology, Faculty of Medicine, University of Augsburg, Augsburg 86156, Germany Find articles by Kristina Krüger 32 , Tina Luther Tina Luther 33 Department of Neurology, University Hospital of the Ludwig-Maximilians-Universität München, Munich 82152, Germany 34 Therapiezentrum Burgau, Hospital for Neurological Rehabilitation, Burgau 89331, Germany Find articles by Tina Luther 33, 34 , Vincent Perlbarg Vincent Perlbarg 35 BRAINTALE SAS, Paris 75013, France Find articles by Vincent Perlbarg 35 , Gustavo Deco Gustavo Deco 36 Center for Brain and Cognition, Computational Neuroscience Group, Universitat Pompeu Fabra, Barcelona 08005, Spain 37 Institució Catalana de la Recerca I Estudis Avançats (ICREA), Barcelona 08010, Spain Find articles by Gustavo Deco 36, 37 , Yonathan Sanz-Perl Yonathan Sanz-Perl 38 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France 39 Institució Catalana de la Recerca I Estudis Avançats (ICREA), Barcelona 08010, Spain Find articles by Yonathan Sanz-Perl 38, 39 , Enzo Tagliazucchi Enzo Tagliazucchi 40 Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET), Ministry of Science, Technology and Innovation, Buenos Aires C1053, Argentina 41 Latin American Brain Health Institute (BrainLat), Universidad Adolfo Ibáñez, Santiago 7941169, Chile Find articles by Enzo Tagliazucchi 40, 41 , Louis Puybasset Louis Puybasset 42 BRAINTALE SAS, Paris 75013, France 43 GRC 29, AP-HP, DMU DREAM, Department of Anaesthesiology and Critical Care Medicine, Pitié-Salpêtrière Hospital, Sorbonne University, Paris 75013, France Find articles by Louis Puybasset 42, 43 , Benjamin Rohaut Benjamin Rohaut 44 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France 45 AP-HP, Hôpital de la Pitié Salpêtrière, Neuro ICU, DMU Neurosciences, Paris 75013, France Find articles by Benjamin Rohaut 44, 45 , Lionel Naccache Lionel Naccache 46 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France 47 AP-HP, Hôpital Pitié - Salpêtrière, Service de Neurophysiologie Clinique, Paris 75013, France Find articles by Lionel Naccache 46, 47 , Angela Comanducci Angela Comanducci 48 IRCCS Fondazione Don Carlo Gnocchi ONLUS, Milan 20148, Italy Find articles by Angela Comanducci 48 , Anat Arzi Anat Arzi 49 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France 50 Department of Medical Neurobiology, Institute for Medical Research Israel-Canada, Faculty of Medicine, The Hebrew University of Jerusalem, Jerusalem 9112102, Israel 51 Department of Cognitive and Brain Sciences, The Hebrew University of Jerusalem, Jerusalem 9112102, Israel Find articles by Anat Arzi 49, 50, 51 , Mario Rosanova Mario Rosanova 52 Department of Biomedical and Clinical Sciences, University of Milano, Milan 20157, Italy Find articles by Mario Rosanova 52 , Andreas Bender Andreas Bender 53 Department of Neurology, University Hospital of the Ludwig-Maximilians-Universität München, Munich 82152, Germany 54 Therapiezentrum Burgau, Hospital for Neurological Rehabilitation, Burgau 89331, Germany 55 Department of Neurorehabilitation, Medical Faculty, University of Augsburg, Augsburg 86156, Germany Find articles by Andreas Bender 53, 54, 55 , Jacobo Diego Sitt Jacobo Diego Sitt 56 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France Find articles by Jacobo Diego Sitt 56, ✉ Author information Article notes Copyright and License information 1 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France 2 Université Paris Cité, Paris 75006, France 3 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France 4 Laboratorio de Inteligencia Artificial Aplicada, Instituto de Ciencias de la Computación, Universidad de Buenos Aires, Buenos Aires C1053, Argentina 5 Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET), Ministry of Science, Technology and Innovation, Buenos Aires C1053, Argentina 6 Department of Neurology, University Hospital of the Ludwig-Maximilians-Universität München, Munich 82152, Germany 7 Therapiezentrum Burgau, Hospital for Neurological Rehabilitation, Burgau 89331, Germany 8 Clinical and Biological Psychology, Institute of Psychology and Education, Ulm University, Ulm 89081, Germany 9 Department of Neurology, University Hospital of the Ludwig-Maximilians-Universität München, Munich 82152, Germany 10 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France 11 Department of Biomedical and Clinical Sciences, University of Milano, Milan 20157, Italy 12 IRCCS Fondazione Don Carlo Gnocchi ONLUS, Milan 20148, Italy 13 Inserm 1266, Institute of Psychiatry and Neurosciences of Paris, Université Paris Cité, Paris F-75014, France 14 Medical Intensive Care Unit, HEGP Hôpital, Assistance Publique - Hôpitaux de Paris-Centre (APHP-Centre), Paris 75014, France 15 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France 16 IRCCS Fondazione Don Carlo Gnocchi ONLUS, Milan 20148, Italy 17 IRCCS Fondazione Don Carlo Gnocchi ONLUS, Milan 20148, Italy 18 Department of Neurology, University Hospital of the Ludwig-Maximilians-Universität München, Munich 82152, Germany 19 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France 20 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France 21 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France 22 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France 23 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France 24 Department of Biomedical and Clinical Sciences, University of Milano, Milan 20157, Italy 25 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France 26 Centre Mémoire de Ressources et de Recherche, Paris Nord/Université Paris-Cité, Paris 75006, France 27 Center for Brain and Cognition, Computational Neuroscience Group, Universitat Pompeu Fabra, Barcelona 08005, Spain 28 IRCCS Fondazione Don Carlo Gnocchi ONLUS, Milan 20148, Italy 29 IRCCS Fondazione Don Carlo Gnocchi ONLUS, Milan 20148, Italy 30 Nuclear Medicine, Faculty of Medicine, University of Augsburg, Augsburg 86156, Germany 31 Diagnostic and Interventional Neuroradiology, Faculty of Medicine, University of Augsburg, Augsburg 86156, Germany 32 Diagnostic and Interventional Neuroradiology, Faculty of Medicine, University of Augsburg, Augsburg 86156, Germany 33 Department of Neurology, University Hospital of the Ludwig-Maximilians-Universität München, Munich 82152, Germany 34 Therapiezentrum Burgau, Hospital for Neurological Rehabilitation, Burgau 89331, Germany 35 BRAINTALE SAS, Paris 75013, France 36 Center for Brain and Cognition, Computational Neuroscience Group, Universitat Pompeu Fabra, Barcelona 08005, Spain 37 Institució Catalana de la Recerca I Estudis Avançats (ICREA), Barcelona 08010, Spain 38 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France 39 Institució Catalana de la Recerca I Estudis Avançats (ICREA), Barcelona 08010, Spain 40 Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET), Ministry of Science, Technology and Innovation, Buenos Aires C1053, Argentina 41 Latin American Brain Health Institute (BrainLat), Universidad Adolfo Ibáñez, Santiago 7941169, Chile 42 BRAINTALE SAS, Paris 75013, France 43 GRC 29, AP-HP, DMU DREAM, Department of Anaesthesiology and Critical Care Medicine, Pitié-Salpêtrière Hospital, Sorbonne University, Paris 75013, France 44 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France 45 AP-HP, Hôpital de la Pitié Salpêtrière, Neuro ICU, DMU Neurosciences, Paris 75013, France 46 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France 47 AP-HP, Hôpital Pitié - Salpêtrière, Service de Neurophysiologie Clinique, Paris 75013, France 48 IRCCS Fondazione Don Carlo Gnocchi ONLUS, Milan 20148, Italy 49 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France 50 Department of Medical Neurobiology, Institute for Medical Research Israel-Canada, Faculty of Medicine, The Hebrew University of Jerusalem, Jerusalem 9112102, Israel 51 Department of Cognitive and Brain Sciences, The Hebrew University of Jerusalem, Jerusalem 9112102, Israel 52 Department of Biomedical and Clinical Sciences, University of Milano, Milan 20157, Italy 53 Department of Neurology, University Hospital of the Ludwig-Maximilians-Universität München, Munich 82152, Germany 54 Therapiezentrum Burgau, Hospital for Neurological Rehabilitation, Burgau 89331, Germany 55 Department of Neurorehabilitation, Medical Faculty, University of Augsburg, Augsburg 86156, Germany 56 Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France ✉ Correspondence to: Dragana Manasova Paris Brain Institute - Institut du Cerveau Hôpital Pitié, 47 Bd de l'Hôpital, Paris 75013, France E-mail: [email protected] ✉ Correspondence may also be addressed to: Jacobo Diego Sitt E-mail: [email protected] Received 2024 Nov 21; Revised 2025 Jul 22; Accepted 2025 Sep 3; Collection date 2026 Apr. © The Author(s) 2026. Published by Oxford University Press on behalf of the Guarantors of Brain. This is an Open Access article distributed under the terms of the Creative Commons Attribution-NonCommercial License ( https://creativecommons.org/licenses/by-nc/4.0/ ), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited. For commercial re-use, please contact [email protected] for reprints and translation rights for reprints. All other permissions can be obtained through our RightsLink service via the Permissions link on the article page on our site—for further information please contact [email protected]. PMC Copyright notice PMCID: PMC13058464  PMID: 41499248 Abstract Severely brain-injured patients may enter a spectrum of conditions collectively known as disorders of consciousness. This spectrum includes clinical conditions such as unresponsive wakefulness syndrome or minimally conscious state, where the behavioural assessment of consciousness can often be deceptive. To bridge this dissociation, neuroimaging techniques are employed to identify the residual brain functions. Each neuroimaging modality imperfectly captures distinct aspects of brain preservation—functional, anatomical, or both. In this study, we adopt a comprehensive approach by integrating the neurophysiology and neuroimaging modalities available from the standard and advanced clinical assessments through interpretable machine learning. The electrophysiological modalities included high-density EEG (resting state and task), whereas neuroimaging modalities included anatomical and resting-state functional MRI, diffusion MRI and 18 F-fluorodeoxyglucose PET. Our investigation reveals that specific modalities, such as functional assessments, provide comprehensive insights into the currently evaluated state of consciousness, the diagnosis of the patients. Conversely, structural modalities offer valuable information about the patient's evolution within the consciousness spectrum. We validate the proposed analysis with data coming from other centres with different acquisition parameters. Importantly, we demonstrate that model performance improves with an increase in the number of modalities. We observe a higher inter-modality disagreement for minimally conscious state patients and those patients who improve. Lastly, we observe a difference in feature importances between diagnosis and prognosis, with an interaction between modality and anatomical structures: some subcortical markers tend to contribute more to prognosis, while other cortical markers are more informative for diagnosis. This integrative multimodal and machine learning methodology presents a promising avenue for a more nuanced understanding of disorders of consciousness, contributing to enhanced diagnostic precision, prognostic capabilities and the personalization of rehabilitative strategies in clinical practice. Keywords: disorders of consciousness, electrophysiology, neuroimaging, multimodal, machine learning Manasova et al. used advanced neuroimaging, electrophysiology, and machine learning to study disorders of consciousness in severely brain-injured patients. They found that different modalities provide complementary diagnostic and prognostic information, and that combining multiple modalities improves predictive accuracy. Introduction Disorders of consciousness (DoC) encompass a spectrum of conditions resulting from various causes of brain injury. Patients with a DoC have a range of sensorimotor deficits that can, to varying extents, impair both their state of consciousness and their capacity to express it. 1 Owing to an unreliable ability to express consciousness through behavioural responses, there can be a dissociation between unresponsiveness (based on following commands with motor outputs) and unconsciousness. 2 , 3 Furthermore, the assessment of patients with DoC can be limited due to the presence of medical devices (such as mechanical ventilation or tracheostomy tubes), acute pain, or medications that affect arousal. 4 All of these aspects pose challenges to the correct assessment of the patient’s consciousness. The main clinical conditions on the DoC spectrum are unresponsive wakefulness syndrome (UWS) and the minimally conscious state (MCS). UWS patients are behaviourally diagnosed by eyes opening during arousal but with no signs of awareness, 5 whereas patients in an MCS show reproducible, though subtle, behavioural signs of consciousness (visual pursuit or a response to simple commands). 6 Although there is no clear consensus where the spectrum of DoC ends and healthy consciousness begins, 7 generally patients are said to be emergent from MCS (EMCS) when they regain some basic communication capacity or when they are capable of functional object use. 6 This recovery of consciousness can occur at any point in the patient’s clinical evolution⸺from the acute to chronic stages. 7 The current diagnostic gold standard in the field is the Coma Recovery Scale-Revised (CRS-R). 8 Although the systematic and repeated use of this scale decreases the rate of misdiagnosis, 8 the remaining uncertainty due to behavioural and neural disparities has yet to be addressed systematically. Thus, despite extensive standardization of the administration of behavioural scales, 9 current guidelines recommend the use of neuronal or physiological signals across distinct modalities to increase the certainty of achieving a correct assessment of a patient’s state. 1 , 4 , 10-13 Additionally, individual electrophysiology or neuroimaging modalities are limited because they occupy a narrow space in the temporal-spatial resolution plane and are designed to evaluate only specific neural structures or activities. A natural question that arises is to what extent each modality is informative in terms of diagnosis and prognosis. Throughout the years, various studies have focused on single modalities and reported the potential of each in improving diagnostic power. 2 , 14-32 However, clinical teams, especially those in hospitals with established expertise in treating DoC patients, have access to and can combine information from various multimodal tests. 1 , 11 , 33 Recent evidence indicates that multimodal assessment enhances neuroprognostics in clinically unresponsive critical-care patients with brain injury. 33 This suggests that latent integration of information by clinicians contributes to improved decision-making outcomes, thereby making a strong argument for the exploration of automatic fusion approaches using machine learning. This is also emphasized by the international guidelines for the clinical approach to DoC, calling for multimodal assessments, especially due to their heterogeneous pathophysiology. 10 , 13 , 34 On the question of how we can systematically make use of the complementary information contained in the different neurophysiological signals, numerous studies have assessed the possibilities of an integrative neuroimaging approach. 1 , 11 , 35-42 These studies highlight the need to investigate various dimensions of brain preservation [e.g. anatomical MRI (aMRI), functional MRI (fMRI), electrophysiology or brain metabolism] to more accurately assess a patient’s current state and progression. However, to date, there has been no large-scale multicentre study involving commonly used neuroimaging and electrophysiological modalities analysed under the same methodological umbrella to evaluate the differences and complementarity of these modalities in assessing a patient’s current condition and evolution. In this study, we took the multimodal integrative neuroimaging and electrophysiological approach one step further and adopted a multicentric, comprehensive approach by separately analysing and then combining six neuroimaging modalities through interpretable machine learning, to investigate the evaluated consciousness (diagnosis) of patients and future change (prognosis). Materials and methods This project is part of an EU-funded 4-year consortium (PerBrain) involving several institutions, including the Pitié Salpetrière Hospital, the University Hospital of the Ludwig-Maximilians-University of Munich, the Therapiezentrum Burgau (hospital for neurological rehabilitation), the University of Milan, Fondazione Don Carlo Gnocchi, and the Weizmann Institute of Science. The consortium’s aims are explained in Willacker et al . 43 The included patients are from three centres and split into Dataset 1 (France), Dataset 2 (Germany) and Dataset 3 (Italy). Ethics statements This research was approved by the ethical committee of the Pitie-Salpetriere under the French label of ‘routine care research’ (Comité de Protection des Personnes no 2013-A01385-40, Ile de France 1, Paris, France under the code ‘Recherche en soins courants’, protocol numbers NEURO-DoC/HAO-006/20130409 and M-NEURO-DoC/ NCT04534777 ); the ethics committee of the medical faculty of Ludwig-Maximilians-Universität München (protocol numbers 20-634 and 20-635); and the ethical committee section of the IRCCS Fondazione Don Carlo Gnocchi (ethics committee IRCCS Regione Lombardia, protocol number 32/2021/CE_FdG/FC/SA). Written informed consent from patients was obtained either through their legal guardian or, in the absence of one, from the closest relative. Table 1 provides an overview of the number of patients from all three datasets who have a particular modality per prediction category (diagnostic or prognostic). The patient inclusion and behavioural assessment (including the CRS-R 8 and the Glasgow Outcome Scale-Extended 44 ) and methodological details can be found in the Supplementary material . Table 1. Overview of the number of patients from all three datasets having a particular neuroimaging modality per prediction category (diagnostic or prognostic) Modality All patients Diagnostic categories (Current state of the patients) Prognostic categories (Evolution of the patients) UWS MCS Not-improved Improved Dataset 1 (France) EEG-RS 120 63 57 50 39 EEG-LG 290 138 152 117 90 fMRI-RS 44 21 23 20 10 FDG-PET 53 21 32 35 9 dMRI 151 79 72 49 46 aMRI 101 54 47 40 26 Dataset 2 (Germany) EEG-RS 50 30 20 – – EEG-LG 42 24 18 – – aMRI 12 10 2 – – fMRI-RS 7 5 2 – – Dataset 3 (Italy) EEG-RS 25 12 13 – – EEG-LG 17 6 11 – – aMRI 12 3 9 – – fMRI-RS 12 3 9 – Open in a new tab The table presents data obtained through various neuroimaging modalities. In Dataset 1 (France), there were a total of 326 patients for diagnosis, and 232 patients with prognostic data; in Dataset 2 (Germany), there were 54 patients; and in Dataset 3 (Italy), 30 patients. Modalities included EEG-resting state (EEG-RS), EEG Local-Global paradigm (EEG-LG), resting-state functional MRI (fMRI-RS), anatomical MRI (aMRI), diffusion MRI (dMRI) and 18 F-fluorodeoxyglucose PET (FDG-PET). The patient population was further categorized based on diagnostic information, distinguishing between those in a vegetative state/unresponsive wakefulness state (UWS) and those in a minimally conscious state (MCS). Prognostic insights into patient evolution are provided, indicating the number of patients showing improvement and those not showing improvement across different modalities. The counts in each cell represent the corresponding number of patients within the specified modality and diagnostic/prognostic category. Not all the patients have prognostic information; thus, the sum of the prognostic categories is not equal to the counts in the ‘All patients’ column. Information on the patients’ sex, aetiology (traumatic brain injury, anoxic damage, stroke and other causes), age, and whether they were acute or chronic at the time of the tests is given in Supplementary Tables 1–4 . Modalities In this study, the included modalities were high-density EEG-resting state (RS) and a two-level local-global (LG) auditory regularity task, aMRI, resting-state functional MRI (fMRI-RS), diffusion MRI (dMRI) and 18 F-fluorodeoxyglucose PET (FDG-PET) ( Table 2 ). The acquisition protocols and parameters, as well as the preprocessing details and analyses of the markers, are provided in the Supplementary material . The markers we extracted from the modalities later used in the machine learning models 45 are given in Table 2 . Table 2. Overview of multimodal neuroimaging markers used as features Modality type Modality Paradigm Metrics Dynamic, functional EEG Resting state Spectral, information theory, connectivity markers Dynamic, functional EEG Local-global auditory task Spectral, information theory, connectivity and evoked markers Dynamic, functional fMRI Resting state Cortical and subcortical functional connectivity Static, functional FDG-PET Resting state Cortical & subcortical metabolic activity Static, anatomical dMRI Resting state White matter tract fractional anisotropy and mean diffusivity Static, anatomical aMRI Resting state Cortical thickness, subcortical volume Open in a new tab This table provides a comprehensive overview of the neuroimaging and electrophysiology modalities used in this work, encompassing both dynamic or static and anatomical or functional properties. Each modality is detailed with its associated paradigm and metric. Dynamic modalities, which have multiple time points, include EEG in the resting state and task local-global paradigms, as well as fMRI in the resting state. Static modalities include FDG-PET, dMRI, and aMRI. The second type refers to whether the modality captures anatomical properties (aMRI and dMRI) or functional ones (EEG, fMRI, FDG-PET). Each modality is associated with specific metrics such as cortical and subcortical functional connectivity, metabolic activity, white matter tract properties, and structural measures like cortical thickness and subcortical volume. aMRI = anatomical MRI; dMRI = diffusion MRI; fMRI = functional MRI; FDG-PET = 18 F-fluorodeoxyglucose PET. The main methodological outline of the paper is given in Fig. 1 . We included two prediction targets: (i) diagnosis; and (ii) prognosis of the patients. We ran models per modality (referred to as unimodal) and then combined the predictions, yielding a multimodal analysis. The acquisition details, preprocessing, marker extraction, machine learning methods and statistical analysis are described in detail in the Supplementary material . Figure 1. Open in a new tab Multimodal assessment methods of diagnostic and prognostic categories of patients with DoC. ( A ) Patients underwent multiple CRS-R assessments during their hospital stay and the best assessment in that range of a week (typically) was taken as the gold-standard clinical diagnosis. The diagnostic clinical categories of patients included in the prediction are UWS and MCS. The prognostic categories are improved or not improved (explained in the ‘Materials and methods’ section). ( B ) For both the diagnostic and prognostic classification, we ran unimodal RFC to obtain probabilistic estimates of each patient belonging to one or another category. The probabilistic outputs are then combined using a second-level RFC either alone or in combination with the aetiologies and demographic information. Missing values are substituted with −1 (a data imputation approach). The final output is a probability of belonging to either a diagnostic or prognostic category. aMRI = anatomical MRI; CRS-R = Coma Recovery Scale-Revised; dMRI = diffusion MRI; DoC = disorders of consciousness; EMCS = emergent minimally conscious state; FDG-PET = 18 F-fluorodeoxyglucose PET; fMRI = functional MRI; GOSE = Glasgow Outcome Scale-Extended; LG = Local Global paradigm; MCS = minimally conscious state; RFC = Random Forest classifiers; RS = resting state; UWS = unresponsive wakefulness syndrome. Results Neuroimaging modalities carry independent diagnostic and prognostic information In the diagnostic classification ( Fig. 2A ), the highest balanced accuracy was observed for PET (0.73 ± 0.13), followed by dMRI (0.69 ± 0.08), EEG-LG (0.69 ± 0.06), EEG-RS (0.66 ± 0.09), fMRI-RS (0.63 ± 0.15), and the lowest for aMRI (0.51 ± 0.1). For the prognostic classification ( Fig. 2B ), the highest balanced accuracy was achieved by dMRI (0.74 ± 0.09), followed by fMRI-RS (0.63 ± 0.18), aMRI (0.58 ± 0.11), EEG-LG (0.55 ± 0.06), PET (0.5 ± 0.16), and the lowest for EEG-RS (0.49 ± 0.1). For both diagnostic and prognostic classification, we calculated the quality of the difference between the model distributions and the surrogate distributions ( Supplementary Fig. 2 ). Figure 2. Open in a new tab Multivariate unimodal models’ accuracy differs depending on the classification target that can be diagnostic or prognostic. ( A ) Unimodal Random Forest classifiers for the six modalities give a diagnostic classification accuracy which is given next to the distributions (median ± standard deviation). The diagnostic classification patient categories are patients in UWS or MCS. ( B ) The same as in A but for the prognostic categories (improved and not improved). ( C and D ) The same as A but for the diagnostic classification trained on Dataset 1: France and tested on the available modalities from Dataset 2: Germany and Dataset 3: Italy. aMRI = anatomical MRI; dMRI = diffusion MRI; fMRI = functional MRI; LG = Local Global paradigm; MCS = minimally conscious state; RS = resting state; UWS = unresponsive wakefulness syndrome. EEG recordings, whether RS or LG, exhibited a classification accuracy close to 0.7 for diagnosis but dropped near the chance level for prognosis. Conversely, aMRI showed an increase from chance level for diagnosis to 0.58 for prognosis, while PET exhibited the opposite trend, becoming non-informative in more of the cross-validated splits for prognosis (in other words, one part of the distribution of the PET prognostic results was at chance level). dMRI and fMRI-RS remained relevant for both diagnosis and prognosis, with dMRI gaining 0.05 points for prognosis. Using two independent datasets (Dataset 2 from Germany and Dataset 3 from Italy), we examined the generalization of diagnostic prediction in four modalities. The balanced accuracy aligned with the training set for fMRI-RS (0.7 ± 0.12) and EEG-LG (0.62 ± 0.03) in Dataset 2 and for EEG-LG (0.75 ± 0.05), EEG-RS (0.64 ± 0.04) and aMRI (0.58 ± 0.12) in Dataset 3 (Italy). The aMRI models for diagnosis for Dataset 1 were at chance level, which implied that we could not test for their generalization. The models that did not generalize above chance level are fMRI-RS (0.44 ± 0.08 median balanced accuracy) in Dataset 3 (Italy) and EEG RS (0.49 ± 0.02 median balanced accuracy) in Dataset 2. Although it was not the focus of this study, we ran additional analyses splitting the unimodal results from Dataset 1 (Paris) per MCS subgroups (MCS− and MCS+) ( Supplementary Fig. 3 ). We observed that the results were stable, with some variation, whereby dMRI and EEG-LG showed better balanced accuracy when comparing UWS and MCS+, and EEG-RS and aMRI exhibited higher performance when contrasting UWS and MCS−. We ran similar splits for three aetiologies: traumatic brain injury, anoxia and other (where other was a combination of stroke and other combinations or occurrences of aetiologies). The diagnostic per-aetiology splits are given in Supplementary Fig. 4 and the prognostic ones in Supplementary Fig. 5 . Owing to the low number of patients in each subcategory per modality being variable and, in some cases, too low, we refrained from further interpretations but provide the supplementary analysis in case it is helpful to future studies. Pairwise disagreements of unimodal models are higher in MCS and improved patients When examining unimodal predictions per patient, we observed a discernible difference in the extent of disagreements across patient groups. Notably, pairwise disagreements were markedly higher in patients classified as being in an MCS compared to those in a UWS ( Fig. 3B ). The trend in diagnosis was primarily driven by combinations involving aMRI, PET, fMRI and EEG-LG, leaving EEG-RS to be more similar to the other modalities. A similar trend, although less pronounced, was evident when comparing patients who showed improvement versus those who did not ( Fig. 3C ), where the modalities contributing significantly to these disagreements were dMRI and the EEG paradigms. Figure 3. Open in a new tab Pairwise disagreements of the classification probabilities are higher for patients in the minimally conscious state and patients who show an improvement in Dataset 1 (France) . ( A ) The pairwise disagreement is calculated per patient per pair of modalities as the absolute difference in the classification probabilities (probability of being in Group 1 versus Group 2, either for the diagnostic or the prognostic groups) between two modalities. ( B ) Pairwise disagreements of the classification probabilities are more common in MCS patients. Results are displayed separately for the two diagnostic groups (UWS, dark red; MCS, light red). ( C ) Pairwise disagreements are higher in improved patients than in not improved patients with a difference statistically less strong than the one of the diagnostic groups. Pairwise disagreements of the classification probabilities are more common in improved patients. Results are displayed separately for the two diagnostic groups (Not improved, dark blue; Improved, light blue). The distributions of the pairwise disagreements are tested using a Wilcoxon signed-rank test to see if two paired samples are from the same distribution. The stars above the distributions denote the significance in the colour related to the diagnostic group (* P < 0.05; ** P < 0.01; *** P < 0.001; **** P < 0.0001). The pairwise disagreements that have a grey background are those that include a model that was at chance level in Fig. 2A or B . aMRI = anatomical MRI; dMRI = diffusion MRI; fMRI = functional MRI; LG = Local Global paradigm; RS = resting state; MCS = minimally conscious state; UWS = unresponsive wakefulness syndrome. Feature importance differs in diagnostic and prognostic models The fluctuations in diagnostic and prognostic accuracy across individual modalities required exploration of the most influential features for prediction, examining whether these varied when assessing the patient's current state or their outcome. To investigate this, we ranked the features per modality based on their average importance scores across all model iterations. Figure 4A, D and F depict the feature importance scores for fMRI, PET and aMRI, showcasing cortical and subcortical regions; whereas Fig. 4B, C and E show the mean feature importance combined into groups for the EEG-RS and -LG and dMRI scans. The diagnostic aMRI models, together with the prognostic EEG-RS and PET models, were all at chance level; thus, we did not analyse their feature importance scores. Figure 4. Open in a new tab Reordering of the feature importance (FI) distributions per group of features for diagnosis and prognosis in Dataset 1 (France) . ( A ) Feature importance of the diagnostic and prognostic prediction using the fMRI-RS scans. The swarm plot shows the feature importance split into within subcortical functional connectivity, within cortical connectivity subdivided into the seven cortical networks, and subcortical to cortical functional connectivity. The brain plots show the feature importance of the within cortical and within subcortical functional connectivity per region of interest. ( B ) Feature importance of the diagnostic prediction using the EEG-RS recordings. The bar plot shows the feature importance split into conceptual families: connectivity (wSMI), information theory (Kolmogorov complexity and permutation entropy), low spectral (delta, theta and alpha frequency bands), high spectral (beta and gamma bands), and other spectral marker summaries. ( C ) Feature importance of the diagnostic and prognostic prediction using the dMRI scan. The swarm plot shows the feature importance split into two measures, fractional anisotropy (FA) and mean diffusivity (MD), which are subdivided into global brain-wide measures and families of tracts: projection fibres, brainstem, commissural fibres and associative fibres. ( D ) Feature importance of the diagnostic prediction using the FDG-PET scan. On the brain plots, the feature importance of the metabolic activity per cortical or subcortical region of interest are shown. The bar plot shows the feature importance split into cortical networks and the subcortical regions, as well as the importance of the half-brain (left or right hemisphere) metabolic activity. ( E ) Feature importance of the diagnostic and prognostic prediction using the LG task-based EEG recordings. The bar plot shows the feature importance split into the same conceptual families as the EEG-RS with the addition of the evoked markers coming from the task-based paradigm. ( F ) Feature importance of the prognostic prediction using the aMRI scan. ( G ) Average AUC value per cortical or subcortical regions of interest (first three plots) or fibre tracts (fourth plot) for the neuroimaging modalities, split per diagnosis and prognosis. The positive classes are MCS and Improved patients, if the AUC value is from 0.5 to 1, the given marker is higher in MCS or Improved patients. If the AUC value is from 0 to 0.5, the given marker is higher in the UWS or not-Improved group of patients. The feature importance of cortical thickness and subcortical volume are shown on the brain plots. All of the bar plots, including the brain plots from four different views, are given in Supplementary Figs 10–12 . AUC = area under the curve; aMRI = anatomical MRI; dMRI = diffusion MRI; FDG-PET = 18 F-fluorodeoxyglucose PET; fMRI = functional MRI; LG = Local Global paradigm; RS = resting state; MCS = minimally conscious state; UWS = unresponsive wakefulness syndrome; wSMI = weighted symbolic mutual information. Examining spectral subcategories of low bands (delta, theta, alpha) and high bands (beta and gamma), we found that, in both paradigms, for diagnosis, low-frequency-based features were most relevant ( Fig. 4B and E and Supplementary Fig. 16G and I ). In diagnostic EEG-LG, connectivity-derived features were also highly relevant, while high-frequency-based features were less informative. Conversely, for EEG-LG prognostic prediction ( Fig. 4E and Supplementary Fig. 16J ), high-frequency features became the most informative. However, the evoked features remained the least important for both diagnosis and prognosis. In the aMRI prognostic models, the most relevant features were the subcortical volume features, followed by salience and visual network cortical thicknesses, with the features from the default mode network (DMN) being the lowest scoring. In the fMRI-RS functional connectivity analysis, subcortical regions to cortical network connectivity features were most informative for diagnostic classification, followed by within-subcortical and cortical-to-cortical region of interest connectivities ( Fig. 4A ). Conversely, for prognosis ( Fig. 4A ), the somatomotor cortical network gained importance, accompanied by an increased relevance of the limbic and frontoparietal networks. The visual and salience networks and subcortical to cortical functional connectivity decreased in significance. In diagnostic FDG-PET analysis ( Fig. 4D ), mean metabolic activity per left or right hemisphere emerged as the most informative feature, followed by cortical networks like the somatomotor and visual networks, while subcortical areas were less informative. In dMRI, the most important feature distinguishing UWS from MCS patients was the combined fractional anisotropy (FA global) ( Fig. 4C ). This was followed by the right superior fronto-occipital fasciculus, right posterior limb of the internal capsule, and left and right corona radiata. When grouping tracts, projection fibres and brainstem fibres were most informative for diagnosis based on FA ( Fig. 4C ). For prognosis, the mean diffusivity of commissural fibres rose in importance ( Fig. 4C ), and the brainstem tracts measured by FA remained among the most informative. Furthermore, there appeared to be an interaction between modality and cortical and subcortical markers, with some subcortical markers being more informative for prognosis and cortical markers contributing more to diagnosis ( Fig. 4G ). In addition, we calculated the univariate area under the curve (AUC) values per feature ( Supplementary Figs 8–13 ), both for diagnosis and prognosis. We found a non-linear relationship between the mean feature importance scores and the feature AUC values ( Supplementary Fig. 19 ). Multimodal integration improves predictive accuracy In this section, we address two questions: (i) whether the model accuracy improves with an increase in the number of modalities; and (ii) whether extended models perform better than basic ones. In the case of diagnostic prediction ( Fig. 5A ), we observed an increase in balanced accuracy for both basic and extended models. The basic model started at chance level and progressively improved to achieve an accuracy above 0.83. For prognosis ( Fig. 5B ), there was an upward trend in accuracy, with a notable deviation when patients had four modalities, leading to a drop in accuracy, particularly for the basic model. In most cases, the extended model demonstrated superior performance, indicating non-redundant information derived from demographic details and aetiological divisions. Figure 5. Open in a new tab Increasing trends in the multimodal model balanced accuracy for the basic models (only neural modalities) and the extended models (neural modalities plus information on the patient aetiologies and demographics) of Dataset 1 (France) . ( A ) Balanced accuracy for the diagnostic models (basic and extended) for patients with 1–5 different neuroimaging modalities ( x -axis). ( B ) Same as A , apart from the prognosis. The error bars represent the first (Q1) and third (Q3) quartile intervals of the distributions. On the x -axis, the top row represents the number of modalities across the function, and the second row represents the number of patients with the given number of modalities. The stars represent significance following Mann–Whitney U-tests (Bonferroni corrected) between the basic and extended models (* P < 0.05, ** P < 0.01, *** P < 0.001, **** P < 0.0001). n = number of elements in the given distribution. When looking into the trends for the models using a Spearman correlation test, for the diagnostic models, we saw an increase in balanced accuracy for the basic model r (2294) = 0.49, P < 0.0001, 95% confidence interval (CI): 0.46–0.52; and for the extended model r (2294) = 0.399, P < 0.0001, 95% CI: 0.36–0.43. For the prognosis, the positive correlation of the balanced accuracy with the number of modalities was less strong r (2037) = 0.105, P < 0.0001, 95% CI: 0.06–0.15, and increased for the extended model to r (2037) = 0.335, P < 0.0001, 95% CI: 0.3–0.37. The statistical difference between the basic and extended model in diagnosis was for n = 1 modalities [U(500,500) = 55562, P < 0.0001], and n = 4 [U(495,495) = 98743, P < 0.0001], whereas in prognosis was for n = 2 [U(500,500) = 107557, P = 0.0007], n = 4 [U(465,465) = 68766, P < 0.0001] and n = 5 [U(74,74) = 1748.5, P = 0.0002] modalities. For the diagnosis for n = 2 [U(500,500) = 124698, P = 1], n = 3 [U(500,500) = 130672, P = 1] and n = 5 [U(301,301) = 42645, P = 0.92], the basic and extended model results were not statistically different; the same was true for n = 1 [U(500,500) = 116245, P = 0.28] and n = 3 [U(500,500) = 124105, P = 1] for prognosis. It is worth noting that the number of patients with five or more modalities was low. When removing the patients with five modalities, the increasing trend remained; however, the correlation was less strong [diagnostic basic model r (1993) = 0.451, P < 0.0001, 95% CI: 0.42–0.49; diagnostic extended model r (1993) = 0.327, P < 0.0001, 95% CI: 0.29–0.37; prognostic basic model r (1963) = 0.09, P = 0.0001, 95% CI: 0.05–0.13; prognostic extended model r (1963) = 0.302, P < 0.0001, 95% CI: 0.26–0.34]. Discussion Interpretable modelling approach for sparse multimodal neuroimaging datasets The assessment of DoC presents significant clinical challenges due to the potential dissociation between behavioural responsiveness and consciousness, as well as the limitations of behavioural scales in the presence of confounding factors. While individual neuroimaging and electrophysiological modalities have shown potential in improving diagnostic and prognostic accuracy, they are inherently limited in scope. Multimodal approaches, which integrate information across various modalities, hold promise for addressing these limitations, but there has been a lack of large-scale, multicentric studies systematically evaluating their complementarity and effectiveness. In this study, we adopted a comprehensive, multicentric approach to overcome these challenges. We integrated six neuroimaging modalities using interpretable machine learning methods. To handle the inherent challenges of sparse and heterogeneous multimodal datasets, we proposed a two-model stacking approach. This methodology effectively addressed issues such as the low patient-to-feature ratio, modality-specific data sparsity, and the heterogeneity of modality combinations, enabling a robust analysis of both diagnostic and prognostic dimensions of DoC. Differentiating unimodal classification scores for a patient’s current state and its evolution Our analysis of neuroimaging and electrophysiological modalities for patients with DoC revealed intriguing modality differences in diagnostic and prognostic accuracy. Modalities that capture the structural preservation of the brain and its networks (aMRI and dMRI) become more relevant in the evolution of a patient’s state, and hence for prognosis. Conversely, modalities that capture electrical activity (EEG) and metabolic activity (PET) are mostly relevant to diagnosis. This shift in accuracy rankings for modalities between diagnostic and prognostic classifications underscores their complementarity. PET displayed the highest accuracy in discriminating between UWS and MCS patients but carried little prognostic information in half of the cross-validated splits ( Fig. 2A and B ). The diagnostic results aligned with previous studies showing that metabolic data distinguish MCS from UWS patients, 35 , 41 , 46 , 47 with one study reporting that the gradient increase continues to EMCS patients and healthy controls. 36 Furthermore, FDG-PET demonstrated better suitability in discriminating DoC diagnoses compared to MRI-derived measures, including active fMRI (where PET had higher sensitivity for identifying MCS patients), 35 aMRI and fMRI-RS. 47 In our work, we observed a drop in the prognostic accuracy of PET compared to diagnostic; however, the distribution was large ( Fig. 2A and B ), indicating that certain cross-validated splits are better predicted whereas other splits show opposing trends to their training subset. This large variation in accuracy could arise due to the large imbalance in patients who improve ( n = 9) versus those who do not ( n = 35) ( Table 1 ). One study from the literature reported a drop in the patient recovery value of PET compared to the diagnostic, albeit still as high as 74%, 35 whereas another study, in a sample of 20 patients, did not show a prognostic value. 48 Comparing the PET with the EEG models, one study showed a higher sensitivity of EEG models compared to FDG-PET, although the AUC of the diagnostic prediction did not differ significantly. 41 In our study, we used the same RS and LG paradigm EEG markers as those reported in two previous studies. 18 , 27 There is a partial overlap in the data with previous studies, and the results are in accordance with our study. Additionally, after collecting the markers into large groups and contrasting their diagnostic and prognostic accuracies, we observed a drop in the accuracy of both EEG paradigms for the evolution of patient states ( Fig. 2A and B ) (in contrast to higher accuracy in the diagnostic models), as previously shown in other studies. 24 , 49 On the contrary, other studies were able to show both the diagnostic and prognostic potential of EEG. 23 , 39 , 40 , 50 In our work, we observed a diagnostic accuracy of aMRI close to chance levels ( Fig. 2A ). One study found similar diagnostic results (balanced accuracy ranging from 0.45 to 0.63). 31 However, Annen et al . 32 showed a high diagnostic area under the receiver operating curve of 96% using grey matter and white matter volume, a prediction comparable to that from FDG-PET. Importantly, we used a different implementation of the cortical thickness and subcortical volume estimates that was created specifically for clinical data with various resolutions and originating from different neuroimaging centres. 51-54 The differences in our findings compared with those reported by Annen et al . 32 may have arisen due to the different methodologies and diverse cohorts, leading to questions that should be answered in future work, comparing both FreeSurfer implementations across different aetiologies (for example, traumatic versus anoxic). In comparison to the diagnostic models, we saw an increase in accuracy when looking into prognosis ( Fig. 2A and B ). When there was physical damage to tissue that could be quantified with neuroimaging, regeneration was slow, and this could dictate patient prognosis. A study by van der Vliet et al . 55 showed that even patients with severe initial deficits can reach favourable outcomes but require longer time constants for recovery, consistent with the importance of anatomical preservation for prognosis. In future studies, it can be tested whether anatomical preservation scales with the time for improvement or recovery in a linear way. The differential relevance of some subcortical markers for prognosis and other cortical markers for diagnosis ( Fig. 4G ) may reflect underlying differences in neuroplasticity mechanisms and their timescales. Importantly, cortical plasticity has been more extensively studied, 56 , 57 in contrast to subcortical plasticity. 56 A few studies suggest that subcortical structures are not passive relays but play a central role in cortical plasticity and cross-modal functional reorganization. 57-59 Thus, anatomical preservation, especially subcortical, could be more indicative of improvement due to its influence on cortical plasticity processes. A follow-up study would be to test the functional preservation at different time points post-injury of patients that show higher subcortical anatomical preservation. For fMRI-RS, we saw a similar median balanced accuracy for diagnosis and prognosis but a larger distribution for prognostic prediction, similar to the results for PET, suggesting an influence depending on the random cross-validated splits. However, the ordering of the modalities’ performances was different, making fMRI-RS one of the most informative modalities for patient state evolution. Previous work has shown the potential of fMRI-RS to discriminate between patients and controls 17 and between UWS and MCS patients. 21 , 29 , 39 The higher accuracy reported in the literature compared to our results could be attributed to methodological differences between seed-based and atlas-based parcellation of RS networks, or the fact that one of the studies 21 used data only from patients for whom the clinical diagnosis based on CRS-R was congruent with PET scans. Two studies have tested the outcome prediction of DoC patients at 3 months with an accuracy range of 0.69–0.78, 40 and 0.81, 22 but no predictive value at 12 months was observed. 40 Given that we looked at a different prognostic metrics, the results are not directly comparable, but there is consistent evidence that fMRI activity does contain prognostically relevant information for DoC patients. Previous diagnostic studies using dMRI have reported accuracies as high as 0.95 60 and in the range 0.81–0.84 using a multivariate searchlight analysis of whole-brain thalamo-cortical tracts. 61 In prognostic studies of cardiac arrest patients, FA values were shown to reach values of 0.95 sensitivity and 1 specificity, 62 0.98 AUC in a larger follow-up study, 28 and 0.93 AUC 1-year prognostic value of global deep white matter metrics in TBI patients. 63 The consistency of our results with previously reported findings emphasizes the importance and potential of using dMRI to aid the diagnostic and prognostic assessment of patients with DoC. Generalization tests across independent datasets demonstrated varying performance ( Fig. 2C and D ). The modalities that could not be generalized include fMRI-RS from Dataset 3. This discrepancy may be attributed to the heterogeneity in the acquisition parameters compared to the training set (see Supplementary material , ‘Methods’ section). Task EEG (EEG-LG) outperformed EEG-RS across all centres, with a notably reduced effect in Dataset 2 from Germany. The higher performance of EEG-LG-based models highlights the critical role of active paradigms 1 in assessing patients with DoC, as these paradigms likely enhance and regulate patients’ attentional states. In contrast, resting-state paradigms may be less robust to cross-centre variability due to their dependence on intrinsic brain activity, which is more susceptible to external and patient-specific factors. These findings, taken together, underscore the importance of accounting for modality-specific and centre-related acquisition parameters to improve model generalizability across centres. Pairwise disagreements between modalities Studying pairwise disagreements across modalities is important, as it can point to cases for whom a dissociation can elucidate the potential for recovery. We observed more pairwise disagreements for MCS patients and for those that improved ( Fig. 3 ), suggesting that some signals may capture a more positive clinical picture, while others do not. The sources of these disagreements can be neural or non-neural. Neural examples include the case of a functional hemispherectomy, when a patient showed almost no metabolic activity in the left hemisphere with preserved white matter tracts, 64 or islands of preserved cortical activity that are posited to exist in this group of patients. 65 Furthermore, UWS patients with unfavourable EEG features have shown an increase in fMRI between-network connectivity and a decrease in DMN within-network connectivity (but not significant). 39 Although EEG and PET have been shown to be highly correlated, EEG connectivity patterns differed in PET-negative and PET-positive patients. 24 Another study found a diagnostic difference (between healthy controls and DoC patients) in metabolic activity and mixed results in positive and negative DMN connectivity, but no significant results in grey matter volume. 36 A disagreement between metabolic activity and grey matter has been found in the left-sided language network of MCS− and MCS+ patients. 37 The first exhibited lower metabolic values in the left middle temporal cortex and a metabolic functional disconnection between the left angular gyrus and the left prefrontal cortex. The authors concluded that brain function and not grey matter structure supports the clinical signs of language processing. 37 In some patients, only the dMRI images showed a consistent loss of white matter compared to the seemingly unchanged appearance of structural images. 66 Furthermore, the reliability of discriminating between MCS and UWS patients is often compromised due to the limited sensitivity of scalp EEG, as demonstrated by instances where pathological brain activity masks normal neuronal patterns in awake individuals, suggesting the potential for complex dissociations in severe cases of brain injury. 10 The proposal to rename the MCS to a cortically mediated state 67 underlines the fact that the MCS encompasses a broad and heterogeneous range of conditions. This spectrum includes unconscious patients who exhibit residual cortical activity leading to observable behaviour, and conscious patients who, despite possibly being self-aware, are hindered by executive deficits that prevent them from effectively using a communication code or responding functionally to commands. 67 , 68 This may explain why there are more pairwise disagreements in the MCS compared to the UWS, and that some specific modalities might not capture the heterogeneity. A complementary perspective is that the differences in prediction can also stem from aetiology specificities, such as EEG alpha power, which has been shown to be suppressed in severely post-anoxic patients and does not differ between patient groups with other aetiologies. 30 , 69 All these examples corroborate the fact that pairwise disagreements from a neural origin are common, and their hierarchical importance in diagnosis and prognosis needs to be further studied. One non-neural source of disagreement was data quality, which, even with stringent exclusion criteria, was lower in the DoC patient group, possibly affecting analyses down the line. Furthermore, the state of the patients fluctuates across various time scales, whereas in our case, we worked with one diagnosis per patient, which could be the source of disagreement between the diagnostic models. This would not be the case for prognostic prediction. A way to surpass this would be to look at the variability of multiple CRS-R tests and check the disagreement in light of the patient’s clinical fluctuations. Differing importance of feature groups within modalities, for diagnostic and prognostic prediction EEG Previous results have shown that the most informative EEG features to differentiate between UWS and MCS patients are absolute alpha power, permutation entropy, Kolmogorov complexity, and a connectivity measure in the theta band (weighted symbolic mutual information, wSMI). 18 , 27 In our work, we found that in EEG, the spectral feature groups were the most informative ( Fig. 4B and E and Supplementary Fig. 16G, I and J ), with low frequencies being important for diagnosis and high frequencies for prognosis. This finding can be related to the mesocircuit hypothesis that provides a framework for understanding the recovery of consciousness after severe brain injuries by focusing on the interconnected roles of cortical and subcortical structures. 7 , 70 Specifically, the ‘ABCD’ model of neuronal recovery proposes that sequential changes in EEG power spectra can be categorized into four broad stages, each reflecting varying degrees of thalamocortical deafferentation severity. 7 , 70 The importance of high beta and gamma frequencies (AUC values in Supplementary Figs 8 and 9 ) in the DoC patients could further help in distinguishing the MCS patients that belong to group C of the ‘ABCD’ model of corticothalamic dynamics. However, for prognosis, the AUC of the various high-frequency features was higher for patients who did not improve ( Supplementary Figs 8 and 9 ). In prognostic EEG studies of prolonged DoC, the presence of dominant delta frequencies and reduced EEG amplitudes was related to worse outcomes, whereas the dominance of alpha frequencies, preserved EEG reactivity, and an increase in the dominant frequency were associated with improvement. 10 , 23 On the contrary, in one study, higher delta power was associated with improved outcomes for patients transitioning from UWS to MCS, 48 which we also found in the EEG-LG results ( Supplementary Fig. 9 ). These discrepancies can come from aetiology-dependent differences such as the slowing of EEG being relevant for prognosis in toxic encephalopathies, a transient increase in slow waves or suppression of sensory stimuli in patients with a traumatic brain injury, contrary to the increase in gamma and alpha frequencies in acute patients with a subarachnoid haemorrhage. 10 We observed that connectivity metrics from LG had a stronger weight compared to RS ( Fig. 4B and E and Supplementary Figs 8 and 9 ). Previous work has demonstrated the importance of network metrics over frequency power, 24 the relevance of coherence across various regions and frequency bands for improving UWS patients, 23 and a stronger delta network connectivity in patients with negative outcomes. 24 On the contrary, one study found no network features related to outcome at 3 or 6 months post-injury, 49 with only relative alpha power improving prediction accuracy at 3 months in contrast to prediction using only clinical features. This is consistent with our findings, where univariate AUC values of the connectivity features were at chance level for prognosis, both for EEG-RS and -LG. When using oddball auditory perception paradigms, event-related potentials such as mismatch negativity and P300 have been reported to have low sensitivity in MCS patients. 10 , 71 Accordingly, in our work, most of the evoked features were the least informative ones ( Fig. 4E ). Neuroimaging Multiple studies using PET or fMRI-RS have reported a brain-wide network difference between UWS and MCS patients, 21 , 41 , 46 with some reporting that the left hemisphere is more impaired in UWS. 31 , 47 In our work, in most neuroimaging modalities, we observed a distributed brain-wide feature importance ( Fig. 4A, D and F , brain plots), with the exception of fMRI-RS prognostic feature importance. However, the DMN has mostly been shown to differ among diagnostic groups, to have lower activity in UWS compared to MCS, 15 , 17 , 22 , 47 both at enrollment and at discharge from the intensive care unit, 39 and to relate to recovery outcome. 20 , 22 , 25 , 26 , 47 Structural information, such as grey matter volume 32 and structural integrity, 31 has also been reported to be highest in DMN regions. In our prognostic results, the DMN features were moderately informative compared to the other networks in the aMRI and fMRI scans ( Fig. 4A and F and Supplementary Fig. 16B and D ). In the fMRI scans, where both diagnostic and prognostic models were comparable, DMN features were more informative for patient prognosis. A few studies have reported the strongest metabolic activity reduction in frontoparietal areas 46 , 47 , 72 or specifically in the medial prefrontal cortex (part of DMN) or lateral parietal cortex. 22 The primary and associative somatosensory areas have been associated with diagnosis, 46 with one FDG-PET study observing this in the best-preserved hemisphere. 68 A few studies have reported specific diagnostic differences only for the auditory network in fMRI-RS 17 , 21 and glucose metabolism to be higher in MCS than in UWS. 48 In PET and fMRI-RS, we also observed the higher relevance of somatomotor and visual networks. In contrast, one study reported that higher-order networks (DMN, salience, dorsal attention network, left and right fronto-parietal network and temporal network) have better diagnostic accuracy than low-order networks (sensorimotor, auditory and visual networks) as derived by their structural integrity, 31 but we could not compare these findings to our results because the aMRI model was at chance level. We found that the subcortical areas are more important for prognosis than diagnosis; however, this difference has not been the focus of neuroimaging investigations. Various subcortical areas have been reported to differ between diagnostic groups, such as lower metabolic activity in the brainstem, 32 , 46 thalamus, 32 , 46 and the caudate and para-hippocampal areas. 32 Thalamic white matter integrity is also affected in patients, 60 along with the pathway linking the posterior cingulate cortex/precuneus with the thalamus, as evidenced by their mean FA values 73 ; whereas brainstem white matter tract preservation has only been observed in ischaemic-hypoxic patients, 66 and no mean diffusivity brainstem differences have been identified between MCS and UWS. 60 Our results suggest that the preservation of subcortical structures (potentially reflecting the integrity of large-scale arousal and integration networks, as well as their aforementioned neuroplasticity) may be a key factor supporting long-term recovery, highlighting its relevance to both outcome prediction and potential therapeutic targets (as previously investigated). 74 , 75 It is important to note that the signal-to-noise ratio when imaging subcortical areas can be lower than that in cortical regions. Multimodal combinations An increase in the balanced accuracy for patients imaged using multiple modalities was expected, according to the literature, for the combinations of EEG and FDG-PET, 41 EEG and fMRI-RS, 39 , 40 and EEG and dMRI, 42 aligning with our results. Similar studies on cardiac arrest patients revealed that a model using only three FA features outperformed models incorporating either only the FA global scores, clinical data or grey matter apparent diffusion coefficient, 62 or an enhanced AUC with the integration of scores and metrics from multiple modalities (EEG, aMRI, dMRI). 28 A case study by Comanducci et al . 11 illustrated how a longitudinal multimodal analysis can reveal covert signs of consciousness in an unresponsive patient. Additionally, Rohaut et al . 33 demonstrated that integrating multimodal observations enhances neuro-prognostication performance. While features across modalities may correlate due to shared neural sources (e.g. slow EEG rhythms and reduced metabolism), each modality also captures distinct noise. This uncorrelated noise can lead to synergy in multimodal models, even when the signals overlap, by improving the overall signal-to-noise ratio. Thus, synergy may reflect both complementary information and the statistical benefits of combining modalities. Benefits, caveats and the future of integrative multimodal neuroimaging for DoC The current electrophysiology and neuroimaging modalities capture some aspect of brain preservation—either the underlying structure or dynamics, and they contain non-redundant information. The importance of having a multidimensional perspective of this clinical group has been increasingly emphasized. 4 , 13 , 33 , 39-41 Overall, numerous reviews have placed focus on the advantage of having multimodal acquisitions 4 , 7 , 12 , 13 , 76 ; however, the practicalities, given the limitations already exposed in this paper, make implementation challenging. Furthermore, machine learning-based approaches, trained on behavioural labels to differentiate between UWS and MCS patients, may overlook conscious but unresponsive individuals, posing a circularity problem; however, there is some robustness to mislabelling if classifiers are trained with a sufficient amount of data. 27 Importantly, random forest classifiers are non-linear models, thus the relationship between feature importance and the models (measured here through the balanced accuracies) is non-linear. While data collection procedures for Dataset 1 (France) were standardized across time for EEG (LG and RS) using the same system, one change occurred in the neuroimaging protocols, due to the installation of a new hospital scanner. Although this may have introduced variability, all acquisition parameters were carefully documented as described in the ‘Materials and methods’ section and Supplementary material . For Datasets 2 and 3, collected across multiple centres, some variability in EEG hardware and MRI acquisition was unavoidable. However, we applied harmonized preprocessing pipelines and used feature extraction methods robust to differences in EEG channel count and MRI preprocessing to reduce site-related bias. A potential shortcoming of our study is that long-term outcome was assessed using a single phone-guided CRS-R evaluation, which, although validated, 33 , 77 may be subject to inter-rater variability and reduced precision. The absence of repeated assessments limited our ability to account for potential fluctuations in clinical state over time. Furthermore, although feature importance can be very informative in understanding how decisions are made, it has important limitations, and further analysis of redundancy and synergy can paint a clearer image of their relationships. Another limitation of the study is the lack of healthy controls, which would render the results more reliable. The reliability of the unimodal model results can be confirmed if the models classify the healthy controls as being in an MCS category and not in UWS. Lastly, future investigations should focus on the distinction between different aetiologies of patients with DoC using combined multimodal approaches (due to interactions of neural signals with aetiology 30 , 69 ). Conclusions In this study, we developed an explainable machine learning approach for the classification of DoC patients from a large dataset of multimodal neuroimaging and electrophysiology recordings. The observed distinctions in accuracy, feature importance and pairwise disagreements underscore the need for tailored strategies in leveraging different modalities for enhanced clinical decision-making. Comparing the current states of patients and their evolution across modalities and features (regions or other signal summaries) may pave the way for more thorough investigations into aetiologies or integrative neuroimaging studies with narrower hypotheses. Supplementary Material awaf412_Supplementary_Data awaf412_supplementary_data.pdf (5.4MB, pdf) Acknowledgements We thank all the participants who took part in the studies. We would like to thank the work and support of the clinicians at the Neuro ICU, DMU Neurosciences, APHP Sorbonne Université, Hôpital de la Pitié Salpêtrière, Paris, France; University Hospital of the Ludwig-Maximilians-University of Munich; Therapiezentrum Burgau; the University of Milan; Fondazione Don Carlo Gnocchi Santa Maria Nascente; the Weizmann Institute of Science, and the patient families whose consent and understanding are essential to the progress of the field. Contributor Information Dragana Manasova, Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France; Université Paris Cité, Paris 75006, France. Laouen Mayal Louan Belloli, Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France; Laboratorio de Inteligencia Artificial Aplicada, Instituto de Ciencias de la Computación, Universidad de Buenos Aires, Buenos Aires C1053, Argentina; Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET), Ministry of Science, Technology and Innovation, Buenos Aires C1053, Argentina. Martin Justinus Rosenfelder, Department of Neurology, University Hospital of the Ludwig-Maximilians-Universität München, Munich 82152, Germany; Therapiezentrum Burgau, Hospital for Neurological Rehabilitation, Burgau 89331, Germany; Clinical and Biological Psychology, Institute of Psychology and Education, Ulm University, Ulm 89081, Germany. Lina Willacker, Department of Neurology, University Hospital of the Ludwig-Maximilians-Universität München, Munich 82152, Germany. Emilia Fló Rama, Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France. Chiara Valota, Department of Biomedical and Clinical Sciences, University of Milano, Milan 20157, Italy; IRCCS Fondazione Don Carlo Gnocchi ONLUS, Milan 20148, Italy. Bertrand Hermann, Inserm 1266, Institute of Psychiatry and Neurosciences of Paris, Université Paris Cité, Paris F-75014, France; Medical Intensive Care Unit, HEGP Hôpital, Assistance Publique - Hôpitaux de Paris-Centre (APHP-Centre), Paris 75014, France. Brigitte Charlotte Kaufmann, Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France. Alice Pirastru, IRCCS Fondazione Don Carlo Gnocchi ONLUS, Milan 20148, Italy. Chiara Camilla Derchi, IRCCS Fondazione Don Carlo Gnocchi ONLUS, Milan 20148, Italy. Theresa Raiser, Department of Neurology, University Hospital of the Ludwig-Maximilians-Universität München, Munich 82152, Germany. Melanie Valente, Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France. Aude Sangare, Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France. Başak Türker, Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France. Nadya Pyatigorskaya, Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France. Benoît Béranger, Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France. Michele Colombo, Department of Biomedical and Clinical Sciences, University of Milano, Milan 20157, Italy. Esteban Munoz-Musat, Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France; Centre Mémoire de Ressources et de Recherche, Paris Nord/Université Paris-Cité, Paris 75006, France. Anira Escrichs, Center for Brain and Cognition, Computational Neuroscience Group, Universitat Pompeu Fabra, Barcelona 08005, Spain. Tiziana Atzori, IRCCS Fondazione Don Carlo Gnocchi ONLUS, Milan 20148, Italy. Francesca Baglio, IRCCS Fondazione Don Carlo Gnocchi ONLUS, Milan 20148, Italy. Constantin Lapa, Nuclear Medicine, Faculty of Medicine, University of Augsburg, Augsburg 86156, Germany. Ansgar Berlis, Diagnostic and Interventional Neuroradiology, Faculty of Medicine, University of Augsburg, Augsburg 86156, Germany. Kristina Krüger, Diagnostic and Interventional Neuroradiology, Faculty of Medicine, University of Augsburg, Augsburg 86156, Germany. Tina Luther, Department of Neurology, University Hospital of the Ludwig-Maximilians-Universität München, Munich 82152, Germany; Therapiezentrum Burgau, Hospital for Neurological Rehabilitation, Burgau 89331, Germany. Vincent Perlbarg, BRAINTALE SAS, Paris 75013, France. Gustavo Deco, Center for Brain and Cognition, Computational Neuroscience Group, Universitat Pompeu Fabra, Barcelona 08005, Spain; Institució Catalana de la Recerca I Estudis Avançats (ICREA), Barcelona 08010, Spain. Yonathan Sanz-Perl, Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France; Institució Catalana de la Recerca I Estudis Avançats (ICREA), Barcelona 08010, Spain. Enzo Tagliazucchi, Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET), Ministry of Science, Technology and Innovation, Buenos Aires C1053, Argentina; Latin American Brain Health Institute (BrainLat), Universidad Adolfo Ibáñez, Santiago 7941169, Chile. Louis Puybasset, BRAINTALE SAS, Paris 75013, France; GRC 29, AP-HP, DMU DREAM, Department of Anaesthesiology and Critical Care Medicine, Pitié-Salpêtrière Hospital, Sorbonne University, Paris 75013, France. Benjamin Rohaut, Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France; AP-HP, Hôpital de la Pitié Salpêtrière, Neuro ICU, DMU Neurosciences, Paris 75013, France. Lionel Naccache, Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France; AP-HP, Hôpital Pitié - Salpêtrière, Service de Neurophysiologie Clinique, Paris 75013, France. Angela Comanducci, IRCCS Fondazione Don Carlo Gnocchi ONLUS, Milan 20148, Italy. Anat Arzi, Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France; Department of Medical Neurobiology, Institute for Medical Research Israel-Canada, Faculty of Medicine, The Hebrew University of Jerusalem, Jerusalem 9112102, Israel; Department of Cognitive and Brain Sciences, The Hebrew University of Jerusalem, Jerusalem 9112102, Israel. Mario Rosanova, Department of Biomedical and Clinical Sciences, University of Milano, Milan 20157, Italy. Andreas Bender, Department of Neurology, University Hospital of the Ludwig-Maximilians-Universität München, Munich 82152, Germany; Therapiezentrum Burgau, Hospital for Neurological Rehabilitation, Burgau 89331, Germany; Department of Neurorehabilitation, Medical Faculty, University of Augsburg, Augsburg 86156, Germany. Jacobo Diego Sitt, Institut du Cerveau - Paris Brain Institute - ICM, Inserm, CNRS, Sorbonne Université, Paris 75013, France. Data availability The data are not publicly available. Codes used in the analyses will be made publicly available upon publication at https://github.com/DraganaMana/multimod_doc . Funding This work was supported by the Ecole Doctorale Frontières de l’Innovation en Recherche et Education–Fondation Bettencourt Schueller (to D.M.). This project is part of the multicentric application for the EU ERAPerMed Joint Translational Call for Proposals for ‘Personalised Medicine: Multidisciplinary research towards implementation’ (ERA PerMed JTC2019). It is funded by local funding agencies of the participating countries (for France it is the Agence Nationale de la Recherche ANR, funding code: ANR-19-PERM-0002, for Germany the Federal Ministry of Education and Research BMBF, funding code: 01KU2003, for Italy, the Fondazione Regionale per la Ricerca Biomedica, funding code: GA 77982), also supported and funded by the Italian Ministry of Health—Ricerca Corrente 2025–2027 (A.C., C.D., T.A.). This project is supported by the MODELDxConsciousness Consortium (Flag-ERA JTC 2023); as well as the ECOS-SUD A20M02 (Argentina/France), and by Paris Brain Institute America’s project on Consciousness mapping. ERC-2022-SYG Grant number 101071900 neurological mechanisms of injury and sleep-like cellular dynamics (NEMESIS) (to M.R.) Competing interests J.D.S. and L.N. are scientific co-founders of NeuroMeters (have scientific advisory activity but no executive or management activity). M.R. is shareholder and scientific advisor of Intrinsic Powers, a spin-off of the University of Milan. This affiliation in no way affects the content of this article. Supplementary material Supplementary material is available at Brain online. References 1. Bodien  YG, Allanson  J, Cardone  P, et al.  Cognitive motor dissociation in disorders of consciousness. N Engl J Med. 2024;391:598–608. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 2. Owen  AM, Coleman  MR, Boly  M, Davis  MH, Laureys  S, Pickard  JD. Detecting awareness in the vegetative state. Science. 2006;313:1402. [ DOI ] [ PubMed ] [ Google Scholar ] 3. Bruno  MA, Vanhaudenhuyse  A, Thibaut  A, Moonen  G, Laureys  S. From unresponsive wakefulness to minimally conscious PLUS and functional locked-in syndromes: Recent advances in our understanding of disorders of consciousness. J Neurol. 2011;258:1373–1384. [ DOI ] [ PubMed ] [ Google Scholar ] 4. Gallucci  A, Varoli  E, Del Mauro  L, et al.  Multimodal approaches supporting the diagnosis, prognosis and investigation of neural correlates of disorders of consciousness: A systematic review. Eur J Neurosci. 2024;59:874–933. [ DOI ] [ PubMed ] [ Google Scholar ] 5. Jennett  B, Plum  F. Persistent vegetative state after brain damage: A syndrome in search of a name. Lancet. 1972;299:734–737. [ DOI ] [ PubMed ] [ Google Scholar ] 6. Giacino  JT, Ashwal  S, Childs  N, et al.  The minimally conscious state: Definition and diagnostic criteria. Neurology. 2002;58:349–353. [ DOI ] [ PubMed ] [ Google Scholar ] 7. Edlow  BL, Claassen  J, Schiff  ND, Greer  DM. Recovery from disorders of consciousness: Mechanisms, prognosis and emerging therapies. Nat Rev Neurol. 2021;17:135–156. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Giacino  JT, Kalmar  K, Whyte  J. The JFK coma recovery scale-revised: Measurement characteristics and diagnostic utility. Arch Phys Med Rehabil. 2004;85:2020–2029. [ DOI ] [ PubMed ] [ Google Scholar ] 9. Wannez  S, Heine  L, Thonnard  M, Gosseries  O, Laureys  S. The repetition of behavioral assessments in diagnosis of disorders of consciousness. Ann Neurol. 2017;81:883–889. [ DOI ] [ PubMed ] [ Google Scholar ] 10. Comanducci  A, Boly  M, Claassen  J, et al.  Clinical and advanced neurophysiology in the prognostic and diagnostic evaluation of disorders of consciousness: Review of an IFCN-endorsed expert group. Clin Neurophysiol. 2020;131:2736–2765. [ DOI ] [ PubMed ] [ Google Scholar ] 11. Comanducci  A, Casarotto  S, Rosanova  M, et al.  Unconsciousness or unresponsiveness in akinetic mutism? Insights from a multimodal longitudinal exploration. Eur J Neurosci. 2024;59:860–873. [ DOI ] [ PubMed ] [ Google Scholar ] 12. Edlow  BL, Fecchio  M, Bodien  YG, et al.  Measuring consciousness in the intensive care unit. Neurocrit Care.  2023;38:584–590. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Kondziella  D, Bender  A, Diserens  K, et al.  European academy of neurology guideline on the diagnosis of coma and other disorders of consciousness. Eur J Neurol. 2020;27:741–756. [ DOI ] [ PubMed ] [ Google Scholar ] 14. Bekinschtein  TA, Dehaene  S, Rohaut  B, Tadel  F, Cohen  L, Naccache  L. Neural signature of the conscious processing of auditory regularities. Proc Natl Acad Sci U S A.  2009;106:1672–1677. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Vanhaudenhuyse  A, Noirhomme  Q, Tshibanda  LJF, et al.  Default network connectivity reflects the level of consciousness in non-communicative brain-damaged patients. Brain. 2010;133:161–171. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Faugeras  F, Rohaut  B, Weiss  N, et al.  Event related potentials elicited by violations of auditory regularities in patients with impaired consciousness. Neuropsychologia. 2012;50:403–418. [ DOI ] [ PubMed ] [ Google Scholar ] 17. Demertzi  A, Gómez  F, Crone  JS, et al.  Multiple fMRI system-level baseline connectivity is disrupted in patients with consciousness alterations. Cortex. 2014;52:35–46. [ DOI ] [ PubMed ] [ Google Scholar ] 18. Sitt  JD, King  JR, El Karoui  I, et al.  Large scale screening of neural signatures of consciousness in patients in a vegetative or minimally conscious state. Brain. 2014;137:2258–2270. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. King  JR, Sitt  JD, Faugeras  F, et al.  Information sharing in the brain indexes consciousness in noncommunicative patients. Curr Biol. 2013;23:1914–1919. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Silva  S, De Pasquale  F, Vuillaume  C, et al.  Disruption of posteromedial large-scale neural communication predicts recovery from coma. Neurology. 2015;85:2036–2044. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Demertzi  A, Antonopoulos  G, Heine  L, et al.  Intrinsic functional connectivity differentiates minimally conscious from unresponsive patients. Brain. 2015;138:2619–2631. [ DOI ] [ PubMed ] [ Google Scholar ] 22. Wu  X, Zou  Q, Hu  J, et al.  Intrinsic functional connectivity patterns predict consciousness level and recovery outcome in acquired brain injury. J Neurosci. 2015;35:12932–12946. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Schorr  B, Schlee  W, Arndt  M, Bender  A. Coherence in resting-state EEG as a predictor for the recovery from unresponsive wakefulness syndrome. J Neurol. 2016;263:937–953. [ DOI ] [ PubMed ] [ Google Scholar ] 24. Chennu  S, Annen  J, Wannez  S, et al.  Brain networks predict metabolism, diagnosis and prognosis at the bedside in disorders of consciousness. Brain. 2017;140:2120–2132. [ DOI ] [ PubMed ] [ Google Scholar ] 25. Kondziella  D, Fisher  PM, Larsen  VA, et al.  Functional MRI for assessment of the default mode network in acute brain injury. Neurocrit Care. 2017;27:401–406. [ DOI ] [ PubMed ] [ Google Scholar ] 26. Threlkeld  ZD, Bodien  YG, Rosenthal  ES, et al.  Functional networks reemerge during recovery of consciousness after acute severe traumatic brain injury. Cortex. 2018;106:299–308. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Engemann  DA, Raimondo  F, King  JR, et al.  Robust EEG-based cross-site and cross-protocol classification of states of consciousness. Brain. 2018;141:3179–3192. [ DOI ] [ PubMed ] [ Google Scholar ] 28. Velly  L, Perlbarg  V, Boulier  T, et al.  Use of brain diffusion tensor imaging for the prediction of long-term neurological outcomes in patients after cardiac arrest: A multicentre, international, prospective, observational, cohort study. Lancet Neurol. 2018;17:317–326. [ DOI ] [ PubMed ] [ Google Scholar ] 29. Demertzi  A, Tagliazucchi  E, Dehaene  S, et al.  Human consciousness is supported by dynamic complex patterns of brain signal coordination. Sci Adv. 2019;5:eaat7603. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 30. Colombo  MA, Comanducci  A, Casarotto  S, et al.  Beyond alpha power: EEG spatial and spectral gradients robustly stratify disorders of consciousness. Cereb Cortex. 2023;33:7193–7210. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 31. Medina Carrion  JP, Stanziano  M, D’Incerti  L, et al.  Disorder of consciousness: Structural integrity of brain networks for the clinical assessment. Ann Clin Transl Neurol. 2023;10:384–396. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. Annen  J, Frasso  G, Crone  JS, et al.  Regional brain volumetry and brain function in severely brain-injured patients. Ann Neurol. 2018;83:842–853. [ DOI ] [ PubMed ] [ Google Scholar ] 33. Rohaut  B, Calligaris  C, Hermann  B, et al.  Multimodal assessment improves neuroprognosis performance in clinically unresponsive critical-care patients with brain injury. Nat Med. 2024;30:2349–2355. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 34. Bender  A, Eifert  B, Rubi-Fessen  I, Jox  RJ, Maurer-Karattup  P, Müller  F. The neurological rehabilitation of adults with coma and disorders of consciousness. Dtsch Arzteblatt Int. 2023;120:605–612. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 35. Stender  J, Gosseries  O, Bruno  MA, et al.  Diagnostic precision of PET imaging and functional MRI in disorders of consciousness: A clinical validation study. Lancet. 2014;384:514–522. [ DOI ] [ PubMed ] [ Google Scholar ] 36. Di Perri  C, Bahri  MA, Amico  E, et al.  Neural correlates of consciousness in patients who have emerged from a minimally conscious state: A cross-sectional multimodal imaging study. Lancet Neurol. 2016;15:830–842. [ DOI ] [ PubMed ] [ Google Scholar ] 37. Aubinet  C, Cassol  H, Gosseries  O, et al.  Brain metabolism but not gray matter volume underlies the presence of language function in the minimally conscious state (MCS): MCS+ versus MCS− neuroimaging differences. Neurorehabil Neural Repair. 2020;34:172–184. [ DOI ] [ PubMed ] [ Google Scholar ] 38. Candia-Rivera  D, Annen  J, Gosseries  O, et al.  Neural responses to heartbeats detect residual signs of consciousness during resting state in postcomatose patients. J Neurosci. 2021;41:5251–5262. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 39. Amiri  M, Fisher  PM, Raimondo  F, et al.  Multimodal prediction of residual consciousness in the intensive care unit: The CONNECT-ME study. Brain. 2022;146:50–64. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 40. Amiri  M, Raimondo  F, Fisher  PM, et al.  Multimodal prediction of 3- and 12-month outcomes in ICU patients with acute disorders of consciousness. Neurocrit Care.  2023;40:718–733. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 41. Hermann  B, Stender  J, Habert  MO, et al.  Multimodal FDG-PET and EEG assessment improves diagnosis and prognostication of disorders of consciousness. Neuroimage Clin. 2021;30:102601. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 42. Altmayer  V, Sangare  A, Calligaris  C, et al.  Functional and structural brain connectivity in disorders of consciousness. Brain Struct Funct. 2024;229:2285–2298. [ DOI ] [ PubMed ] [ Google Scholar ] 43. Willacker  L, Raiser  TM, Bassi  M, et al.  PerBrain: A multimodal approach to personalized tracking of evolving state-of-consciousness in brain-injured patients: Protocol of an international, multicentric, observational study. BMC Neurol.  2022;22:468. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 44. Wilson  JTL, Pettigrew  LEL, Teasdale  GM. Structured interviews for the Glasgow outcome scale and the extended Glasgow outcome scale: Guidelines for their use. J Neurotrauma. 1998;15:573–580. [ DOI ] [ PubMed ] [ Google Scholar ] 45. Pedregosa  F, Varoquaux  G, Gramfort  A, et al.  Scikit-learn: Machine learning in python. J Mach Learn Res. 2011;12:2825–2830. [ Google Scholar ] 46. Stender  J, Kupers  R, Rodell  A, et al.  Quantitative rates of brain glucose metabolism distinguish minimally conscious from vegetative state patients. J Cereb Blood Flow Metab. 2015;35:58–65. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 47. Rosazza  C, Andronache  A, Sattin  D, et al.  Multimodal study of default-mode network integrity in disorders of consciousness. Ann Neurol. 2016;79:841–853. [ DOI ] [ PubMed ] [ Google Scholar ] 48. Golkowski  D, Kiel  T, Schorr  B, et al.  Simultaneous EEG-PET-fMRI measurements in disorders of consciousness: An exploratory study on diagnosis and prognosis. J Neurol. 2017;264:1986–1995. [ DOI ] [ PubMed ] [ Google Scholar ] 49. O’Donnell  A, Pauli  R, Banellis  L, et al.  The prognostic value of resting-state EEG in acute post-traumatic unresponsive states. Brain Commun. 2021;3:fcab017. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 50. Stefan  S, Schorr  B, Lopez-Rolon  A, et al.  Consciousness indexing and outcome prediction with resting-state EEG in severe disorders of consciousness. Brain Topogr. 2018;31:848–862. [ DOI ] [ PubMed ] [ Google Scholar ] 51. Gopinath  K, Greve  DN, Das  S, Arnold  S, Magdamo  C, Iglesias  JE. Cortical analysis of heterogeneous clinical brain MRI scans for large-scale neuroimaging studies. In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2023. MICCAI 2023. Lecture Notes in Computer Science . Vol 14227. Springer;2023:35-45. 52. Billot  B, Greve  DN, Puonti  O, et al.  SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining. Med Image Anal. 2023;86:102789. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 53. Billot  B, Magdamo  C, Cheng  Y, Arnold ID  SE, Das  SI, Eugenio Iglesias  J. Robust machine learning segmentation for large-scale analysis of heterogeneous clinical brain MRI datasets. Proc Natl Acad Sci U S A.  2023;120:e2216399120. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 54. Iglesias  JE, Billot  B, Balbastre  Y, et al.  SynthSR: A public AI tool to turn heterogeneous clinical brain scans into high-resolution T1-weighted images for 3D morphometry. Sci Adv. 2023;9:eadd3607. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 55. van der Vliet  R, Selles  RW, Andrinopoulou  ER, et al.  Predicting upper limb motor impairment recovery after stroke: A mixture model. Ann Neurol.  2020;87:383–393. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 56. Duffau  H. Does post-lesional subcortical plasticity exist in the human brain?  Neurosci Res.  2009;65:131–135. [ DOI ] [ PubMed ] [ Google Scholar ] 57. Nudo  RJ. Recovery after brain injury: Mechanisms and principles. Front Hum Neurosci.  2013;7:887. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 58. Jones  EG. Cortical and subcortical contributions to activity-dependent plasticity in primate somatosensory Cortex. Annu Rev Neurosci.  2000;23:1–37. [ DOI ] [ PubMed ] [ Google Scholar ] 59. Ewall  G, Parkins  S, Lin  A, Jaoui  Y, Lee  HK. Cortical and subcortical circuits for cross-modal plasticity induced by loss of vision. Front Neural Circuits.  2021;15:665009. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 60. Fernández-Espejo  D, Bekinschtein  T, Monti  MM, et al.  Diffusion weighted imaging distinguishes the vegetative state from the minimally conscious state. NeuroImage. 2011;54:103–112. [ DOI ] [ PubMed ] [ Google Scholar ] 61. Zheng  ZS, Reggente  N, Lutkenhoff  E, Owen  AM, Monti  MM. Disentangling disorders of consciousness: Insights from diffusion tensor imaging and machine learning. Hum Brain Mapp. 2017;38:431–443. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 62. Luyt  CE, Galanaud  D, Perlbarg  V, et al.  Diffusion tensor imaging to predict long-term outcome after cardiac arrest: A bicentric pilot study. Anesthesiology. 2012;117:1311–1321. [ DOI ] [ PubMed ] [ Google Scholar ] 63. Puybasset  L, Perlbarg  V, Unrug  J, et al.  Prognostic value of global deep white matter DTI metrics for 1-year outcome prediction in ICU traumatic brain injury patients: An MRI-COMA and CENTER-TBI combined study. Intensive Care Med. 2022;48:201–212. [ DOI ] [ PubMed ] [ Google Scholar ] 64. Bruno  MA, Fernández-Espejo  D, Lehembre  R, et al.  Multimodal neuroimaging in patients with disorders of consciousness showing “functional hemispherectomy”. In: Progress in brain research. Vol 193. Elsevier B.V.; 2011:323–333. [ DOI ] [ PubMed ] [ Google Scholar ] 65. Bayne  T, Seth  AK, Massimini  M. Are there islands of awareness?  Trends Neurosci. 2020;43:6–16. [ DOI ] [ PubMed ] [ Google Scholar ] 66. Newcombe  VFJ, Williams  GB, Scoffings  D, et al.  Aetiological differences in neuroanatomy of the vegetative state: Insights from diffusion tensor imaging and functional implications. J Neurol Neurosurg Psychiatry. 2010;81:552–561. [ DOI ] [ PubMed ] [ Google Scholar ] 67. Naccache  L. Minimally conscious state or cortically mediated state?  Brain. 2018;141:949–960. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 68. Hermann  B, Sangaré  A, Munoz-Musat  E, et al.  Importance, limits and caveats of the use of “disorders of consciousness” to theorize consciousness. Neurosci Conscious. 2021;2021:niab048. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 69. Maschke  C, Belloli  L, Manasova  D, Sitt  JD, Blain-Moraes  S. The role of etiology in the identification of clinical markers of consciousness: Comparing EEG alpha power, complexity, and spectral exponent. Cereb Cortex. 2025;35:bhaf254. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 70. Schiff  ND. Mesocircuit mechanisms underlying recovery of consciousness following severe brain injuries: Model and predictions. In: Brain function and responsiveness in disorders of consciousness. Springer International Publishing; 2016:195–204. [ Google Scholar ] 71. Daltrozzo  J, Wioland  N, Mutschler  V, Kotchoubey  B. Predicting coma and other low responsive patients outcome using event-related brain potentials: A meta-analysis. Clin Neurophysiol. 2007;118:606–614. [ DOI ] [ PubMed ] [ Google Scholar ] 72. Bodart  O, Gosseries  O, Wannez  S, et al.  Measures of metabolism and complexity in the brain of patients with disorders of consciousness. Neuroimage Clin. 2017;14:354–362. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 73. Fernandez-Espejo  D, Soddu  A, Cruse  D, et al.  A role for the default mode network in the bases of disorders of consciousness. Ann Neurol. 2012;72:335–343. [ DOI ] [ PubMed ] [ Google Scholar ] 74. Tasserie  J, Uhrig  L, Sitt  JD, et al.  Deep brain stimulation of the thalamus restores signatures of consciousness in a nonhuman primate model. Sci Adv.  2022;8:eable5547. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 75. Schiff  ND, Giacino  JT, Butson  CR, et al.  Thalamic deep brain stimulation in traumatic brain injury: A phase 1, randomized feasibility study. Nat Med.  2023;29:3162–3174. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 76. Coleman  MR, Bekinschtein  T, Monti  MM, Owen  AM, Pickard  JD. A multimodal approach to the assessment of patients with disorders of consciousness. Prog Brain Res. 2009;177(C):231–248. [ DOI ] [ PubMed ] [ Google Scholar ] 77. Sterling  A, Bodien  Y, Bergin  M, et al.  Validity of the telephone-administered coma recovery scale-revised and confusion assessment protocol for standardized remote assessment of persons with disorders of consciousness. Arch Phys Med Rehabil.  2024;105:e23–e24. [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials awaf412_Supplementary_Data awaf412_supplementary_data.pdf (5.4MB, pdf) Data Availability Statement The data are not publicly available. Codes used in the analyses will be made publicly available upon publication at https://github.com/DraganaMana/multimod_doc . Articles from Brain are provided here courtesy of Oxford University Press ACTIONS View on publisher site PDF (1.0 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top