Thèse du Niveau à la Fonction Phénotypage Computationnel des Troubles de la Conscience à Partir de Réponses Neuronales en un Seul Essai Stand-Out H/F - Doctorat.Gouv.Fr
- CDD
- Doctorat.Gouv.Fr
Les missions du poste
Établissement : Université de Toulouse École doctorale : CLESCO - Comportement, Langage, Éducation, Socialisation, Cognition Laboratoire de recherche : ToNIC-Toulouse NeuroImaging Center (UMR 1214) Direction de la thèse : Stein SILVA ORCID 0000000196907413 Début de la thèse : 2027-09-01 Date limite de candidature : 2026-11-16T23:59:59 L'évaluation clinique de la conscience après une lésion cérébrale sévère demeure fondamentalement comportementaliste : les patients sont classés selon ce qu'ils peuvent manifester, et non selon ce qu'ils peuvent traiter. Ce paradigme est à la fois statique et faillible : statique, car il impose une catégorie fixe à un état dont la fluctuation constitue précisément une caractéristique essentielle ; faillible, car un patient peut percevoir, apprendre et anticiper tout en étant privé des voies motrices lui permettant de le démontrer. Les chiffres révèlent l'ampleur de cette impasse conceptuelle : malgré l'utilisation de l'ensemble des outils recommandés, 50 à 77 % des patients à la phase aiguë demeurent dans une zone grise pronostique. Or, la majorité des décès après un coma surviennent à la suite d'une décision de limitation ou d'arrêt des traitements prise au coeur même de cette incertitude. Autrement dit, le diagnostic repose encore sur un pari fondé sur l'absence de preuve.
La neuro-imagerie a partiellement répondu à cette difficulté en affinant les marqueurs du niveau de conscience - signatures spectrales, intégration des réseaux, entropie des états cérébraux dynamiques. Ces outils présentent désormais une réelle valeur pronostique clinique, reproductible entre modalités et centres. Toutefois, détecter et quantifier la conscience ne revient pas à en comprendre les mécanismes. Un indice de complexité peut indiquer l'intensité de l'activité cérébrale, mais non la manière dont le cerveau traite les informations reçues. Le patient accumule-t-il encore des preuves sensorielles ? Maintient-il un modèle prédictif du monde auditif, pondérant la surprise selon la fiabilité de ses attentes ? Ces questions portent sur la fonction, et non sur l'amplitude. Aucune mesure de connectivité ou de complexité n'a été conçue pour y répondre, car aucune ne modélise explicitement la dynamique neuronale comme un processus d'inférence.
Les paradigmes auditifs passifs se rapprochent le plus de cette frontière. Le paradigme local-global distingue les régularités préservées en l'absence de conscience des violations associées à l'accès conscient ; leur détection prédit la récupération dans de vastes cohortes prospectives. Des approches plus récentes de décodage et d'apprentissage profond étendent cette démarche à la prédiction de l'éveil au lit du patient. Pourtant, une revue systématique publiée en 2026 met en évidence une limite persistante : le traitement prédictif est régulièrement invoqué comme cadre théorique, tandis que les analyses reposent sur des classifications génériques utilisant des caractéristiques spectrales ou entropiques. La théorie est citée, mais jamais ajustée aux données. Aucune étude n'a encore estimé, dans cette population, les paramètres individuels d'un modèle d'inférence. L'écart entre métaphore et mécanisme demeure donc non quantifié.
STAND-OUT vise à combler cette lacune. En ajustant des modèles bayésiens hiérarchiques d'inférence aux réponses neuronales essai par essai, le projet ne cherche plus à déterminer ce qu'un patient non communicant peut manifester, mais ce qu'il demeure capable de calculer. Son ambition est triple : approfondir notre compréhension de l'architecture neuronale de l'inférence et de son altération après une lésion cérébrale ; fournir aux cliniciens une base fonctionnelle, au-delà du seul comportement, pour évaluer la conscience des patients incapables de s'exprimer ; et ouvrir, à terme, la voie à une neuromodulation personnalisée. À l'interface des neurosciences fondamentales, de la modélisation computationnelle et de la validation clinique, STAND-OUT reformule la question centrale de la médecine du coma : non plus « quelqu'un est-il encore là ? », mais « que peut-il encore, précisément, connaître ? ». The clinical assessment of consciousness after severe brain injury rests on behaviour: patients are classified according to what they can demonstrate to an examiner. This is both static and fallible. It is static because it assigns a fixed diagnostic category to states whose most consistent property is fluctuation. It is fallible because a patient may perceive, learn and anticipate without possessing the motor pathway required to show it. The documented misdiagnosis rates in this population, and the fact that between 50% and 77% of acute patients remain in a prognostic grey zone even when all guideline-recommended tools are combined, are the clinical expression of a conceptual problem: current assessment measures output capacity, not processing capacity. The stakes are immediate - most deaths after coma follow decisions to withdraw life-sustaining therapy taken under precisely this uncertainty.
Neuroimaging has addressed this problem by developing markers of the level of consciousness: spectral and complexity measures, network integration, dynamic brain state repertoires. Our two groups have contributed directly to this effort, showing that weighted entropy of dynamic brain states predicts three-month outcome in acute patients in fMRI and that the same framework generalises across three clinical centres in high-density EEG with real-time bedside feasibility. This generation of markers has largely accomplished what it set out to do: detect and grade.
It leaves untouched a different question, and arguably the one that determines both clinical management and functional prognosis. Is this patient still accumulating sensory evidence, and how efficiently? Does he/she maintain a predictive model of his/her auditory environment? Does he/she weight incoming surprise according to how reliable his/her expectations are? These are questions about function, not level. No complexity or connectivity measure can answer them, because such measures describe neural dynamics without interpreting them as inference.
Passive auditory paradigms already give the field an entry point. The local-global paradigm dissociates local regularity violations, which survive unconsciousness, from global violations, which are far more closely tied to conscious access; the presence of a global effect has been shown to predict recovery of overt consciousness in large prospective cohorts. Studies in comatose patients have used multivariate decoding of single-trial responses to global violations and found that the progression of decoding performance predicts awakening, and more recent work has applied deep-learning classifiers to auditory EEG responses in 134 comatose patients to predict awakening and survival.
What these approaches share is that they classify. They ask whether patient and control responses are separable, not what computation the response reflects. A 2026 systematic review of predictive-processing approaches in disorders of consciousness makes this explicit: the existing literature invokes predictive processing as a theoretical motivation while implementing generic machine-learning pipelines over spectral, entropy or connectivity features. The theory is cited; the model is never fitted. No study to date has estimated subject-specific parameters of an inference model in this population.
The STAND-OUT project directly addresses the BEST strategic theme by advancing the understanding and assessment of brain health in individuals with severe alterations of consciousness. Consciousness is fundamental to human well-being, autonomy and dignity, yet in patients who cannot produce a behavioural response it cannot be reliably assessed at all. This creates a major risk of diagnostic error, inappropriate prognostication and unequal access to therapeutic interventions. STAND-OUT addresses this gap by fitting hierarchical Bayesian models of inference to single-trial neural responses, characterising what a non-communicating patient can still process rather than what they can display.
The project will promote well-being at several complementary levels. Scientifically, it will deepen understanding of the neural mechanisms underpinning conscious processing and their disruption after severe brain injury. Clinically, it may enable more accurate assessment of brain-injured patients who cannot communicate, reducing uncertainty and supporting better-informed decisions - and, in the longer term, opening the way to individually tailored neuromodulation strategies. By combining fundamental neuroscience, computational modelling and clinically meaningful validation, the project establishes a translational pathway from understanding brain function to promoting neurological recovery, autonomy and quality of life after brain injury. The method this thesis proposes is not new - it is established, validated, and has simply never been brought to this population. In the tradition of computational single-trial EEG analysis, a hierarchical Bayesian model of learning generates trial-wise trajectories of beliefs and precision-weighted prediction errors, and these trajectories are used to explain single-trial variation in the evoked response. The Hierarchical Gaussian Filter is explicitly constructed so that its response model can link trial-wise perceptual quantities to any observable - choices, reaction times, eye movements, or evoked response amplitude. The neural signal takes the place ordinarily occupied by the reaction time. No behavioural output is required.
This approach has an established track record on exactly the paradigms available at the bedside. Trial-wise modelling of the mismatch negativity was formalised with explicit model comparison by Lieder et al. (2013); Stefanics et al. (2018) demonstrated that single-trial ERP amplitudes track precision-weighted prediction errors derived from a Bayes-optimal agent; pharmacological studies have shown these quantities to be sensitive to NMDA-receptor antagonism. Most relevant here, the same pipeline has been used clinically and prognostically: in individuals at clinical high risk for psychosis, low-level precision-weighted prediction errors distinguished those who later converted from those who did not Structurally, that design is the one proposed here, with recovery in place of conversion. The term itself is already in use in this literature - computational phenotyping of belief updating .
What this is not. The dominant computational tradition in disorders of consciousness is whole-brain generative modelling - mean-field and neural-mass models fitted to regional dynamics, estimating coupling parameters and simulating transitions between states. That work asks how the substrate behaves. This thesis asks what the patient infers. The models are cognitive, not biophysical; the estimated parameters are perceptual and epistemic rather than physiological. The two lines are complementary, and their integration is a longer-term horizon, but they should not be confused.
Three experiments:
The thesis is built as a validation ladder: each study establishes what the next one is entitled to assume.
Study 1 - Healthy volunteers: establishing neural-only estimation
Participants perform local-global and roving oddball paradigms with a behavioural task. Inference parameters are estimated from single-trial EEG alone and benchmarked against parameters estimated from behaviour in the same participants. This is the validation step that no clinical study can provide: it is the only setting in which the neural-only estimate can be checked against a behavioural ground truth. Acquisition in Toulouse and Córdoba; modelling in Córdoba.
Study 2 - Graded sedation: characterising orderly degradation
The validated pipeline is applied to graded, reversible, dose-controlled sedation, where consciousness is lost and recovered within the same individual. The test is whether specific parameters - sensory precision, volatility learning, prior weighting - degrade in an orderly, dissociable way rather than the evoked response simply diminishing in amplitude. This study bridges the gap between a cooperative participant and a patient who cannot respond, in a population where the cause of the change is known.
Study 3 - Acute comatose patients: prognostic value
In acutely comatose patients from the Toulouse ICU cohort, the thesis tests whether latent inference parameters predict three-month neurological outcome beyond entropy-based markers, ERP-based effects and standard clinical predictors, and - the clinically decisive question - whether they resolve patients whom existing markers leave in the grey zone.
Methodological challenge:
The central technical difficulty should be stated plainly, since it is also the thesis's main methodological contribution. In most published applications, belief trajectories are generated either by a Bayes-optimal agent with fixed parameters or by a model fitted to behaviour, and are then entered as regressors on the EEG. Estimating subject-specific parameters purely from physiology, with no behavioural anchor, is a substantially harder inverse problem and remains far less established. It raises identifiability questions - which parameters are recoverable, under what signal-to-noise ratio, with how many trials - that must be settled through systematic parameter recovery on simulated and real data before any clinical claim can be made. Study 1 exists precisely to bound this problem in a population where ground truth is available.
La neuro-imagerie a partiellement répondu à cette difficulté en affinant les marqueurs du niveau de conscience - signatures spectrales, intégration des réseaux, entropie des états cérébraux dynamiques. Ces outils présentent désormais une réelle valeur pronostique clinique, reproductible entre modalités et centres. Toutefois, détecter et quantifier la conscience ne revient pas à en comprendre les mécanismes. Un indice de complexité peut indiquer l'intensité de l'activité cérébrale, mais non la manière dont le cerveau traite les informations reçues. Le patient accumule-t-il encore des preuves sensorielles ? Maintient-il un modèle prédictif du monde auditif, pondérant la surprise selon la fiabilité de ses attentes ? Ces questions portent sur la fonction, et non sur l'amplitude. Aucune mesure de connectivité ou de complexité n'a été conçue pour y répondre, car aucune ne modélise explicitement la dynamique neuronale comme un processus d'inférence.
Les paradigmes auditifs passifs se rapprochent le plus de cette frontière. Le paradigme local-global distingue les régularités préservées en l'absence de conscience des violations associées à l'accès conscient ; leur détection prédit la récupération dans de vastes cohortes prospectives. Des approches plus récentes de décodage et d'apprentissage profond étendent cette démarche à la prédiction de l'éveil au lit du patient. Pourtant, une revue systématique publiée en 2026 met en évidence une limite persistante : le traitement prédictif est régulièrement invoqué comme cadre théorique, tandis que les analyses reposent sur des classifications génériques utilisant des caractéristiques spectrales ou entropiques. La théorie est citée, mais jamais ajustée aux données. Aucune étude n'a encore estimé, dans cette population, les paramètres individuels d'un modèle d'inférence. L'écart entre métaphore et mécanisme demeure donc non quantifié.
STAND-OUT vise à combler cette lacune. En ajustant des modèles bayésiens hiérarchiques d'inférence aux réponses neuronales essai par essai, le projet ne cherche plus à déterminer ce qu'un patient non communicant peut manifester, mais ce qu'il demeure capable de calculer. Son ambition est triple : approfondir notre compréhension de l'architecture neuronale de l'inférence et de son altération après une lésion cérébrale ; fournir aux cliniciens une base fonctionnelle, au-delà du seul comportement, pour évaluer la conscience des patients incapables de s'exprimer ; et ouvrir, à terme, la voie à une neuromodulation personnalisée. À l'interface des neurosciences fondamentales, de la modélisation computationnelle et de la validation clinique, STAND-OUT reformule la question centrale de la médecine du coma : non plus « quelqu'un est-il encore là ? », mais « que peut-il encore, précisément, connaître ? ». The clinical assessment of consciousness after severe brain injury rests on behaviour: patients are classified according to what they can demonstrate to an examiner. This is both static and fallible. It is static because it assigns a fixed diagnostic category to states whose most consistent property is fluctuation. It is fallible because a patient may perceive, learn and anticipate without possessing the motor pathway required to show it. The documented misdiagnosis rates in this population, and the fact that between 50% and 77% of acute patients remain in a prognostic grey zone even when all guideline-recommended tools are combined, are the clinical expression of a conceptual problem: current assessment measures output capacity, not processing capacity. The stakes are immediate - most deaths after coma follow decisions to withdraw life-sustaining therapy taken under precisely this uncertainty.
Neuroimaging has addressed this problem by developing markers of the level of consciousness: spectral and complexity measures, network integration, dynamic brain state repertoires. Our two groups have contributed directly to this effort, showing that weighted entropy of dynamic brain states predicts three-month outcome in acute patients in fMRI and that the same framework generalises across three clinical centres in high-density EEG with real-time bedside feasibility. This generation of markers has largely accomplished what it set out to do: detect and grade.
It leaves untouched a different question, and arguably the one that determines both clinical management and functional prognosis. Is this patient still accumulating sensory evidence, and how efficiently? Does he/she maintain a predictive model of his/her auditory environment? Does he/she weight incoming surprise according to how reliable his/her expectations are? These are questions about function, not level. No complexity or connectivity measure can answer them, because such measures describe neural dynamics without interpreting them as inference.
Passive auditory paradigms already give the field an entry point. The local-global paradigm dissociates local regularity violations, which survive unconsciousness, from global violations, which are far more closely tied to conscious access; the presence of a global effect has been shown to predict recovery of overt consciousness in large prospective cohorts. Studies in comatose patients have used multivariate decoding of single-trial responses to global violations and found that the progression of decoding performance predicts awakening, and more recent work has applied deep-learning classifiers to auditory EEG responses in 134 comatose patients to predict awakening and survival.
What these approaches share is that they classify. They ask whether patient and control responses are separable, not what computation the response reflects. A 2026 systematic review of predictive-processing approaches in disorders of consciousness makes this explicit: the existing literature invokes predictive processing as a theoretical motivation while implementing generic machine-learning pipelines over spectral, entropy or connectivity features. The theory is cited; the model is never fitted. No study to date has estimated subject-specific parameters of an inference model in this population.
The STAND-OUT project directly addresses the BEST strategic theme by advancing the understanding and assessment of brain health in individuals with severe alterations of consciousness. Consciousness is fundamental to human well-being, autonomy and dignity, yet in patients who cannot produce a behavioural response it cannot be reliably assessed at all. This creates a major risk of diagnostic error, inappropriate prognostication and unequal access to therapeutic interventions. STAND-OUT addresses this gap by fitting hierarchical Bayesian models of inference to single-trial neural responses, characterising what a non-communicating patient can still process rather than what they can display.
The project will promote well-being at several complementary levels. Scientifically, it will deepen understanding of the neural mechanisms underpinning conscious processing and their disruption after severe brain injury. Clinically, it may enable more accurate assessment of brain-injured patients who cannot communicate, reducing uncertainty and supporting better-informed decisions - and, in the longer term, opening the way to individually tailored neuromodulation strategies. By combining fundamental neuroscience, computational modelling and clinically meaningful validation, the project establishes a translational pathway from understanding brain function to promoting neurological recovery, autonomy and quality of life after brain injury. The method this thesis proposes is not new - it is established, validated, and has simply never been brought to this population. In the tradition of computational single-trial EEG analysis, a hierarchical Bayesian model of learning generates trial-wise trajectories of beliefs and precision-weighted prediction errors, and these trajectories are used to explain single-trial variation in the evoked response. The Hierarchical Gaussian Filter is explicitly constructed so that its response model can link trial-wise perceptual quantities to any observable - choices, reaction times, eye movements, or evoked response amplitude. The neural signal takes the place ordinarily occupied by the reaction time. No behavioural output is required.
This approach has an established track record on exactly the paradigms available at the bedside. Trial-wise modelling of the mismatch negativity was formalised with explicit model comparison by Lieder et al. (2013); Stefanics et al. (2018) demonstrated that single-trial ERP amplitudes track precision-weighted prediction errors derived from a Bayes-optimal agent; pharmacological studies have shown these quantities to be sensitive to NMDA-receptor antagonism. Most relevant here, the same pipeline has been used clinically and prognostically: in individuals at clinical high risk for psychosis, low-level precision-weighted prediction errors distinguished those who later converted from those who did not Structurally, that design is the one proposed here, with recovery in place of conversion. The term itself is already in use in this literature - computational phenotyping of belief updating .
What this is not. The dominant computational tradition in disorders of consciousness is whole-brain generative modelling - mean-field and neural-mass models fitted to regional dynamics, estimating coupling parameters and simulating transitions between states. That work asks how the substrate behaves. This thesis asks what the patient infers. The models are cognitive, not biophysical; the estimated parameters are perceptual and epistemic rather than physiological. The two lines are complementary, and their integration is a longer-term horizon, but they should not be confused.
Three experiments:
The thesis is built as a validation ladder: each study establishes what the next one is entitled to assume.
Study 1 - Healthy volunteers: establishing neural-only estimation
Participants perform local-global and roving oddball paradigms with a behavioural task. Inference parameters are estimated from single-trial EEG alone and benchmarked against parameters estimated from behaviour in the same participants. This is the validation step that no clinical study can provide: it is the only setting in which the neural-only estimate can be checked against a behavioural ground truth. Acquisition in Toulouse and Córdoba; modelling in Córdoba.
Study 2 - Graded sedation: characterising orderly degradation
The validated pipeline is applied to graded, reversible, dose-controlled sedation, where consciousness is lost and recovered within the same individual. The test is whether specific parameters - sensory precision, volatility learning, prior weighting - degrade in an orderly, dissociable way rather than the evoked response simply diminishing in amplitude. This study bridges the gap between a cooperative participant and a patient who cannot respond, in a population where the cause of the change is known.
Study 3 - Acute comatose patients: prognostic value
In acutely comatose patients from the Toulouse ICU cohort, the thesis tests whether latent inference parameters predict three-month neurological outcome beyond entropy-based markers, ERP-based effects and standard clinical predictors, and - the clinically decisive question - whether they resolve patients whom existing markers leave in the grey zone.
Methodological challenge:
The central technical difficulty should be stated plainly, since it is also the thesis's main methodological contribution. In most published applications, belief trajectories are generated either by a Bayes-optimal agent with fixed parameters or by a model fitted to behaviour, and are then entered as regressors on the EEG. Estimating subject-specific parameters purely from physiology, with no behavioural anchor, is a substantially harder inverse problem and remains far less established. It raises identifiability questions - which parameters are recoverable, under what signal-to-noise ratio, with how many trials - that must be settled through systematic parameter recovery on simulated and real data before any clinical claim can be made. Study 1 exists precisely to bound this problem in a population where ground truth is available.
Le profil recherché
Le projet convient à un ou une étudiant-e ayant une formation en bioingénierie, en sciences de la vie, en neurosciences ou dans un domaine connexe. Le ou la candidat-e idéal-le aura de l'expérience dans l'acquisition et l'analyse de l'EEG, un vif intérêt pour le développement méthodologique et un enthousiasme pour les données neurophysiologiques multimodales, les interactions cerveau-corps et les interfaces cerveau-ordinateur. Des compétences en traitement du signal, en programmation et/ou en apprentissage automatique seraient précieuses, tout comme la capacité de travailler en collaboration dans un environnement de recherche clinique interdisciplinaire. Une bonne connaissance de l'anglais est requise; la connaissance du français n'est pas obligatoire.
Compétences requises
- Electro-encéphalographie (EEG)
- Programmation
- Anglais
- Traitement du signal
- Machine learning