2025​​Activity reportProject-TeamBOOST​​​‌

3.1.1‌ Mathematical Analysis and Extension‌​‌ of Semi-Classical Signal Processing​​ Methods

Participants: Taous Meriem​​​‌ Laleg.

Within this‌ axis, BOOST investigates quantum-inspired‌​‌ signal processing methods based​​ on the semi-classical analysis​​​‌ of the Schrödinger operator,‌ with the objective of‌​‌ developing mathematically grounded and​​ interpretable tools for neurophysiological​​​‌ signal and image analysis‌ 62. In particular,‌​‌ the team builds on​​ the Semi-Classical Signal Analysis​​​‌ (SCSA) framework, which represents‌ signals through the spectral‌​‌ decomposition of a Schrödinger​​ operator whose potential is​​​‌ defined by the signal‌ itself. Unlike classical decomposition‌​‌ techniques relying on fixed​​ bases, SCSA provides signal-adaptive​​​‌ representations that capture fine‌ morphological variations, making it‌​‌ especially well suited for​​ pulse-shaped and continuously recorded​​​‌ physiological signals 67,‌ 64. 58,‌​‌ 63.

A first​​ research objective concerns the​​​‌ mathematical analysis of the‌ SCSA method, including the‌​‌ study of the role​​ and properties of the​​​‌ squared eigenfunctions involved in‌ the reconstruction. BOOST aims‌​‌ to better understand the​​ influence of the semi-classical​​​‌ parameter and other tuning‌ parameters on signal representation,‌​‌ and to develop systematic​​ and reliable parameter selection​​​‌ strategies 65, 68‌. This includes investigating‌​‌ links with semi-classical and​​ quantum perturbation theory, with​​​‌ the goal of improving‌ the theoretical foundations and‌​‌ robustness of the method.​​

A second objective focuses​​​‌ on the extension of‌ SCSA to two-dimensional signals‌​‌ and images, with applications​​ to image reconstruction and​​​‌ contrast enhancement. Image data‌ play a central role‌​‌ in health, movement analysis,​​ and sports, for instance​​​‌ in video-based biomechanical analysis‌ or spectrogram-based signal representations‌​‌ 61.

Finally, BOOST​​ aims to support the​​​‌ dissemination and usability of‌ these methods through the‌​‌ development of user-friendly and​​ efficient software tools for​​​‌ SCSA in both one‌ and two dimensions. While‌​‌ one-dimensional implementations are already​​ well optimized, extending the​​​‌ method to images raises‌ significant computational challenges due‌​‌ to the cost of​​ solving large-scale eigenvalue problems.​​​‌ BOOST therefore investigates advanced‌ numerical solvers to improve‌​‌ computational efficiency and scalability,​​ enabling the practical use​​​‌ of SCSA in signal‌ and image analysis applications.‌​‌

3.1.2 Non-Intrusive Monitoring of​​ Physiological Signals for Health,​​​‌ Sport, and Well-Being

Participants:‌ Ioannis Bargiotas, Arnaud‌​‌ Boutin, François Cottin​​, Taous Meriem Laleg​​​‌.

This research axis‌ addresses the development of‌​‌ non-intrusive signal monitoring methods​​ for health and sports​​​‌ performance analysis, with a‌ focus on neurophysiological and‌​‌ cardiovascular signals. The objective​​ is to better characterize​​​‌ physiological variability during exercise‌ and daily activities in‌​‌ order to support performance​​ optimization, prevent overtraining and​​​‌ injury, and assess both‌ physical and mental load.‌​‌

BOOST investigates readily available​​ signals such as EEG,​​​‌ ECG, blood pressure, and‌ PPG, as well as‌​‌ physiological variables that are​​ not directly measurable in​​​‌ practice. In this context,‌ the team develops virtual‌​‌ sensing approaches, combining signal​​ processing and artificial intelligence​​​‌ techniques, to estimate key‌ indicators such as heart‌​‌ rate variability and arterial​​ stiffness in a continuous​​​‌ and non-invasive manner.

A‌ first research focus concerns‌​‌ the non-invasive assessment of​​​‌ arterial stiffness, a major​ marker of cardiovascular health​‌ and adaptation to physical​​ activity. Since direct measurement​​​‌ of central arterial stiffness​ is impractical, BOOST aims​‌ to derive reliable stiffness​​ indicators from peripheral and​​​‌ wearable signals, including PPG​ and continuous blood pressure​‌ recordings. Advanced signal processing​​ and AI-based methods are​​​‌ investigated to improve robustness​ and accuracy, enabling the​‌ study of the relationship​​ between arterial stiffness, training​​​‌ load, fatigue, and overtraining​ in health and sports​‌ contexts.

A second research​​ focus addresses the evaluation​​​‌ of mental and physiological​ stress through the joint​‌ analysis of brain and​​ cardiovascular signals. BOOST investigates​​​‌ the definition of simple,​ continuous, and non-invasive stress​‌ indicators based on the​​ simultaneous recording of EEG​​​‌ and cardiovascular signals under​ different cognitive and physiological​‌ conditions. Signal processing and​​ AI methods are combined​​​‌ to extract relevant features​ and analyze brain–heart interactions,​‌ with the objective of​​ supporting objective stress assessment​​​‌ in both health and​ sports settings.

Finally, BOOST​‌ aims to contribute to​​ the development of a​​​‌ monitoring framework for health​ and sports applications, based​‌ on continuous recordings of​​ non-invasive signals which play​​​‌ a key role in​ training monitoring and personalization​‌ but are strongly affected​​ by movement artifacts and​​​‌ activity-dependent variability. BOOST therefore​ focuses on designing robust​‌ signal processing tools and​​ indices that remain reliable​​​‌ under realistic acquisition conditions,​ supported by both public​‌ datasets and dedicated experimental​​ data collection.

3.2.1​‌ Modeling Bio-Informed Systems

Participants:​​ Arnaud Boutin, Bastien​​​‌ Berret, François Cottin​, Taous Meriem Laleg​‌.

BOOST develops mathematical​​ models for bio-informed systems​​​‌ involving humans, with the​ objective of supporting the​‌ interpretation of physiological signals,​​ the inference of internal​​​‌ states, and the design​ of estimation and control​‌ strategies. These models provide​​ structured representations of physiological​​​‌ and neuromechanical processes and​ form a foundation for​‌ monitoring, prediction, and biofeedback​​ in health and sports​​​‌ applications.

Bio-informed systems are​ characterized by nonlinear, uncertain,​‌ and heterogeneous dynamics, as​​ well as multiscale behaviors,​​ delays, and partial observability.​​​‌ Classical modeling approaches may‌ be insufficient to capture‌​‌ these properties. BOOST therefore​​ investigates advanced modeling frameworks​​​‌ that preserve physiological interpretability‌ while remaining compatible with‌​‌ estimation and control design.​​

A first modeling focus​​​‌ concerns brain-heart interactions, with‌ the objective of improving‌​‌ the understanding of mental​​ and physiological stress mechanisms.​​​‌ Stress responses involve complex,‌ bidirectional coupling between neural‌​‌ and cardiovascular systems 20​​. BOOST investigates model-based​​​‌ and data-informed approaches to‌ represent these interactions and‌​‌ to support the definition​​ of simple, continuous, and​​​‌ non-invasive indicators of mental‌ and physiological state.

A‌​‌ second modeling focus addresses​​ human neuromechanical control, which​​​‌ plays a central role‌ in movement execution, stability,‌​‌ and adaptation 57,​​ 56, 55.​​​‌ BOOST investigates mathematical models‌ describing the interplay between‌​‌ anticipatory (open-loop) and reactive​​ (closed-loop) control mechanisms in​​​‌ the human motor system.‌ These models account for‌​‌ nonlinear musculoskeletal dynamics, delays,​​ noise, and inter-individual variability,​​​‌ and aim to predict‌ both average motor behavior‌​‌ and its variability. Such​​ modeling efforts provide a​​​‌ basis for improved identification‌ methods and for the‌​‌ design of adaptive control​​ strategies in human-robot interaction​​​‌ systems.

3.2.2 Estimation and‌ Observer Design for Bio-Informed‌​‌ Systems

Participants: Bastien Berret​​, Taous Meriem Laleg​​​‌.

Estimation plays a‌ central role in bio-informed‌​‌ modeling and control, as​​ many physiological, biomechanical, and​​​‌ neuromechanical variables of interest‌ are not directly accessible‌​‌ to measurement. In health,​​ movement, and sports applications,​​​‌ estimation is required to‌ calibrate models, infer internal‌​‌ states and parameters, and​​ support control and decision-making,​​​‌ often under conditions of‌ uncertainty, noise, partial observability,‌​‌ and inter-individual variability.

BOOST​​ focuses on the development​​​‌ of estimation methods and‌ observer-based algorithms for a‌​‌ wide range of dynamical​​ systems, including finite and​​​‌ infinite-dimensional models 72,‌ 54, 73.‌​‌ Particular attention is given​​ to observer design, which​​​‌ provides a structured and‌ interpretable framework for state‌​‌ and parameter estimation in​​ systems governed by ordinary​​​‌ or partial differential equations,‌ and can be naturally‌​‌ integrated with control strategies​​ and virtual sensing approaches.​​​‌

A first research direction‌ addresses the design of‌​‌ asymptotic and non-asymptotic estimators,​​ including both classical observers​​​‌ with asymptotic convergence and‌ finite-time estimation methods. Finite-time‌​‌ approaches are particularly appealing​​ in practical contexts, as​​​‌ they provide fast convergence‌ and robustness to unknown‌​‌ initial conditions and noise​​ 53, 59,​​​‌ 69, 60.‌ BOOST investigates such methods‌​‌ for nonlinear and distributed​​ systems, with an emphasis​​​‌ on convergence analysis, robustness‌ properties, and numerical implementation.‌​‌

A second research direction​​ focuses on learning-based observer​​​‌ frameworks, which combine data-driven‌ components with model-based estimation.‌​‌ These approaches aim to​​ overcome limitations of classical​​​‌ observer design when models‌ are incomplete, uncertain, or‌​‌ affected by unknown disturbances​​ 74. 15,​​​‌ 71. BOOST investigates‌ hybrid learning observers that‌​‌ retain the structure and​​ feedback properties of observers​​​‌ while exploiting learning mechanisms‌ to approximate unknown dynamics,‌​‌ adapt observer gains, or​​ self-tune estimation parameters. A​​​‌ particular emphasis is placed‌ on robustness, convergence, and‌​‌ generalizability, in order to​​​‌ move beyond purely application-driven​ learning approaches.

Finally, BOOST​‌ investigates inverse estimation approaches,​​ where the objective is​​​‌ to infer hidden internal​ quantities such as control​‌ objectives or cost function​​ parameters from observed behavior​​​‌ 66, 70.​ In particular, inverse optimal​‌ control and inverse differential​​ game formulations are explored​​​‌ to estimate the goals​ and strategies underlying human​‌ movement and interaction. These​​ approaches provide a compact​​​‌ and interpretable representation of​ complex behaviors and are​‌ especially relevant in human-robot​​ interaction scenarios, where predicting​​​‌ user intent and adaptation​ is essential for personalized​‌ assistance and control.

Together,​​ these research directions contribute​​​‌ to the development of​ robust, adaptive, and interpretable​‌ estimation methodologies that bridge​​ modeling, learning, and control,​​​‌ and that are well​ suited to the challenges​‌ of bio-informed human-in-the-loop systems.​​

3.3.1 Combined​‌ AI and Control-Theoretic Design​​

Participants: Michel-Ange Amorim,​​​‌ Bastien Berret, Taous​ Meriem Laleg.

BOOST​‌ investigates control methodologies that​​ combine artificial intelligence with​​​‌ control theory, aiming to​ leverage the complementary strengths​‌ of model-based and data-driven​​ approaches. In particular, the​​​‌ team explores hybrid frameworks​ that integrate optimal control​‌ and model predictive control​​ (MPC) with reinforcement learning​​​‌ (RL). MPC provides robustness,​ stability, and constraint handling,​‌ while learning-based approaches offer​​ adaptability and the ability​​​‌ to cope with modeling​ uncertainties and inter-individual variability.​‌

BOOST studies control architectures​​ in which learning mechanisms​​​‌ are used to complement​ MPC, for instance by​‌ adapting cost functions, reward​​ structures, or control parameters,​​​‌ while preserving control-theoretic guarantees.​ These approaches are particularly​‌ relevant for personalized biofeedback​​ systems, where control laws​​​‌ must adapt to individuals​ with different physiological states,​‌ health conditions, or performance​​ objectives.

Within this context,​​​‌ BOOST also explores links​ with dynamic treatment regimes,​‌ which formalize sequential decision-making​​ strategies based on individual​​​‌ histories and outcomes. Such​ formulations naturally connect reinforcement​‌ learning and control and​​ provide a principled framework​​​‌ for personalization in health​ and sports applications, including​‌ time-dependent decisions, delayed effects,​​ and multi-objective optimization 75​​​‌.

3.3.2 Movement Assistance,​ Training, and Rehabilitation

BOOST​‌ also focuses on control​​ design for movement assistance,​​​‌ training, and rehabilitation, with​ a particular emphasis on​‌ human-robot interaction systems such​​ as assistive exoskeletons. In​​​‌ these applications, real-time constraints,​ limited data availability, and​‌ human variability pose major​​ challenges to control design​​​‌ 57, 56.​

To address these issues,​‌ BOOST investigates approaches that​​ combine model-based representations of​​​‌ human neuromechanical control with​ learning-based methods. In particular,​‌ simulated or artificial motion​​ data generated from biomechanical​​ models can be used​​​‌ to learn compact representations‌ of movement variability offline,‌​‌ enabling efficient real-time control.​​ Learning-based motion representations, such​​​‌ as probabilistic movement primitives,‌ are explored to predict‌​‌ short-horizon human intent and​​ support adaptive assistance.

BOOST​​​‌ further investigates interaction control‌ strategies adapted to different‌​‌ objectives, including assistance, training,​​ and rehabilitation. Both trajectory-based​​​‌ approaches, which exploit predictions‌ of upcoming motion, and‌​‌ trajectory-free approaches, which rely​​ on force, torque, or​​​‌ EMG feedback, are considered.‌ Game-theoretic formulations provide a‌​‌ natural framework to model​​ shared effort and cooperation​​​‌ between humans and robots,‌ while human-in-the-loop optimization approaches‌​‌ offer model-free alternatives for​​ controller tuning based on​​​‌ physiological or biomechanical performance‌ criteria.

Together, these contributions‌​‌ aim to develop versatile,​​ adaptive, and personalized control​​​‌ strategies that integrate modeling,‌ estimation, learning, and interaction,‌​‌ enabling effective assistance and​​ training while accounting for​​​‌ human adaptation and variability.‌

7.1.1 pyscsa​​​‌

• Name:
  Semi-classical Signal Analysis‌
• Keywords:
  Signal processing, Spectral‌​‌ analysis, Schrödinger equation, Denoising​​
• Functional Description:
  PySCSA is​​​‌ an advanced library that‌ implements the Semi-Classical Signal‌​‌ Analysis (SCSA) method for​​ signal and image processing.​​​‌ It generates data-adaptive atoms,‌ enabling accurate reconstruction, effective‌​‌ denoising, and highly robust​​ feature extraction. PySCSA also​​​‌ includes dedicated tools for‌ optimizing SCSA parameters, making‌​‌ the method efficient and​​ intuitive to use for​​​‌ both research and practical‌ applications.
• Contact:
  Taous Laleg‌​‌ Kirati

8.1.1 Image contrast enhancement‌​‌ using the SCSA

Participants:​​ Taous Meriem Laleg,​​​‌ Juan Manuel Vargas Garcia‌.

A novel image‌​‌ contrast enhancement method based​​ on the spectral properties​​​‌ of the Schrödinger operator‌ is proposed in 49‌​‌, extending the semi-classical​​ signal analysis (SCSA) framework​​​‌ to two-dimensional image data.‌ The proposed approach formulates‌​‌ image contrast enhancement as​​ a spectral manipulation problem,​​​‌ where image intensity is‌ treated as the potential‌​‌ of a Schrödinger operator​​ and the enhanced image​​​‌ is reconstructed from its‌ negative eigenvalues and associated‌​‌ eigenfunctions. Unlike conventional contrast​​ enhancement techniques, the method​​​‌ is adaptive to image‌ content and preserves fine‌​‌ structural details while enhancing​​ contrast. The approach is​​​‌ validated on synthetic and‌ real images, demonstrating improved‌​‌ visual quality and robustness​​ to noise, and highlighting​​​‌ the potential of quantum-inspired‌ spectral methods for advanced‌​‌ image processing applications in​​​‌ health and biomedical imaging.​

8.1.2 Modeling, Representation Learning,​‌ and Robust Processing of​​ Neurophysiological Signals

Participants: Israel​​​‌ Jesus Santos Filho,​ Juan Manuel Vargas Garcia​‌, Jiahao Hu,​​ Ege Kibrislioglu, Taous​​​‌ Meriem Laleg.

This​ research axis addresses the​‌ development of signal processing​​ and AI-based methods for​​​‌ modeling and analyzing neurophysiological​ signals acquired in realistic​‌ and unconstrained conditions. The​​ work focuses on structured​​​‌ signal representations, long-range temporal​ modeling, and robustness to​‌ noise and motion artifacts,​​ which are essential for​​​‌ the reliable exploitation of​ neurophysiological signals in health​‌ and sport applications.

Long-range​​ and structured modeling of​​​‌ EEG and ECG signals:​ Mamba-based sequence models combined​‌ with dynamic graph learning​​ and channel attention mechanisms​​​‌ were developed to capture​ long-range temporal dependencies and​‌ inter-channel relationships in multichannel​​ EEG and ECG signals.​​​‌ These approaches provide flexible​ representations of spatiotemporal dynamics​‌ and improve modeling and​​ classification performance in tasks​​​‌ such as sleep staging​ 25 and physiological signal​‌ modeling 13.

Robust​​ processing of EEG signals​​​‌ under motion: A deep​ learning-based approach, termed Motion-Net,​‌ was proposed for the​​ removal of motion artifacts​​​‌ in EEG recordings. The​ method improves signal quality​‌ in the presence of​​ subject movement, enabling more​​​‌ reliable analysis of EEG​ data acquired in ambulatory​‌ and wearable settings 14​​.

Graph-based feature extraction​​​‌ from photoplethysmographic signals: Weighted​ visibility graph representations were​‌ introduced to extract robust​​ features from photoplethysmographic signals.​​​‌ These representations enable the​ reliable estimation of cardiovascular​‌ indicators, such as pulse​​ wave velocity, and remain​​​‌ effective under signal variability​ and noise 23.​‌

Digital twin modeling of​​ photoplethysmographic signals from wireless​​​‌ sensing:

A neural representation​ framework was developed to​‌ synthesize photoplethysmographic signals from​​ 6G/WiFi integrated sensing and​​​‌ communication signals, enabling the​ construction of Radio-PPG digital​‌ twins and supporting contactless​​ physiological monitoring 39.​​​‌

8.1.3 Physiological State Estimation​ and Health Monitoring from​‌ Wearable, Mobile, and Ambient​​ Sensors

Participants: Tareq Alnafouri​​​‌ , Taous Meriem Laleg​, Juan Manuel Vargas​‌ Garcia.

This research​​ axis focuses on the​​​‌ estimation of physiological states​ and health-related indicators from​‌ signals acquired using wearable,​​ mobile, and contactless sensing​​​‌ technologies. The work aims​ at translating low-level neurophysiological​‌ and cardiovascular signals into​​ meaningful indicators relevant for​​​‌ health monitoring, sports performance,​ and well-being, using signal​‌ processing and AI-based estimation​​ techniques.

Noninvasive hydration monitoring​​​‌ using wearable and mobile​ sensors: In collaboration with​‌ Prof Tareq Alnafouri 's​​ team, signal processing and​​​‌ machine learning methods were​ developed to estimate hydration​‌ status in fasting and​​ sports contexts using skin​​​‌ capacitance measurements acquired from​ wearable sensors 21.​‌ In parallel, a smartphone-based​​ approach relying on camera-derived​​​‌ signals was proposed to​ assess hydration levels in​‌ a fully noninvasive manner.​​ These works demonstrate the​​​‌ feasibility of continuous hydration​ monitoring using low-cost and​‌ widely available sensing modalities​​ 8.

AI-based analysis​​​‌ of medical imaging for​ clinical classification: A hybrid​‌ deep transfer learning framework​​ was developed for the​​​‌ classification of symptomatic and​ asymptomatic carotid plaques from​‌ CT images. The proposed​​ approach integrates feature extraction​​ and classification stages and​​​‌ demonstrates the potential of‌ AI-based imaging analysis for‌​‌ supporting clinical decision-making 28​​.

8.2.1 Models design‌

Participants: Arnaud Boutin,‌​‌ François Cottin, Taous​​ Meriem Laleg, Maria​​​‌ Sara Nour Sadoun.‌

Modeling brain-heart interactions: A‌​‌ comprehensive review of mechanistic​​ dynamical models describing brain-heart​​​‌ interactions was conducted, with‌ the objective of supporting‌​‌ the modeling of coupled​​ neurocardiovascular systems relevant to​​​‌ mental and physical health‌ monitoring. The study analyzes‌​‌ existing linear and nonlinear​​ modeling frameworks used to​​​‌ capture bidirectional interactions between‌ neural and cardiovascular dynamics,‌​‌ as well as their​​ ability to represent physiological​​​‌ mechanisms and multiscale behaviors.‌ By identifying current limitations‌​‌ and open challenges, this​​ work provides a structured​​​‌ foundation for the development‌ of bio-informed models enabling‌​‌ the inference of physiological​​ and mental states from​​​‌ jointly acquired brain and‌ cardiovascular signals 20.‌​‌

8.2.2 Learning-Based Observer Design​​ for Nonlinear Dynamical Systems​​​‌

Participants: Mohamed Boukaf,‌ Zehor Belkhatir, Mohamed‌​‌ Chadli, Taous Meriem​​ Laleg, Yasmine Marani​​​‌ .

Deep learning-based extensions‌ of the Kazantzis-Kravaris-Luenberger (KKL)‌​‌ observer were developed for​​ discrete-time nonlinear systems subject​​​‌ to time-varying output delays,‌ enabling accurate state reconstruction‌​‌ under delayed and partial​​ measurements. Neural networks are​​​‌ used to approximate observer‌ dynamics while maintaining the‌​‌ observer formulation 15.​​

These learning-based KKL observers​​​‌ were further applied to‌ a variety of estimation‌​‌ problems involving complex nonlinear​​ dynamics and limited sensing.​​​‌ In particular, they were‌ used for state estimation‌​‌ in mixed-autonomy traffic systems,​​ enabling the reconstruction of​​​‌ vehicle position and speed‌ from heterogeneous and incomplete‌​‌ measurements 26. In​​ parallel, a combined deep​​​‌ KKL observer and nonlinear‌ model predictive control (NMPC)‌​‌ framework was proposed for​​ the modeling of muscle​​​‌ co-contraction under measurement delays,‌ where the observer provides‌​‌ real-time estimates of internal​​ muscle states that are​​​‌ exploited within a control‌ scheme 35. Together,‌​‌ these contributions illustrate how​​ learning-enhanced observer design can​​​‌ address estimation problems in‌ complex nonlinear systems that‌​‌ are beyond the reach​​ of classical observer approaches.​​​‌

8.2.3 Scientific learning-based observer‌ design

Participants: Mohamed Boukaf‌​‌, Israel Jesus Santos​​ Filho, Taous Meriem​​​‌ Laleg, Yasmine Marani‌.

Complementing the observer-based‌​‌ approaches presented above, BOOST​​ investigates physics-informed and learning-control​​​‌ estimation methods for nonlinear‌ systems, with the objective‌​‌ of estimating states, parameters,​​ and unknown inputs from​​​‌ sparse and noisy measurements.‌ While observer-based frameworks rely‌​‌ on specific dynamical embeddings,​​ these approaches address estimation​​​‌ problems for more general‌ nonlinear system structures, for‌​‌ which classical observer design​​​‌ often requires restrictive assumptions.​

In this context, unsupervised​‌ physics-informed neural networks (PINNs)​​ were developed for blood​​​‌ flow estimation and cardiovascular​ parameter assessment, by embedding​‌ physiological models, such as​​ Windkessel representations, directly into​​​‌ the learning process. Parallel​ PINN architectures were introduced​‌ to improve scalability and​​ robustness when dealing with​​​‌ heterogeneous data sources and​ limited measurements 32.​‌

Additionnally, learning-based nonlinear observers​​ incorporating contraction-inspired principles were​​​‌ proposed to address estimation​ problems for general nonlinear​‌ system structures, thereby relaxing​​ the strong structural conditions​​​‌ typically required by classical​ nonlinear observer design. By​‌ ensuring contracting estimation dynamics,​​ these approaches facilitate stable​​​‌ state reconstruction and enable​ the integration of learning-based​‌ estimators within control-oriented frameworks​​ 27.

8.2.4 Physics-Informed​​​‌ Estimation of Autonomic Cardiac​ Dynamics

Participants: Arnaud Boutin​‌, François Cottin,​​ Taous Meriem Laleg,​​​‌ Maria Sara Nour Sadoun​.

BOOST has obtained​‌ new results on the​​ estimation of internal states​​​‌ and unknown inputs in​ autonomic cardiac dynamics, addressing​‌ key challenges related to​​ delayed dynamics, partial observability,​​​‌ and physiological interpretability. These​ works introduce physics-informed neural​‌ estimation frameworks that explicitly​​ incorporate left-invertibility constraints, a​​​‌ structural property that guarantees​ the identifiability of states​‌ and unknown inputs from​​ available measurements.

The proposed​​​‌ approaches combine neural network-based​ estimators with physiological dynamical​‌ models of cardiac autonomic​​ regulation, enabling the simultaneous​​​‌ estimation of hidden cardiac​ states and unmeasured autonomic​‌ inputs. By embedding physical​​ constraints directly into the​​​‌ learning process, the methods​ overcome limitations of purely​‌ data-driven estimators, ensuring consistency​​ with cardiovascular dynamics and​​​‌ improving robustness to delays​ and modeling uncertainties 47​‌.

The second contribution​​ extends this framework to​​​‌ delayed autonomic cardiac systems,​ which are particularly challenging​‌ due to time-lagged responses​​ and coupling between neural​​​‌ and cardiovascular processes. The​ results demonstrate that enforcing​‌ left-invertibility within the neural​​ estimator allows reliable state​​​‌ and input reconstruction despite​ delays, providing a principled​‌ solution for non-invasive cardiac​​ monitoring 48.

Together,​​​‌ these works contribute novel​ methodologies at the intersection​‌ of estimation theory, physiological​​ modeling, and machine learning,​​​‌ and demonstrate the potential​ of physics-informed neural estimators​‌ for interpretable and reliable​​ cardiovascular monitoring in health​​​‌ and sports applications.

8.3.1​​ Control Design for Human-Robot​​​‌ Interaction and Assistive Systems​

Participants: Bastien Berret,​‌ Mohamed Boukaf, Taous​​ Meriem Laleg.

BOOST​​ develops control frameworks for​​​‌ systems involving direct interaction‌ between humans and robotic‌​‌ or neuromuscular systems, with​​ an emphasis on adaptation,​​​‌ safety, and interaction-aware control.‌ In such systems, control‌​‌ laws must regulate system​​ dynamics while accounting for​​​‌ uncertainty, delays, and coupling‌ effects arising from human‌​‌ interaction.

In this context,​​ a combined deep KKL​​​‌ observer and nonlinear model‌ predictive control (NMPC) framework‌​‌ was proposed to model​​ and regulate muscle co-contraction​​​‌ under measurement delays. The‌ learning-based observer provides real-time‌​‌ estimates of system states​​ and/or internal muscle states,​​​‌ which are exploited within‌ a predictive control scheme‌​‌ to handle delayed measurements​​ and complex neuromuscular dynamics.​​​‌ This contribution illustrates how‌ estimation and control can‌​‌ be tightly integrated to​​ address human-in-the-loop control problems​​​‌ where having predictive models‌ of human sensorimotor control‌​‌ are crucial 35.​​

8.3.2 Learning-Based and Game-Theoretic​​​‌ Control for Personalized Assistance‌

Participants: Bastien Berret,‌​‌ Abdelwaheb Hafs.

To​​ address inter-individual variability and​​​‌ shared control objectives, BOOST‌ investigates learning-based and game-theoretic‌​‌ control approaches for personalized​​ assistance. These methods aim​​​‌ to balance assistance and‌ user autonomy by explicitly‌​‌ modeling the interaction between​​ the human and the​​​‌ assistive system.

Game-theoretic formulations‌ were introduced to describe‌​‌ human-robot interaction as a​​ dynamic game, enabling the​​​‌ design of model predictive‌ game control strategies for‌​‌ targeted and personalized assistance.​​ These approaches explicitly account​​​‌ for the coupled volitional‌ control processes of the‌​‌ human and the robotic​​ system 29, 42​​​‌.

In parallel, learning-based‌ control methods were developed‌​‌ to exploit human movement​​ variability for improved assistance.​​​‌ Probabilistic movement primitives and‌ artificially generated demonstrations were‌​‌ used to enhance exoskeleton​​ flow control and adapt​​​‌ assistance to diverse movement‌ patterns. Early intent classification‌​‌ from kinematic data was​​ also investigated to anticipate​​​‌ human motion and adjust‌ control actions accordingly 43‌​‌, 44, 45​​. In addition, EMG-based​​​‌ control methods have been‌ developed by estimating human‌​‌ muscle forces via learning​​ methods and using those​​​‌ estimation in adaptive controllers‌ to improve movement assistance‌​‌ using robotic exoskeletons 18​​, 17.

8.4.1​​ Stress and Neurocardiovascular Interaction​​​‌ Monitoring

Participants: Arnaud Boutin‌, François Cottin,‌​‌ Taous Meriem Laleg,​​ Maria Sara Nour Sadoun​​​‌.

BOOST has obtained‌ new results in the‌​‌ objective monitoring of mental​​ and physiological stress through​​​‌ the joint analysis of‌ neurophysiological and cardiovascular signals.‌​‌ Dedicated experiments were conducted​​ in a controlled laboratory​​​‌ environment, involving 30 subjects‌ exposed to three complementary‌​‌ stressors:

• physical stress,
• physiological​​ stress induced by a​​​‌ cold pressor test, and‌
• cognitive stress induced by‌​‌ a singing-to-singing task.

During​​ these experiments, EEG, ECG,​​​‌ and arterial pressure signals‌ were recorded simultaneously, resulting‌​‌ in a multimodal dataset​​ specifically designed to study​​​‌ brain-heart interactions under stress.‌ This dataset has been‌​‌ used to validate signal​​ processing and AI-based methods​​​‌ for extracting stress-related features‌ and indicators.

The proposed‌​‌ approaches enable the association​​​‌ of low-level signal characteristics​ with higher-level markers of​‌ mental and physiological load,​​ contributing to objective stress​​​‌ assessment. These results complement​ subjective evaluations, and demonstrate​‌ the relevance of multimodal,​​ bio-informed methodologies for stress​​​‌ monitoring.

8.4.2 Prediction and​ Personalized Monitoring of Vascular​‌ Diseases

Participants: Jean Michel​​ Davaine, Juan Manuel​​​‌ Vargas Garcia, Taous​ Meriem Laleg.

BOOST​‌ has also produced results​​ in applications related to​​​‌ the prediction and personalized​ monitoring of vascular diseases,​‌ providing clinically relevant validation​​ of its methodologies.

A​​​‌ first line of work​ addresses the prediction of​‌ vulnerable carotid plaques, a​​ major risk factor for​​​‌ ischemic stroke. Hybrid AI-based​ image analysis methods, combining​‌ deep learning and transfer​​ learning techniques, have been​​​‌ applied to vascular CT​ images to distinguish between​‌ symptomatic and asymptomatic carotid​​ plaques. These results demonstrate​​​‌ the ability of data-driven​ approaches to capture subtle​‌ morphological features associated with​​ plaque vulnerability, despite strong​​​‌ inter-patient variability 28.​

In parallel, BOOST has​‌ contributed to the estimation​​ of pulse wave velocity​​​‌ (PWV) from photoplethysmographic signals,​ enabling non-invasive assessment of​‌ arterial stiffness, a key​​ biomarker of vascular health​​​‌ 23. These signal-based​ approaches complement imaging-based plaque​‌ characterization and illustrate the​​ benefits of combining multimodal​​​‌ vascular indicators for improved​ risk stratification.

These results​‌ highlight the potential of​​ bio-informed signal processing and​​​‌ AI-based tools to support​ personalized monitoring and prevention​‌ strategies in vascular diseases.​​

8.4.3 Human Movement, Muscle​​​‌ Co-Contraction, and Assistance

Participants:​ Bastien Berret, Mohamed​‌ Boukaf, Taous Meriem​​ Laleg.

A second​​​‌ major set of results​ concerns human movement analysis​‌ and neuromuscular function, with​​ a particular focus on​​​‌ muscle co-contraction, which plays​ a critical role in​‌ movement control, stability, and​​ injury prevention. Muscle co-contraction​​​‌ involves internal states that​ are difficult to measure​‌ directly and is affected​​ by delays, nonlinearities, and​​​‌ strong coupling between neural​ and muscular dynamics.

BOOST​‌ has validated its methodologies​​ in this context by​​​‌ combining modeling, observer-based estimation,​ and predictive control to​‌ infer muscle co-contraction dynamics​​ from partial and delayed​​​‌ measurements. These results demonstrate​ the effectiveness of learning-enhanced​‌ observers and control strategies​​ in human-in-the-loop systems, where​​​‌ accurate estimation of internal​ states is essential for​‌ assistance and adaptation.

8.4.4​​ Sports Performance and Injury​​​‌

Participants: Ioannis Bargiotas,​ François Cottin, François​‌ Saulnier.

BOOST proposed​​ a novel approach for​​​‌ injury prevention in athletes​ through artificial intelligence frameworks.​‌ The primary objective is​​ to develop a longitudinal​​​‌ machine learning framework for​ multimodality settings to predict​‌ future injury risk from​​ repeated training reports, under​​​‌ realistic conditions of heterogeneity,​ completely missing modalities, and​‌ inter-individual variability that challenge​​ the generalization and interpretability​​​‌ of complex models.

Setting​ the core ideas of​‌ the analysis, the project​​ emphasize the utility of​​​‌ the Rashomon effect by​ explicitly exploring sets of​‌ quasi-optimal longitudinal models, thereby​​ capturing multiple plausible explanations​​​‌ of injury dynamics consistent​ with both data and​‌ various expert knowledge. Extending​​ already developed packages, BOOST​​​‌ proposition (TREEFARMS+) 31 improves​ scalability, supports flexible evaluation​‌ metrics suited to imbalanced​​ injury data, and incorporate​​ interactive visualization tools to​​​‌ enable expert-in-the-loop model exploration.‌

Experiments on synthetic and‌​‌ real-world datasets demonstrate robust​​ classification performance, approaching that​​​‌ of state-of-the-art. Feature importance‌ analyses confirm the method's‌​‌ ability to recover meaningful​​ predictors despite discretization, and​​​‌ athlete-level analyses reveal systematic,‌ individualized misclassification patterns that‌​‌ may reflect chronic overload​​ or prolonged high-risk phases​​​‌ rather than model failure.‌ Overall, TREEFARMS+ on the‌​‌ one hand, enhances scalability,​​ interpretability, and usability of​​​‌ Rashomon set analysis in‌ longitudinal settings, and on‌​‌ the other hand empowers​​ domain experts and users​​​‌ to dynamically make their‌ trade-offs between performance and‌​‌ simplicity. Although validated in​​ sports science, the framework​​​‌ is domain-agnostic and particularly‌ useful in multidiscuplinary studies.‌​‌ Future works targeting the​​ extension of Rashomon concepts​​​‌ to sequential and attention-based‌ models to more explicitly‌​‌ capture temporal dependencies.

8.4.5​​ Performance and Well being​​​‌ and Mental Workload

Participants:‌ Ioannis Bargiotas.

Ocular‌​‌ analysis has been proposed​​ as a powerful approach​​​‌ for understanding cognitive states‌ in complex human–machine or‌​‌ human-environment interaction contexts, such​​ as mental workload (MWL),​​​‌ decision making, and well-being‌ . BOOST team actively‌​‌ participated in recent work​​ on MWL prediction under​​​‌ realistic conditions (operationally speaking).‌ The results from these‌​‌ efforts showed that data-driven​​ models can estimate mental​​​‌ workload in real time,‌ showing that task- and‌​‌ operation-related behavioral signals may​​ be sometimes more informative​​​‌ than well-known physiological measures‌ 19. Moreover, advancements‌​‌ in eye-tracking devices, have​​ enabled the investigation of​​​‌ gaze behavior beyond laboratory‌ conditions. This investigation highlighted‌​‌ the close link between​​ oculomotor behaviour, body motion,​​​‌ and environmental interaction, information‌ extremely important in many‌​‌ special contexts (high-level sports,​​ battlefields etc.). Generally, the​​​‌ latter developments enhanced the‌ importance of moving from‌​‌ static or aggregated metrics​​ toward continuous dynamic characterization​​​‌ of longitudinal ocular activity.‌

9.1.1‌ Participation in other International‌​‌ Programs

• Bastien Berret:​​ Collaborative Research in Computational​​​‌ Neuroscience (CRCNS), ANR/NSF, France-US‌ [NEUROPT project]
• Taous Meriem‌​‌ Laleg: Collaboration with​​ Dr. Hacene Serrai from​​​‌ Carle Health in the‌ US.
• Israel Jesus Santos‌​‌ Filho: in cotutelle​​ agreement with King Abdullah​​​‌ University of Science and‌ Technology, co-advised by Prof.‌​‌ Tareq Alnafouri .
• Mohamed​​ Boukaf: PhD co-advised​​​‌ by Prof. Zehor Belkhatir‌ from Southampton University, UK.‌​‌
• Elham Rostami : PhD​​ coadvised by Prof. Hamidou​​​‌ Tembine from Université du‌ Québec à Trois-Rivières, Canada.‌​‌

9.1.2 Visits of international​​ scientists

9.1.3 Visits to international​ teams

10.1.1​​​‌ Scientific events: organisation

• Taous​ Meriem Laleg: Organization​‌ of two invited open-track​​ IFAC World Congress, August​​​‌ 2026.
• Michel-Ange Amorim:​ Organization of FéDeV Day​‌ (Demenÿ-Vaucanson Federation of Movement​​ Sciences), Friday, November 14,​​​‌ 2025, Inria Center in​ Saclay (Alan Turing Building),​‌ approximately one hundred participants.​​

10.1.2 Scientific events: selection​​​‌

10.1.3​​ Journal

10.1.4‌​‌ Invited talks

• Taous Meriem​​ Laleg: Keynote speaker,​​​‌ "Semi-Classical Signal Analysis: From‌ Theory to Biomedical Signal‌​‌ and Image Processing", ICAECCS'2025,​​ Université Blida 1, Algeria,​​​‌ December 9-10, 2025.
• Arnaud‌ Boutin: Sleep rhytms‌​‌ underlying motor memory consolidation.​​ Advances in Motor Learning​​​‌ II, University of Birmingham,‌ Birmingham, 11-12 décembre 2025.‌​‌
• Taous Meriem Laleg:​​ Learning-based observer design for​​​‌ dynamical systems estimation, EUReCA‌ day, Centrale Supélec, December‌​‌ 3rd, 2025.
• Bastien Berret​​: Stochastic optimal feedforward-feedback​​​‌ control as a theory‌ of human sensorimotor control,‌​‌ Workshop on Healthcare Applications​​ of Interactive Robotics, King's​​​‌ College London, November 2025.‌
• Bastien Berret: Modeling‌​‌ human sensorimotor control for​​ optimizing human-robot interaction, R4​​​‌ conference, Nov. 2025, virtual‌ meeting.
• Taous Meriem Laleg‌​‌: Invited speaker, "Advanced​​ Observer Design for Nonlinear​​​‌ Systems: From KKL to‌ Contraction Theory", International League‌​‌ of Mathematics, Bauhausuniversität Weimar,​​ October 2025.
• Taous Meriem​​​‌ Laleg: Invited speaker,‌ "Computer-aided assessment of the‌​‌ Cortico-Cardiovascular Interaction for monitoring​​ the stress of athletes​​​‌ ", Insep-Inria agreement signature,‌ Insep, October 2025.
• Bastien‌​‌ Berret « The Cost​​ of Time in Action​​​‌ », CNRS Summer School‌ “Le temps dans tous‌​‌ ses états”, September 2025,​​ Hyères-Porquerolles.
• Bastien Berret «​​​‌ Understanding human motor control‌ through optimality principles »,‌​‌ Mini symposium on neural​​ control of movement, University​​​‌ of Leeds, June 2025.‌
• Boukaf Mohamed: Congress‌​‌ speaker, "Windkessel model-based Physics​​ Informed Neural Networks for​​​‌ Blood Flow Estimation and‌ Cardiovascular Parameters Assessment", SAGIP'2025,‌​‌ Mulhouse, France, May 2025.​​

10.1.5 Scientific expertise

10.1.6​​ Research administration

• Michel-Ange Amorim​​​‌ Ciams Laboratory Vice President‌
• Bastien BerretVice-dean for‌​‌ research (Faculty of Sport​​ Sciences)

Ioannis Bargiotas

• Introduction to​​ Machine learning, Level:M2 ,​​​‌ ISMH (Ingénierie, Data Science,‌ Mouvement Humain)

10.2.1 Supervision‌​‌

Taous Meriem Laleg

PhD​​ students

• Elham Rostami (2025-),​​​‌ Université Paris-Saclay, Optimal control‌ for robust and stable‌​‌ generative AI (50%).
• Abdelbaki​​ Guir (2025-), Inria, Modeling​​​‌ the brain-heart interaction for‌ the Assessment of Athlete‌​‌ stress (50%).
• Israel J​​ Jesus Santos Filho (2024-),​​​‌ Université Paris-Saclay/Cotutelle KAUST, Biosignals‌ analysis and characterization using‌​‌ signal processing and AI​​​‌ (50%).
• Juan Manuel Vargas​ (2024-), Inria, Signal processing​‌ based on the squared​​ eigenfunctions of the Schrodinger​​​‌ operator: Mathematical analysis and​ application to the identification​‌ of vulnerable carotid plaques​​ (50%).
• Mohamed Boukaf (2024-),​​​‌ Université Paris-Saclay, State Estimation​ for Nonlinear Dynamical Uncertain​‌ System using Combined Neural​​ Network and Observer Design​​​‌ (40%).
• Maria Sara Nour​ Sadoun (2023-), Université Paris-Saclay,​‌ Computer aided characterization of​​ the corticocardiovascular interaction (40%).​​​‌
• Jiahao Hu (2021-2025), KAUST,​ Contributions to classification of​‌ biomedical signals (50%).
• Yasmine​​ Marani (2021-2025), KAUST, Contribution​​​‌ to observer design for​ nonlinear dynamical systems (50%).​‌
• Fahad Aljehani (2020-2025), KAUST,​​ Contribution to learning-based control​​​‌ and estimation (50%).

Master​ students

• Elham Rostami (2025),​‌ M2, Université Paris-Saclay, Seizure​​ onset zone localization using​​​‌ EEG signals.
• Abdelbaki Guir​ (2025), M2, Université Paris-Saclay,​‌ AI-Enhanced Prediction of Diabetic​​ Foot Risk: Integrating NIRS​​​‌ and TcPO2 Data for​ Early Vascular Health Assessment​‌ coadvised with Marie Gernigon​​ .
• Hajar Elhaimer (2025),​​​‌ M2, Université Mohammed VI​ Polytechnic, Morocco, AI-based Assessment​‌ of the CorticoCardiovascular Interaction​​ for monitoring stress. (50%)​​​‌ co-advised with Maria Sara​ Nour Sadoun .
• Aya​‌ Harkat (2025), M2, Université​​ Paris-Saclay, Combined observer design​​​‌ and Model predictive control​ framwork for co-contraction mdoeling​‌ co-advised with Bastien Berret​​
• Damya Mansouri (2025), Université​​​‌ Paris-Saclay co-advised with Juan​ Vargas -
• Antoine Pierrard​‌ (2025,), M1, Télécom Physique​​ Strasbourg.

Bastien Berret

• Aymeric​​​‌ Orhan, 2021-2025 (30%), then​ post-doc
• Abdelwaheb Hafs, 2021-2025​‌ (50%), then ATER at​​ Université Paris-Saclay
• Anais Farr,​​​‌ 2024-, sur le rôle​ du bruit et des​‌ délais sensoriels sur la​​ planification motrice, bourse CDSN​​​‌ - ENS Rennes (100%)​
• Duy Hoang, 2024-, sur​‌ la création de contrôleur​​ d'exosquelette basé EMG et​​​‌ machine learning, bourse J.P.​ Aguilar auprès de la​‌ fondation CFM pour la​​ recherche (30%)

Michel-Ange Amorim​​​‌ PhD students

• Meriem Younssi​ (2025-) Université Paris-Saclay, Determinants​‌ of perceived intentionality and​​ action in response to​​​‌ uncertain visual trajectories.
• Olga​ Polezhaeva (2022-2025) Université Paris-Saclay,​‌ Perceiving and interacting with​​ an object whose trajectory​​​‌ is difficult to predict.​

Master students

• Clarisse Ancel​‌ (2025), M1, Université Paris-Saclay,​​ Effects of the spatial​​​‌ structure of uncertain visual​ trajectories and perceived final​‌ position on decision-making in​​ a visuomotor extrapolation task.​​​‌
• Alexandre Trstenjak (2025), M1,​ Université Paris-Saclay, Perceiving and​‌ reacting to an object​​ whose trajectory is difficult​​​‌ to predict.

Ioannis Bargiotas​

PhD students

• Quentin Laborde​‌ (2022-), ENS Paris-Saclay, Development​​ of machine learning models​​​‌ to identify traffic safety​ issues related to driver​‌ mental workload (30%).

M2​​ students

• François Saulnier (2025),​​​‌ M2, ENS Paris-saclay, Machine​ Learning and risk of​‌ future injuries in mid-​​ and long-range runners: Single​​​‌ point of view is​ never enough

M1 students​‌

• Olympe Pann (2025), M1,​​ Université Paris-saclay, Standardization of​​​‌ estimation for Posture and​ Gait Assessment Variables

10.2.2​‌ Juries

10.2.3 Educational and pedagogical‌​‌ outreach

International journals

• 8​​​‌ articleR.Rose Alaslani​, L.Levina Perzhilla​‌, M. M.Muhammad​​ Mahboob Ur Rahman,​​​‌ T.-M.Taous-Meriem Laleg-Kirati and​ T.Tareq Al-Naffouri.​‌ You Can Monitor Your​​ Hydration Level Using Your​​​‌ Smartphone Camera.IEEE​ Transactions on Instrumentation and​‌ Measurement742025,​​ 1-14HALDOIback​​​‌ to text
• 9 article​E.Emilie Bollache,​‌ I.Ioannis Bargiotas,​​ A.Alain Giron,​​​‌ A.Alain de Cesare​, G.Gilles Soulat​‌, E.Elie Mousseaux​​ and N.Nadjia Kachenoura​​​‌. Noninvasive quantification of​ aortic wave reflection timing​‌ indices in aging: a​​ study combining MRI and​​​‌ applanation tonometry.Journal​ of HypertensionJuly 2025​‌HALDOI
• 10 article​​G.Gabriel Chan,​​​‌ J.Julia Ruiz-Fernández,​ M.-L.Marie-Laure Paillère Martinot​‌, I.Ioannis Bargiotas​​, N.Nicolas Vayatis​​​‌, E.Eric Artiges​ and J.-L.Jean-Luc Martinot​‌. 413. Resilience and​​ Adolescent Brain Structure Protective​​​‌ Features: A Confirmation and​ Machine Learning Study.​‌Biological Psychiatry979​​May 2025, S267-S268​​​‌HALDOI
• 11 article​J.Jianxu Chen,​‌ I.Ibrahima N’doye,​​ Y.Yevhen Myshkevych,​​​‌ F.Fahad Aljehani,​ M. K.Mohammad Khalil​‌ Monjed, T.-M.Taous-Meriem​​ Laleg-Kirati and P.-Y.Pei-Ying​​​‌ Hong. Viral particle​ prediction in wastewater treatment​‌ plants using nonlinear lifelong​​ learning models.npj​​​‌ Clean Water81​April 2025, 28​‌HALDOI
• 12 article​​W.Waldez Gomes,​​​‌ L.Lucas Quesada,​ B.Bastien Berret,​‌ N.Nicolas Vignais and​​ D.Dorian Verdel.​​​‌ How task-relevant vibratory feedback​ from an active exoskeleton​‌ can lead to ergonomic​​ postures.Communications Engineering​​​‌42152025HAL​DOI
• 13 articleJ.​‌Jiahao Hu, M.​​ M.Muhammad Mahboob Ur​​ Rahman and T.-M.Taous-Meriem​​​‌ Laleg-Kirati. Adaptive long-range‌ modeling of EEG and‌​‌ ECG with Mamba and​​ dynamic graph learning.​​​‌Scientific Reports151‌November 2025, 38762‌​‌HALDOIback to​​ text
• 14 articleE.​​​‌Ege Kibrislioglu and T.-M.‌Taous-Meriem Laleg-Kirati. Motion‌​‌ Artifact Removal from EEG​​ Signals Using the Motion-Net​​​‌ Deep Learning Algorithm.‌Biomedical Signal Processing and‌​‌ Control113March 2026​​, 108877HALDOI​​​‌back to text
• 15‌ articleY.Yasmine Marani‌​‌, I.Ibrahima N’doye​​ and T. M.Taous​​​‌ Meriem Laleg-Kirati. Deep-learning‌ based KKL chain observer‌​‌ for discrete-time nonlinear systems​​ with time-varying output delay​​​‌.Automatica171January‌ 2025, 111955HAL‌​‌DOIback to text​​back to text
• 16​​​‌ articleY.Yevhen Myshkevych‌, I.Ibrahima N’doye‌​‌, J. S.Julie​​ Sanchez Medina, F.​​​‌Fahad Aljehani, Y.‌Yanghui Xiong, T.-M.‌​‌Taous-Meriem Laleg-Kirati and P.-Y.​​Pei-Ying Hong. Combining​​​‌ flow virometry with tree-based‌ machine learning models for‌​‌ rapid virus particle estimation​​ in different wastewater matrices​​​‌.Water Research284‌May 2025, 123905‌​‌HALDOI
• 17 article​​L.Lucas Quesada,​​​‌ D.Dorian Verdel,‌ O.Olivier Bruneau,‌​‌ B.Bastien Berret,​​ M.-A.Michel-Ange Amorim and​​​‌ N.Nicolas Vignais.‌ EMG feature extraction and‌​‌ muscle selection for continuous​​ upper limb movement regression​​​‌.Biomedical Signal Processing‌ and Control1032025‌​‌, 107323HALDOI​​back to text
• 18​​​‌ articleL.Lucas Quesada‌, D.Dorian Verdel‌​‌, O.Olivier Bruneau​​, B.Bastien Berret​​​‌, M. A.Michel‌ Ange Amorim and N.‌​‌Nicolas Vignais. EMG-to-torque​​ models for exoskeleton assistance:​​​‌ a framework for the‌ evaluation of in situ‌​‌ calibration.The International​​ Journal of Robotics Research​​​‌January 2026HALDOI‌back to text
• 19‌​‌ articleA.Axel Roques​​, D.Dimitri Keriven​​​‌ Serpollet, A.Alice‌ Nicolaï, S.Stéphane‌​‌ Buffat, Y.Yannick​​ James, N.Nicolas​​​‌ Vayatis, I.Ioannis‌ Bargiotas and P.-P.Pierre-Paul‌​‌ Vidal. Continuous Assessment​​ of Mental Workload During​​​‌ Complex Human–Machine Interaction: Inferring‌ Cognitive State from Signals‌​‌ External to the Operator​​.Sensors2512​​​‌June 2025, 3624‌HALDOIback to‌​‌ text
• 20 articleS.​​ N.Sara Nour Sadoun​​​‌, A.Arnaud Boutin‌, F.François Cottin‌​‌ and T.-M.Taous-Meriem Laleg-Kirati​​. Modeling Brain-Heart Interaction:​​​‌ A Review of Mechanistic‌ Dynamical Models.IEEE‌​‌ Reviews in Biomedical Engineering​​2025, 1-17HAL​​​‌DOIback to text‌back to text
• 21‌​‌ articleS.Soumia Siyoucef​​, M.Mourad Adnane​​​‌, M. M.Muhammad‌ Mahboob Ur Rahman,‌​‌ T.-M.Taous-Meriem Laleg-Kirati and​​ T.Tareq Al-Naffouri.​​​‌ Noninvasive Monitoring of Dehydration‌ of Fasting and Sportspeople‌​‌ Subjects via Skin Capacitance​​.IEEE Sensors Journal​​​‌257April 2025‌, 11428-11440HALDOI‌​‌back to text
• 22​​ articleA.Anthony Supiot​​​‌, N.Nicolas Roche‌, B.Bastien Berret‌​‌ and D.Didier Pradon​​. Evaluation of Muscle​​​‌ Synergy Flexibility Induced by‌ a Muscle Nerve Block‌​‌ in Chronic Stroke Patients​​​‌.Biomechanics52​2025, 27-27HAL​‌DOI
• 23 articleJ.​​ M.Juan M Vargas​​​‌, M. A.Mohamed​ A Bahloul, M.​‌ M.Mohamed M Boularas​​, K.Kaan Yuceel​​​‌, S.Slaheddine Aridhi​ and T.-M.Taous-Meriem Laleg-Kirati​‌. Assessment of pulse​​ wave velocity through weighted​​​‌ visibility graph metrics from​ photoplethysmographic signals.Scientific​‌ Reports151August​​ 2025, 31325HAL​​​‌DOIback to text​back to text
• 24​‌ articleD.Dorian Verdel​​, B.Bastien Berret​​​‌ and E.Etienne Burdet​. How individual vigor​‌ shapes human-human physical interaction​​.eLifeDecember 2025​​​‌HALDOI

International peer-reviewed​ conferences

• 25 inproceedingsH.​‌Hu Jiahao, M.​​ M.M. Mahboob Ur​​​‌ Rahman, T.Tareq​ Al-Naffouri and T.-M.Taous-Meriem​‌ Laleg-Kirati. Mamba-CAM-Sleep: A​​ Mamba-based Channel Attention Model​​​‌ for Sleep Staging Classification​.2025 47th Annual​‌ International Conference of the​​ IEEE Engineering in Medicine​​​‌ and Biology Society (EMBC)​Copenhagen, DenmarkIEEEDecember​‌ 2025, 1-6HAL​​DOIback to text​​​‌
• 26 inproceedingsY.Yasmine​ Marani, Z.Zhe​‌ Fu, I.Ibrahima​​ N’doye, E.Eric​​​‌ Feron, T.-M.Taous-Meriem​ Laleg-Kirati and A. M.​‌Alexandre M Bayen.​​ Position and Speed Estimation​​​‌ Using Deep Learning-Based KKL​ Observer in Mixed Autonomy​‌ Traffic Systems.2025​​ IEEE 64th Conference on​​​‌ Decision and Control (CDC)​Rio de Janeiro (BR),​‌ BrazilDecember 2025HAL​​back to text
• 27​​​‌ inproceedingsY.Yasmine Marani​, F.Filho Israel​‌, A.Alnafouri Tareq​​ and T.-M.Taous-Meriem Laleg-Kirati​​​‌. Unsupervised Physics-Informed Neural​ Network-Based Nonlinear Observer Design​‌ for Autonomous Systems Using​​ Contraction Analysis..2025​​​‌ 23rd European Control Conference​ (ECC)Thessaloniki, GreeceJune​‌ 2025HALback to​​ text
• 28 inproceedingsJ.​​​‌Juan Vargas, J.​Jean Michel Davaine and​‌ T.-M.Taous-Meriem Laleg-Kirati.​​ Symptomatic and Asymptomatic Carotid​​​‌ Plaques Classification using CT​ Images and Hybrid Deep​‌ Transfer Learning.2025​​ 47th Annual International Conference​​​‌ of the IEEE Engineering​ in Medicine and Biology​‌ Society (EMBC)Copenhagen, Denmark​​IEEEDecember 2025,​​​‌ 1-4HALDOIback​ to textback to​‌ text

Conferences without proceedings​​

• 29 inproceedingsA.Abdelwaheb​​​‌ Hafs, D.Dorian​ Verdel, O.Olivier​‌ Bruneau and B.Bastien​​ Berret. Game-Theoretic Interaction​​​‌ Control for Assistive Exoskeletons:​ a 2-DOF Simulation Study​‌.IFAC 2025 -​​ Joint 10th IFAC Symposium​​​‌ on Mechatronic Systems and​ 14th Symposium on Robotics​‌Paris, FranceJuly 2025​​HALback to text​​​‌

Edition (books, proceedings, special​ issue of a journal)​‌

• 30 proceedingsA Symbolic​​ Approach for Task-related Gaze​​​‌ Classification.47th Annual​ International Conference of the​‌ IEEE Engineering in Medicine​​ & Biology Society,Copenaghen,​​​‌ Denmark2025. In​ press. HAL
• 31 proceedings​‌Risk of Future Injuries​​ in Mid/Long-Range Runners: Feasible​​​‌ Rashomon Set of Sparse​ Decision Trees.2025​‌ IEEE 37th International Conference​​ on Tools with Artificial​​​‌ Intelligence (ICTAI)Athens, France​IEEENovember 2025,​‌ 1499-1503HALDOIback​​ to text

Reports &​​​‌ preprints

• 32 miscM.​Mohamed Boukaf, Z.​‌Zehor Belkhatir, M.​​ A.Mohamed A Bahloul​​, M.Mohamed Chadli​​​‌ and T.-M.Taous-Meriem Laleg-Kirati‌. Unsupervised Parallel Physics‌​‌ Informed Neural Networks for​​ Blood Flow Estimation and​​​‌ Cardiovascular Parameters Assessment ⋆‌.November 2025HAL‌​‌back to text
• 33​​ miscM.Mohamed Boukaf​​​‌, Z.Zehor Belkhatir‌, M.Mohamed Chadli‌​‌ and T.Taous‐meriem Laleg-Kirati​​. PINN-Based Parameters And​​​‌ Input Joint Estimation Of‌ Nonlinear Systems With Non-Gaussian‌​‌ Noise.February 2025​​HAL
• 34 miscM.​​​‌Mohamed Boukaf, Z.‌Zehor Belkhatir, M.‌​‌Mohamed Chadli and T.-M.​​Taous-Meriem Laleg-Kirati. Windkessel​​​‌ model-based Physics Informed Neural‌ Networks for Blood Flow‌​‌ Estimation and Cardiovascular Parameters​​ Assessment.May 2025​​​‌HAL
• 35 miscM.‌Mohamed Boukaf, Z.‌​‌Zehor Belkhatir, A.​​Aya Harkat, M.​​​‌Mohammed Chadli, B.‌Bastien Berret and T.‌​‌ M.Taous Meriem Laleg​​. Combined Deep KKL​​​‌ and NMPC for Modeling‌ Muscle Co-contraction Under Measurement‌​‌ Delays.January 2026​​HALback to text​​​‌back to text
• 36‌ miscA.Arnaud Boutin‌​‌ and A.Arnaud Badets​​. Conscious awareness of​​​‌ actions shapes motor memory‌ consolidation.June 2025‌​‌HALDOI
• 37 misc​​A.Adrien Conessa,​​​‌ A.Arnaud Boutin and‌ B.Bastien Berret.‌​‌ Evidence for separate processes​​ underlying movement and decision​​​‌ vigor in a reward-oriented‌ task.November 2025‌​‌HALDOI
• 38 misc​​S.Sylvain Famié and​​​‌ M.-A.Michel-Ange Amorim.‌ Different spatio-temporal strategies for‌​‌ controlling a striking gesture​​ to slide an object​​​‌ toward a target distance‌.September 2025HAL‌​‌DOI
• 39 miscI.​​ J.Israel Jesus Santos​​​‌ Filho, M. M.‌Muhammad Mahboob Ur Rahman‌​‌, T.-M.Taous-Meriem Laleg-Kirati​​ and T.Tareq Al-Naffouri​​​‌. Radio-PPG: photoplethysmogram digital‌ twin synthesis using deep‌​‌ neural representation of 6G/WiFi​​ ISAC signals.2025​​​‌HALback to text‌
• 40 miscA.Abdelbaki‌​‌ Guir, I. J.​​Israel Jesus Santos Filho​​​‌ and T.-M.Taous-Meriem Laleg-Kirati‌. L.Latifa Hamami-Mitiche‌​‌, eds. CycleGAN for​​ EEG artifact removal.​​​‌2025HAL
• 41 misc‌A.Abdelbaki Guir,‌​‌ C.Chloe French,​​ D.Dan Robbins,​​​‌ D.Dan Gordon,‌ M.Marie Gernigon and‌​‌ M.Meriem Taous.​​ Signal-Based Monitoring for Tissue​​​‌ Oxygenation & Diabetes Characterization‌.2025HAL
• 42‌​‌ miscA.Abdelwaheb Hafs​​, A.Anaïs Farr​​​‌, D.Dorian Verdel‌, O.Olivier Bruneau‌​‌, E.Etienne Burdet​​ and B.Bastien Berret​​​‌. Model predictive game‌ control for personalized and‌​‌ targeted interactive assistance.​​February 2025HALDOI​​​‌back to text
• 43‌ miscA.Aymeric Orhan‌​‌, D.Duy Hoàng​​, O.Olivier Bruneau​​​‌, B.Bastien Berret‌ and F.Franck Geffard‌​‌. Early Classification of​​ Human Motion Intent for​​​‌ Exoskeleton Assistance Using Kinematics‌.March 2025HAL‌​‌back to text
• 44​​ miscA.Aymeric Orhan​​​‌, D.Duy Hoàng‌, O.Olivier Bruneau‌​‌, B.Bastien Berret​​ and F.Franck Geffard​​​‌. Enhancing Exoskeleton Assistance‌ Flow Control by Leveraging‌​‌ Human Variability and Probabilistic​​ Movement Primitives.March​​​‌ 2025HALback to‌ text
• 45 miscA.‌​‌Aymeric Orhan, D.​​​‌Duy Hoàng, O.​Olivier Bruneau, B.​‌Bastien Berret and F.​​Franck Geffard. Using​​​‌ Artificial Demonstrations Emulating Human​ Movement Variability for a​‌ Learning-based Exoskeleton Flow Controller​​.March 2025HAL​​​‌DOIback to text​
• 46 miscL.Lucas​‌ Quesada, D.Dorian​​ Verdel, O.Olivier​​​‌ Bruneau, B.Bastien​ Berret, M.-A.Michel-Ange​‌ Amorim and N.Nicolas​​ Vignais. Less is​​​‌ more: uncompensated gravity torques​ for intuitive EMG-based assistance​‌ with a robotic exoskeleton​​.November 2025HAL​​​‌DOI
• 47 miscS.​ N.Sara Nour Sadoun​‌, G. A.Giuseppe​​ Alessio d'Inverno, A.​​​‌Arnaud Boutin, F.​François Cottin and T.-M.​‌Taous-Meriem Laleg-Kirati. Physics-Informed​​ Neural Estimation of State​​​‌ and Unknown Input inAutonomic​ Cardiac Dynamics with Left-Invertibility​‌ Constraints.January 2026​​HALback to text​​​‌
• 48 miscS. N.​Sara Nour Sadoun,​‌ G. A.Giuseppe Alessio​​ d'Inverno, A.Arnaud​​​‌ Boutin, F.François​ Cottin and T.-M.Taous-Meriem​‌ Laleg-Kirati. State and​​ Unknown Input Estimation using​​​‌ a Left-Invertibility Constrained NeuralEstimator​ in Delayed Autonomic Cardiac​‌ Dynamics.January 2026​​HALback to text​​​‌
• 49 miscJ. M.​Juan Manuel Vargas Garcia​‌ and H.Hacene Serrai​​. MRI contrast enhancement​​​‌ using schrodinger spectrum and​ unsupervised CNN: Amélioration du​‌ contraste IRM par le​​ spectre de Schrödinger et​​​‌ un CNN non supervisé​.January 2026HAL​‌back to text
• 50​​ miscJ. M.Juan​​​‌ Manuel Vargas Garcia,​ L.Louise Wang,​‌ A.Alix Piedelievre,​​ G.Guillaume Goudot,​​​‌ J.-M.Jean-Michel Davaine and​ T.-M.Taous-Meriem Laleg-Kirati.​‌ CT-Based Classification of Symptomatic​​ vs. Asymptomatic Carotid Plaques​​​‌ Using Schrodinger Spectrum Features​.January 2026HAL​‌

Other scientific publications

• 51​​ inproceedingsL.Leonie Hirsch​​​‌, A. L.Anna​ Luisa Maier, N.​‌Nuno de Sá Teixeira​​, M. A.Michel​​​‌ Ange Amorim, C.​Christoph von Castell and​‌ H.Heiko Hecht.​​ Objects in vista space​​​‌ are misrepresented to be​ closer in a spatial​‌ updating task.47th​​ European Conference on Visual​​​‌ PerceptionMainz, GermanyAugust​ 2025, https://converia.uni-mainz.de/frontend/index.php?page_id=4562&v=List&do=15&day=all&ses=3422#HAL​‌
• 52 inproceedingsO.Olga​​ Polezhaeva, S.Stefan​​​‌ Glasauer and M. A.​Michel Ange Amorim.​‌ Perceptual Strategies for Extrapolating​​ Noisy Visual Trajectories.​​​‌47th European Conference on​ Visual PerceptionMainz, Germany​‌August 2025, https://converia.uni-mainz.de/frontend/index.php?page_id=4562&v=List&do=15&day=all&ses=3361#​​HAL