NEMO-BMI description of work (deliverable D33)

One medical application of Brain Machine Interfaces (BMIs)s is to translate brain signals into commands for effectors e.g. exoskeleton or electrical stimulation of the spinal cord for motor substitution.

Our application targets severely disabled users, in particular, Spinal Cord Injury (SCI) patients.

SCI leads to severe neurological deficits. Locomotion impairment or the loss of upper limb functions requires caregivers and dramatically impacts the patient’s autonomy in their daily life. Each year, there are approximately 31,000 new cases of SCI in EU⁠ and 18,000 in the US, with dramatic human, societal and economical costs. Let us keep in mind that currently there is no approved therapy to improve motor recovery from SCIs. Consequently, there is no BCI system commercially available and the main results published relate to clinical proof of concepts [1], [3].

These revolutionary BMI technologies provides SCI patient with control over orthosis or (neuro)prostheses overcoming some of these disabilities. These recent clinical trials raise great hopes for SCI patients, effectively assessing the feasibility of chronic Electrocorticography (ECoG)-based motor BMI. Wireless minimally invasive ECoG recording systems used in these clinical trials are far more appropriate for long-term application. In terms of effectors, to compensate motor disabilities our consortium supplies patients with a spinal cord stimulation therapy (Targeted Epidural Spinal Stimulation from EPFL) and a dedicated chronic implantable device for spinal cord stimulation (ARC^IM from ONWARD).

These clinical trials are very promising but also highlight numerous barriers to bringing Motor BMI to clinical practice or home use. Indeed, brain activity decoding to supply motor functions requires a model in the decoding software. This model is patient specific and has to be trained for each patient in a supervised manner and in a well-controlled environment. For BMI decoding model training, users are requested to perform specific actions: imaging / tentative motor tasks. The neuronal activity is recorded simultaneously. Generally, numerous repetitions are needed to identify the model, which correctly decodes neuronal activity into motor tasks. Machine learning or AI algorithms are generally used for the model training. When the decoding model is stable, it can be used by the patent autonomously.

However, the performances of BCI decoders can degrade over time due to variations in the user’s psychological state, brain plasticity, micro-movements of the brain or other concept drifts. The BMI decoding efficiency degradation requires updating the decoders or training new decoders from scratch, when the performance has degraded excessively. The necessity of regular decoder retraining due to neuronal data instability or variability, is one of the major barriers for long-term use of BMIs at home.

Moreover, targeted epidural spinal cord stimulation can restore movement of paralyzed muscles after spinal cord injury. However, the configuration of the electrical stimulation relies on an almost infinite number of possible combinations of electrodes, stimulation frequency and amplitude. These parameters are currently optimized manually by a team of experts.

In the longer term, the goal is to have devices that are portable, discreet and with a long battery life. To this end, we have started the optimization of the decoding algorithms targeting their integration in a low power application specific integrated circuit (ASIC) (WP5).

All these technological developments are patient centered and embed ethics by design. Their performance will be assessed off-line using existing clinical databases and on-line in new clinical situations after approval of the amendments.

Auto-Adaptive Brain Signal Decoder (WP2)

Facing the constraints of regular Motor BMI retraining, the objective of WP2 is the development of a auto-adaptive BMI (A-BMI) algorithms. After initial BMI decoder training, the BMI decoder will be updated during the BMI free use. The core idea is to infer the training task information directly from the brain signals rather than from the environment, eliminating or reducing the necessity for the special training sessions. The newly proposed Neural Response (NR) decoder will classify the motor actions executed during the BMI free use as ‘correct’ or ‘erroneous’. E.g. if the BMI motor action was considered by the NR decoder as a correct one, with a high level of certainty, this action is added as a label to the corresponding epochs as training data. An adaptive incremental learning algorithm developed at CEA is able to update the current model in real-time using the newly collected data. The A-BMI concept was initially explored off-line to validate the concept.

The objective of the project is (a) to develop A-BMI in more complex and realistic real-life scenarios, (b) to implement A-BMI decoding algorithms, (c) to test them with existing brain recordings, (d) and, finally, to test them in clinical trials, which are already in progress.

Auto-adaptive brain guided electrical stimulation of spinal cord (WP3)

The NEMO-BMI project proposes to use artificial intelligence algorithms to efficiently explore the stimulation space and guarantee the convergence to optimal parameters that can restore movements such as opening the hand, closing the hand, or move the arm in three dimensions.

In particular, the AI algorithm can use a reward function that is based on muscle activity, kinematics, and feedback of the physiotherapist and, of course, feedback of the patient. This feedback from the participant can be measured objectively from the brain signals.

The NEMO-BMI project will develop the tools to measure and decode patient satisfaction and drive the exploration of stimulation patterns to restore the most natural movements after spinal cord injury.

Neuromorphic Approach for Brain Signal Decoding (WP4)

The neuromorphic approach exploits the knowledge about the brain structure and functioning. It consists of two main blocks: a three dimensional spike timing neural network (3D-SNN) and a reservoir of randomly connected firing rate neurons called Echo state network (ESN). The 3D-SNN structure is composed of leaky integrate and fire neurons modelled to mimic the brain neural cells dynamics. Their positions in 3D space were determined according to the personal brain shape of the patient while the connections between them were set initially based on their mutual distance and tuned to reflect the personal brain activity by the brain-inspired spike timing dependent plasticity (STDP) rule. The 3D-SNN structure receives the raw ECoG signals and transforms them to spike trains having time-varying firing rates. The ESN module is a pool of firing rate neurons whose random recurrent connectivity is inspired by the neocortex in the brain. It receives as input firing rates of a group of neurons from the 3D-SNN structure and learns in real time to decode patient’s movement intent. The algorithms are intended to be implemented on a mobile neuromorphic platform.

Miniaturized Hardware for On-Line Decoding (WP5)

The team’s long-term goal is to provide the patient with a solution which is virtually invisible and with a battery autonomy of at least an entire day, if not longer. Currently, the decoding of the brain signals requires a powerful computer, which can extract meaningful features from the raw signals and can then apply a model to decode the features and to identify the patient’s movement intent. Currently, this was done using a high-level compute framework (‘Matlab’) running on a laptop, which the patient carried in a backpack, but this is obviously not an ergonomic solution. The team has already completed a first step of optimization, which consisted of simplifying the algorithm and porting the code to the C++ language. With these optimizations, the decoding algorithm can now operate in real-time on a small, off-the-shelf, portable device. This initial step has resulted in more than a 10X reduction in power consumption. The embedded computer, however, uses a general-purpose processor, which is optimized for running Linux applications, rather than signal processing. The next step which, is already underway, is the development of a dedicated integrated (IC) which is optimized to perform the feature extraction algorithm with minimum power consumption. This circuit will execute the current decoding algorithm, while providing sufficient flexibility to execute new algorithms, which may emerge. In the future, this IC could be integrated in the hardware worn by the patient, or potentially

Scientific publications of the consortium

• Rusev G, Yordanov S, Nedelcheva S, Banderov A, Sauter-Starace F, Koprinkova-Hristova P, Kasabov N. Decoding Brain Signals in a Neuromorphic Framework for a Personalized Adaptive Control of Human Prosthetics. Biomimetics. 2025; 10(3):183. https://doi.org/10.3390/biomimetics10030183
• Saad J, Evans A, Jaoui I, Roux-Sibillon V, Hardy E and Anghel L (2025) Comparison metrics and power trade-offs for BCI motor decoding circuit design. Front. Hum. Neurosci. 19:1547074. doi: 10.3389/fnhum.2025.1547074
• Rusev G, Yordanov S, Nedelcheva S, Banderov A, Lafaye de Micheaux H, Sauter-Starace F, Aksenova T, Koprinkova-Hristova P, Kasabov N. NEuroMOrphic Neural-Response Decoding System for Adaptive and Personalized Neuro-Prosthetics’ Control. Biomimetics. 2025; 10(8):518. https://doi.org/10.3390/biomimetics10080518
• Lafaye de Micheaux H, Martel F, Sauter-Starace F, Charvet G, Lorach H, Aksenova T, Auto-adaptation of ECoG-based motor BCI using neural response decoder: a cross-patient study, 9th Graz Brain-Computer Interface Conference 2024, https://doi.org/10.3217/978-3-99161-014-4-013
• Penkov D, Koprinkova-Hristova P, Kasabov N, Nedelcheva S, Ivanovska S & Yordanov S, Grid Search Optimization of Novel SNN-ESN Classifier on a Supercomputer Platform, Lecture Notes in Computer Science, https://doi.org/10.1007/978-3-031-56208-2_45
• Souriau R, Moly A, Martel F, Struber L, Karakas S, Juillard V, Charvet G, Aksenova T, Hierarchical Hidden Markov Model for Online Decoding in Brain-Computer Interface, 2024 32nd European Signal Processing Conference (EUSIPCO), https://doi.org/10.23919/EUSIPCO63174.2024.10715255
• Koprinkova-Hristova P, Penkov D, Nedelcheva S, Yordanov S, & Kasabov N, On-line Learning, Classification and Interpretation of Brain Signals using 3D SNN and ESN, International Joint Conference on Neural Networks (IJCNN) 2023, https://doi.org/10.1109/IJCNN54540.2023.10191974
• Struber L, Martel F, Karakas S, Juillard V, Bellicha A, Sauter-Starace F, Chabardès S, Lorach H, Charvet G, Aksenova T, Predictors of ECoG-BCI performances across subjects and sessions derived from idle state characteristics, 9th Graz Brain-Computer Interface Conference 2024, https://doi.org/10.3217/978-3-99161-014-4-015

NEMO-BMI Status at closing (September 2026) (D1.9: Public summary)

Every year, 108’000 Europeans sustain a spinal cord injury (SCI) that paralyzes their legs[1]. SCI has no cure and rehabilitation is largely unable to restore motor function[2]. Consequently, 3.1 million Europeans currently live with permanent paraplegia due to SCI. Huge costs of medical and social care coupled with severely reduced physical and financial independence leave a severe emotional, economic and social impact on the people with SCI, their families and society[3]. We have developed the world’s first clinically validated therapy capable of restoring voluntary mobility after paralysis caused by SCI. The brain-spine interface creates a digital bridge between the brain and the spinal cord that bypasses the injury and restores functional leg movements[4][5][6][7].

To provide intention-driven control, brain activity recorded through implanted electrocorticography (ECoG from the WIMAGINE implants[8]) was coupled to Epidural Electrical Stimulation (EES), thereby re-establishing communication across the injury. The feasibility and clinical impact of brain-spine interfaces have been demonstrated in humans in the STIMO-BSI (NCT04632290) and Think2Go (NCT06243952) clinical trials, where four participants with chronic SCI regained voluntary control of leg movements and the ability to stand, walk, navigate stairs and traverse complex terrain (Fig. 1). These studies established the clinical proof of concept for brain-spine interfaces and highlighted a clear path toward clinical deployment. Despite these promising outcomes, the clinical therapy prototypes relied on external research computers and required lengthy expert-led calibration procedures, impeding adoption in routine clinical practice. The inspiration of AI from the brain organization led to development of neuromorphic models[9] promising creation of low energy consumption devices in near future. NEMO-BMI addressed these barriers through innovations in embedded neural decoding, neuromorphic decoder, automated therapy personalization, and neuroprosthesis architecture design. The principal outcomes of NEMO-BMI are:

• A low-power neural decoding ASIC capable of inferring movement intentions from ECoG signals in real time on a 3.6mm² silicon chip. Compared with the research-grade computing platform currently used in the clinical prototype, the ASIC reduces power consumption by approximately 10,000-fold while maintaining real-time decoding performance. This breakthrough enables replacement of bulky external hardware with embedded artificial intelligence, representing a critical step toward a portable and clinically deployable BSI.
• An automated EES configuration platform based on Gaussian Process Bayesian Optimization (GP-BO). Current clinical configuration requires expert operators, multiple days of calibration sessions, and extensive electromyographic (EMG) recordings. The developed optimization framework identifies optimal stimulation parameters within one to two hours and has demonstrated superior performance compared with manually configured settings, substantially reducing clinical burden while improving therapeutic outcomes.
• Auto-adaptive features were successfully integrated into the brain signal decoder developed by CEA which allows for a dramatically limited need for decoding model recalibration
• An initial integrated, lightweight prototype of the wearable components of the brain-spine interface
• Neuromorphic architecture was explored for the aims of brain signals decoding. It possesses auto-adaptive ability of the brain and could be embedded in future in some of fast developing nowadays neuromorphic chips.

Together, these results address the principal barriers to clinical adoption of brain-spine interfaces: system complexity and personalization burden. Next step for the team is an integration in a patient-oriented redesign of the brain-spine interface prototype into a minimal viable product, moving the whole system from TRL 4 to 6, hopefully with an EIC Transition Challenge support, enabling its full functional evaluation in clinical setting and in daily life.

---

[1] Safdarian, M et al. Global, regional, and national burden of spinal cord injury, 1990-2019: a systematic analysis for the Global Burden of Disease Study 2019. The Lancet Neurology 22, (2023)

[2] Nas, K, Yazmalar, L, Şah, V, Aydın, A & Öneş, K. Rehabilitation of spinal cord injuries. World J Orthop 6, (2015)

[3] Foundation, CDR. One Degree of Separation, Paralysis and Spinal Cord Injury in the United States.  (2009)

[4] Adams, M & Cavanagh, JFR. International Campaign for Cures of Spinal Cord Injury Paralysis (ICCP): another step forward for spinal cord injury research. Spinal cord 42, (2004).

[5] Lorach, H et al. Walking naturally after spinal cord injury using a brain–spine interface. Nature 618, (2023).

[6] Courtine, G, van den Brand, R & Musienko, P. Spinal cord injury: time to move. Lancet 377, (2011).

[7] Mehrholz, J, Kugler, J & Pohl, M. Locomotor training for walking after spinal cord injury. Cochrane Database Syst Rev 11, (2012).

[8] Mestais, C. S., Charvet, G., Sauter-Starace, F., Foerster, M., Ratel, D., & Benabid, A. L. (2014). WIMAGINE: wireless 64-channel ECoG recording implant for long term clinical applications. IEEE transactions on neural systems and rehabilitation engineering, 23(1), 10-21.

[9] Kasabov, N., Time-Space, Spiking Neural Networks and Brain-Inspired Artificial Intelligence, Springer Nature (2019) 750p