TL;DR
NeuroRestore’s success comes from fusing academia, hospital care and industry into one translational pipeline for spinal cord injury — a template other neurotech labs would do well to study.
Background
Academic labs pursue projects with no clinical potential. Start-ups can’t foster a stable environment for rigorous long-term R&D. And hospitals lack the innovation to keep up with medtech advancements.
Stereotypically, these are the factors holding back the transition of promising technology into clinical care; the challenges many research entities are beginning to tackle. One such example is NeuroRestore, the disruptive lab overlooking the magnificent shores of Lake Léman in Switzerland. Co-directed by neuroscientist Grégoire Courtine and neurosurgeon Jocelyne Bloch, NeuroRestore is a joint center founded by EPFL, Lausanne University Hospital (CHUV), the University of Lausanne’s (UNIL) Faculty of Biology and Medicine, and the Defitech Foundation — a genuine cross-institutional venture from day one, not simply an EPFL lab. Its focus is interventional treatment for the motor recovery of patients with spinal cord injuries (SCIs) — namely epidural electrical stimulation (EES) of the dorsal spinal roots. Its work has already produced remarkable results (Kathe et al. 2022).
NeuroRestore has found a happy medium in their engagement in industry, academia and hospital settings, thereby serving as a leading example of how translational research can be carried from idea stage to clinical application within the one lab. Through my recent internship, I sought insight into their regulatory, clinical, and technological operations to better understand just what makes them tick.
Institutional Collaborations
My previous exposure to Resonait’s operational focus on grant applications had led me to expect a similar organisational structure. Yet I was still struck by the extent to which NeuroRestore relies (and thrives) on collaborative relationships and external funding.
The group’s early successes led to a close collaboration with Swiss neurosurgeon Jocelyne Bloch to later enable advanced clinical trials at the Centre Hospitalier Universitaire Vaudois (CHUV) in Lausanne. Later, additional preclinical research activities at Geneva’s Campus Biotech helped support and extend the translational program. What’s more, NeuroRestore’s desire to wield more operational control over the manufacturing of its device hardware led to the creation of a dedicated hardware partner, Onward, now a thriving enterprise whose BCI hardware is sold to research labs globally, of course including NeuroRestore.
This collaborative model is also reflected in NeuroRestore’s funding. Beyond its institutional foundations, the centre attracts targeted philanthropic support capable of directly accelerating clinical translation. The Michael J. Fox Foundation and Defitech Foundation are two such examples. Other sources of funding, whilst not always publicly announced, were doggedly chased within the walls of the lab.
Today’s well-oiled NeuroRestore machine can therefore pursue an idea from its pre-clinical conception all the way through to neurologically-indicated clinical approval, and in the future, its market sales. Their ability to work on this entire production line (preclinical proof of concept, hardware design, regulatory approval, device implantation and data acquisition) in-house, or at least within their Lausanne ecosystem, represents an enviable operational advantage. Similarly, first-hand exposure to the users of the technology — the patients themselves — offers unfiltered feedback, crucial to iterative development.
And so NeuroRestore has enabled itself to go toe-to-toe with pioneering BCI enterprises, all the while benefitting from Swiss public funding and project-based student labour. Importantly, their discoveries sit in the public domain, open for all clinicians and potential patients to investigate and critique, rejecting the all-too-predictable trend of scientific discovery becoming hidden behind the closed-source curtain of private companies.
In-House Data Acquisition
I was impressed by how quickly the engineers moved through each feedback cycle. New ideas for neural modelling or stimulation parameters, derived from statistical analysis, can be implemented in the next day’s clinical session and promptly assessed for their real-world viability.
The lab boasts an impressive clinical team of neurosurgeons, specialist physiotherapists and nurses, and engineers trained and qualified to work with the implanted patients in both French and English. They conduct their own clinical sessions with patients for the rehabilitation and restoration of upper-limb movement, lower-limb movement or autonomic spinal cord function. In each session type, the patient is accompanied by a lead-engineer to orchestrate the patient tasks and ensure the recording of data real-time, a note-taker to label this data with cue timestamps and patient commentary, and typically a physiotherapist to aid in the safe completion of the commands given by the engineer(s).
The team’s joint understanding of the data acquisition needed to overcome their week-to-week engineering challenges, infused with an intimate understanding of the clinical realities of patient sessions, are the focus of weekly team meetings. The motor tasks scheduled to be performed in the daily clinical sessions are dynamically adjusted, ensuring the engineers are empowered with the data needed to train the inference models. The end goal remains fixed: empowering each patient to overcome the difficulties induced by their SCI in a meaningful way, respective to the realities of their at-home living. The metrics used by the lab to quantify patient improvement reflect this philosophy.
Due to the mental load of these sessions, patient time is limited. Start-of-session recalibration to account for neural drift consumes valuable time and technical problems which require on-the-spot problem solving are not uncommon. Engineers think on their feet, physios offer precise instructions, and note-takers furiously record key events, all the while ensuring that the patient is treated with the love and respect befitting a person pioneering the next discoveries of neuroscience.
The proprietary datasets acquired from these sessions are invaluable. They represent the nectar which feeds NeuroRestore’s powerful motor-intention inference models and perhaps the core factor behind NeuroRestore’s success.
Computational Underpinnings
My own internship project (done alongside involvement in regular clinical sessions, as is the norm for every engineer) focused on using computational modelling to tailor Transcutaneous Spinal Cord Stimulation (TSCS) for precision medicine. Outside the more invasive EES, TSCS is a promising treatment option for SCI indication, but its effects vary considerably between subjects as per their unique anatomy. Creating a ‘digital twin’ of each subject based off MRI scans allows us to simulate what the stimulation would look like and whether current would safely hit the dorsal roots required for motor recovery.
This represents one of a range of NeuroRestore projects focused on modelling session-based stimulation for optimisation, the intention being to reduce error and delays in clinical settings. Given the breadth of different severities of SCI, how could a one-size-fits-all treatment regiment ever work? NeuroRestore knows that it needs to find creative ways to scale success in human trials, where patients are worked with very closely, to a clinical product applicable to many. It needs to prove that the observable and quantifiable patient recovery is a direct result of their stimulatory technology; that the improvement is not only attributable to the intense physiotherapist attention the patient receives during rehab.
This represents another pillar of the lab’s success — the resources and manpower allocated to the computational coal-face is critical for the challenge of scalability.
One recent marvel to come out of the lab is BrainGPT, the inference model tasked with matching neural recordings with motor intention in real time — the same real-time decoding challenge at the heart of the lab’s brain–spine interface work (Lorach et al. 2023). Under the hood, its architecture is reminiscent of the BERT model which powers modern LLMs, and inspired by the JEPA training paradigm championed by Yann LeCun. As the engineering team revealed to me, this architecture effectively converted the raw multichannel ECoG into a much more semantically meaningful sequence of vectors. Once the model learns from these vectors (rather than raw signals), it is more robust to the aforementioned challenge of between-patient variability.
Evidently, the lab benefits from its focus on recruiting talent with experience in both neural data and in the cutting-edge trends of machine learning at large.
Now, patient scalability is not the only future hurdle faced by NeuroRestore (and similar BCI labs at their stage of clinical trial success) — achieving powerful edge-computing, reading sufficiently high-resolution neural signals and modelling neural drift between sessions are but a few more. But the wide-reaching and session-focused template set out by NeuroRestore provides the framework of effective translational research for the neurotech world at large.