NVIDIA FLARE
Federated learning across clinics: models train on local data and only model updates travel.
Technology
AI-ready by architecture.
Every SyneuraX session produces structured clinical data, and that data stays on a station inside the clinic. The same station is designed to become the clinic's AI node, running models locally on GPU hardware.
Data never leaves the clinic. Models travel.
The architecture
Six layers, from the headset to the learning network.
The first four layers are built and running in clinics today. The last two are the next generation, designed around the NVIDIA AI stack.
- 01Built
Capture
Standalone headsets record every response, score, timing and movement as the patient trains.
- 02Built
Private network
Five dedicated channels between headsets and the station over the clinic's own network. No internet needed.
- 03Built
Edge station
The Therapist Station runs and records every session locally and is the natural home for on-site AI.
Next on this layer: RTX GPU inference with TensorRT and NVIDIA NIM.
- 04Built
Clinical data
14 data layers and 113 clinical measures across 9 cognitive domains, with full session history. Programs sync between stations; patient data does not.
Next on this layer: GPU data science with NVIDIA RAPIDS.
- 05Next
Intelligence
Measurement models, a clinician copilot, adaptive therapy, and speech and movement AI, running at the edge.
Designed for NVIDIA NIM, NeMo Guardrails, NeMo Retriever, Riva and Holoscan.
- 06Next
Learning network
Models improve across clinics and regions while every patient record stays where it was created.
Designed for NVIDIA FLARE federated learning.
Anatomy of a session
From one movement to a clinical decision.
Every step below runs today in clinics, inside the clinic's own network.
- Move and respond
The patient reaches, grabs, remembers and decides inside the exercise.
- Stream
Each response and movement travels instantly over the private clinic network.
- Record
The station stores every response with its timing, on site.
- Measure
Each level becomes clinical measures: accuracy, speed, smoothness, range and more.
- Score
Measures roll up into nine cognitive domains, with trends across sessions.
- Decide
The clinician reads the report and shapes the next session.
From movement to meaning
A measurement engine, not a scoreboard.
Raw responses are distilled, layer by layer, into a profile a clinician can act on. Nine more exercises are being mapped into the same structure.
13exercises mappedEach records its own responses and timings
113clinical measuresAccuracy, timing, errors, strategy
48subdomainsSuch as working memory and response inhibition
9cognitive domainsFrom attention to numerical cognition
1clinical pictureOne profile per patient, over time
- Executive functions28
- Memory26
- Attention20
- Perception11
- Processing speed7
- Temporal processing6
- Sensorimotor6
- Visuospatial5
- Numerical cognition4
29 live metrics from a single motor exercise
Performance
ScoreTimeAccuracySuccessfulFailedEvents
Motion quality
Smoothness indexLeft jerkRight jerkLeft tremorRight tremorLeft efficiencyRight efficiency
Range and reach
Total distanceLeft range of motionRight range of motionMax left velocityMax right velocityAvg left velocityAvg right velocity
Laterality
Left usageRight usageMidline crossings
Timing and learning
Active timePause timePause countReaction variabilityResponse consistencyLearning slope
The data foundation
Fourteen layers of data, ready for AI.
The AI map pairs each of these data layers with the algorithm families that can learn from it.
BehaviourBodyLanguageClinicalPlatform
Per-trial responsesReaction time, accuracy, error type
Level parametersDifficulty and exercise design
Session historyLongitudinal course
Cognitive profile113 indices across 9 domains
Hand and arm motionController and hand tracking
Head and posture6-DoF headset pose
GazeEye-tracking headsets
PhysiologyWearable sensors
Speech and languageVoice and word tasks
Clinician decisionsPlans, notes, overrides
Clinical records and scalesDiagnosis, MoCA, Fugl-Meyer
Content and stimuliItems, environments, voice
Technical telemetryTracking, frames, network
Multi-centre dataPooled across clinics
Layers 01 to 03, live
How the data moves today.
Designed for the NVIDIA stack
Where each NVIDIA technology fits.
We chose an edge-first architecture so that AI can run where the data is created. These are the NVIDIA technologies that layer maps to, and what each one will do in SyneuraX.
RTX + TensorRT
Fast, offline inference on a GPU inside the Therapist Station.
NVIDIA NIM
Packaged language and vision models running on site for the clinician copilot.
NeMo Guardrails and Retriever
Keeps the copilot on clinical rails and answers from approved guidelines only.
Riva speech AI
Speech recognition and voice for language exercises and guidance, starting with English and Arabic.
Holoscan
Real-time processing of motion and sensor streams for movement analysis.
RAPIDS
GPU data science for reliability studies, norms and long-term outcome analysis.
CloudXR
Streams high-fidelity environments from the station's GPU to wireless headsets.
Omniverse and OpenUSD
Faster creation of exercise worlds and digital twins of treatment rooms.
Clara and IGX
A path toward medical-grade edge AI as SyneuraX moves toward clinical claims.
Where the compute lives
Edge first. Cloud optional.
Today the work is split between the headset, the station and the optional central library. AI adds to the station, so it can run where the data is.
Headset
Standalone, wireless
- Exercise logic and scoringBuilt
- Motion captureBuilt
- Voice guidanceBuilt
- Simple on-device quality checksNext
Therapist Station
Inside the clinic
- Session control and live viewBuilt
- Recording every responseBuilt
- Measures, domains and reportsBuilt
- GPU inference for AI modelsNext
- Clinician copilotNext
Central library
Optional, online
- Programs and level setsBuilt
- Domain designsBuilt
- Federated model updatesNext
The learning network
Models travel. Patient data stays.
With federated learning, each clinic trains on its own data and shares only model updates. The shared model improves for everyone, and no patient record leaves the building.
Designed for NVIDIA FLARE.
Local model update, sent outImproved model, sent backPatient data never moves
Already on GPU
Our content factory runs on NVIDIA RTX today.
The 3D renders, exercise worlds and visuals on this site are produced in-house with GPU ray tracing on NVIDIA RTX hardware. The same habits carry over to AI: build on the GPU, keep the work in-house, measure everything.
Why it matters
The moat is the data, not the model.
- Consented, longitudinal and multi-modal rehabilitation data
- Collected in clinics in under-served languages
- Owned by the clinic and usable for privacy-preserving learning
- Ready for any model the field produces next
See the full picture
The AI map shows 62 applications with their published evidence. The investor presentation shows the plan.
SyneuraX is an independent company and is not affiliated with or endorsed by NVIDIA. NVIDIA, RTX, TensorRT, NIM, NeMo, Riva, Holoscan, RAPIDS, CloudXR, Omniverse, Clara, IGX and FLARE are trademarks of NVIDIA Corporation. Items marked Next are planned and not yet part of the product.