SyneuraX
The SyneuraX AI science programme
SyneuraX captures responses, movement and care trajectories in virtual reality. This map pairs each data layer with each algorithm family: its purpose, the performance reported in the literature, and how strong that evidence is.
Each cell pairs a data layer with an algorithm family, and its number counts the applications. Select a cell to see them.
| Bayesian adaptive | Psychometrics | Cognitive modelling | Hierarchical and longitudinal | Signal processing | Machine learning | Deep learning | Bandits and RL | Language models | RAG and knowledge | Speech AI | Generative content | Anomaly detection | Causal and trial design | Privacy-preserving | Simulation and twins | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Per-trial responses | ||||||||||||||||
| Level parameters | ||||||||||||||||
| Session history | ||||||||||||||||
| Cognitive profile | ||||||||||||||||
| Hand and arm motion | ||||||||||||||||
| Head and posture | ||||||||||||||||
| Gaze | ||||||||||||||||
| Physiology | ||||||||||||||||
| Speech and language | ||||||||||||||||
| Clinician decisions | ||||||||||||||||
| Clinical records and scales | ||||||||||||||||
| Content and stimuli | ||||||||||||||||
| Technical telemetry | ||||||||||||||||
| Multi-centre data |
Figures come from published studies of comparable systems and populations. They are not SyneuraX results.
62 applications
Picks the most informative next stimulus in a short assessment block, reaching each person's threshold in few trials.
In the method paper, under 30 trials sufficed for threshold and about 300 for slope.
Optimises the test design using group and individual priors, shortening assessment.
Moderate evidence, mostly from vision science.
Uses the person's trend to suggest the next session's level to the clinician, supporting adherence.
In a 224-patient multicentre trial, adaptive training improved compliance (p=0.009) while the primary outcome was equal (OR 1.02).
Estimates the probability of skill mastery from success and error sequences.
Established in education; little tested in rehabilitation.
Learns learning patterns from long sequences to predict next performance.
Its advantage has been seen only on very large datasets.
Learns the best option per person for low-risk choices such as exercise order.
Clinical evidence is limited: one study of 27 people with diabetes. Personalisation must be tested as real rather than spurious.
Learns a long-horizon difficulty policy from care trajectories, tested first in simulation.
Estimates from observational data are prone to bias and need very careful evaluation design.
Aligns exercise intensity with the person's physiological signals.
Our review found no randomised trial of this approach.
Shows whether an algorithm truly beats a transparent rule at equal dose.
Reference effect: in stroke, VR versus alternative therapy has a small effect (SMD 0.20, low certainty). SMART and micro-randomised designs suit an effect of this size.
Models measurement error and returns a value with an uncertainty interval for each index.
Analysis choice alone moved reliability from about 0 to above 0.9.
Measures each index's stability across repeat sessions so only dependable ones reach the report.
Accepted ICC bands: below 0.5 poor, 0.5–0.75 moderate, 0.75–0.9 good, above 0.9 excellent. Cognitive-task reliability has been reported from 0 to 0.82.
Shows how many low scores are normal in a healthy person, preventing false alarms.
In one battery, 45.9% of people had at least one score at or below the 16th percentile.
Separates real patient change from measurement noise and practice effects.
A standard clinical method; higher reliability makes smaller changes detectable.
Estimates each person's trajectory despite irregular sessions and missing data.
An established method for analysing clinical change.
Separates processing speed, caution and motor time in two-choice tasks.
Test-retest above r=0.70; about 68% accuracy for MCI detection.
Estimates how quickly an initiated response can be stopped.
Consensus calls for at least 50 stop trials; the best estimator reached r=0.57 in simulation.
Achieves the precision of a full item bank with far fewer items.
11 items reached r=0.94 against the full bank, in simulation.
Tests the domain structure against data and builds valid composite scores.
Established; the unity of some constructs such as attention and inhibition is debated.
Places each person against peers of similar age, sex and education.
Strong in brain imaging; applying it to VR behaviour is a new research direction.
A simple, interpretable model every more complex model must beat.
Across 282 comparisons, machine learning showed no advantage over logistic regression in low-bias studies. External validation needs at least 100 events.
Flags signs of mild cognitive impairment from exercise performance for further assessment.
Pooled AUC of about 0.95 is reported, mostly from small case-control samples with little external validation. Machine learning did not raise pooled accuracy.
Combines behaviour and movement to predict disease course or treatment response.
In one review, 20 of 21 abstracts recommending clinical use lacked external validation.
Simulates the response to different exercise scenarios from each person's model parameters.
Our review found no direct clinical evidence; a long-term research direction.
Keeps each patient's data on one side of the train/test split so true accuracy is measured.
Data leakage can overstate accuracy by 29–55%.
Extracts movement time, peak speed and hand path to track motor performance.
In 39 healthy adults, Quest 3 against OptiTrack: 0.68–3% movement-time error and about 4.8–8.8% speed error.
Follows session-to-session range of motion with no added equipment.
Controller error up to 13.52 mm, mostly in static conditions.
Measures motor control quality beyond speed and range.
On Vicon in 31 stroke patients, ICC was 0.912 and 0.911; validation on Quest is the next research step.
Estimates the clinical motor score from movement data, enabling assessment between visits.
With wearable IMUs: 7% normalised error in 11 people. A Quest-only estimator has not been published.
Measures manual dexterity with a task mirroring the clinical test in VR.
Correlation with the physical test is reported as moderate to high; agreement between the two still needs establishing.
Detects when a patient substitutes trunk movement for arm movement.
In a preprint, macro-F1 fell from 0.80 in healthy people to 0.43 in patients; a trunk sensor may help.
Estimates static balance from head motion.
A force-plate study reported correlations that vary by metric.
Extracts hand tremor features to follow movement disorders.
No Quest validity study was found; a wrist sensor is a better candidate.
Learns movement representations from unlabelled data so models need fewer clinical labels.
No direct comparison with simple models on VR data was found; must be judged on patient-level splits.
Adds fixation, dwell time and antisaccade error alongside behavioural responses.
In a preprint on Quest Pro, about 1.04° accuracy at 89 Hz. Quest 3 and 3S have no eye tracking.
Follows session physiological load with a wrist or chest sensor.
In 931 people, optical sensing estimated HRV with non-uniform bias; ECG is needed for clinical outcomes.
Estimates emotional state from several signals.
Reported accuracy is limited, and this use is a distinct high-risk category under the EU AI Act.
Software computes the numbers, the model writes the prose and the clinician signs off.
In the NOHARM benchmark (preprint) over 80% of severe errors were omissions, so a mandatory-findings checklist is central to the design.
Turns a clinician's request into valid parameters, passing a validator before use.
No clinician study was found yet; the validator and human approval block invalid output.
Answers clinicians with direct quotes from approved sources and declines when none applies.
Hallucination rose from 5 per 100 in one turn to 40 per 100 after 10 turns; single-turn design is safer.
Shows the clinician several reasoned options; the decision always stays with the clinician.
No randomised trial in rehabilitation was found, and automation bias must be designed against.
Shows clinicians unexpected changes in a patient's course earlier.
Should first be evaluated silently on multi-patient longitudinal data.
Generates report text on the clinic computer, so data stays on site.
On PersianMedQA, Persian-branded models scored 34.87% and 24.22% against 83.09% for GPT-4.1, so model choice needs Persian testing.
Converts speech to text and timing so the language domain can be measured.
Preliminary Persian norms exist. Whisper can invent text during pauses, so human review is required.
Studies speech features as early-sign research markers.
Samples are small: ADReSS-M had only 46 test speakers.
Builds a consented Persian speech collection, which is scarce in this field.
Published Persian studies have around 50 participants, and Persian and English patterns differ.
Turns performance in the Persian word exercise into language measurement indices.
Needs a validity study against standard language tests.
Delivers exercise instructions in the patient's language with natural voice, important for low literacy.
Pre-produced, human-reviewed text removes the risk of live errors.
Guides and encourages patients with clinically approved lines.
One study covered 21 healthy people outside VR; no study with older adults or patients in VR was found.
Free conversation with patients for engagement and support.
A meta-analysis found no low-bias trial, and data on older adults was very scarce.
Creates equivalent versions of each exercise so repetition does not become memorisation.
There is real evidence for semi-parallel forms; equivalence must be calibrated on people.
Drafts new stimulus ideas for designer review.
No evidence was found for psychometric equivalence of LLM-generated cognitive items.
Produces varied, locally relevant environments faster.
The effect of environment changes on measurement must be tested so scores stay comparable.
Speeds translation and cultural adaptation, then tests psychometric equivalence.
Machine translation alone does not carry over test validity; local norms are needed.
Removes invalid sessions and trials before indices are computed.
A limit of about 3.2 m/s for implausible hand speed is a useful starting point.
Suggests a break before symptoms worsen.
A systematic review found no model predicting a first event.
Identifies patients at risk of stopping care for earlier follow-up.
A real need: in FINGER only 12% completed all sessions.
Separates sessions with outside help from measurement.
No validated method was found; manual recording is the better starting point.
Evaluates model performance separately by age, sex, education and accent.
Model performance drops in low-resource languages (preprint).
Detects tracking, performance and network problems before they affect a session.
An engineering practice, judged by fewer lost sessions.
Combines aggregate clinic results without moving raw data.
Benefit is limited with few sites, and differential privacy can reduce utility (preprint).
Tests software and parameter recovery without patient data.
Defensible for testing; not a substitute for clinical samples, and no privacy guarantee on its own.
Shows which type of exercise worked better for this particular patient.
Our review found no study in VR rehabilitation.
Deterministic code computes every clinical number. Language models write prose only.
Algorithms suggest. Clinicians approve or change the care plan.
Complex models must outperform a transparent method in a registered comparison with patient-level data splits.
Reliability and validity come before prediction, because no model beats the measurements it learns from.