Sense
Bedside optical capture
NIR light and a forehead probe acquire blood-flow signal
SafeICP is not only a deviceIt is a measurement pipeline
SafeICP combines bedside optical sensing, signal interpretation, and machine learning-based ICP estimation into one non-invasive monitoring approach.
Read
Dynamic signal patterns
Speckle dynamics reveal cerebral hemodynamic behavior
Infer
ICP-relevant output
ML maps temporal patterns to pressure estimates
Probe Context
Light goes inUseful dynamics come back
SafeICP uses safe near-infrared illumination and a forehead probe to read out blood-flow-related behavior without placing the sensing hardware inside the skull.
Placement
Forehead probe
The optical interface sits outside the head, at the bedside.
Illumination
Near-infrared light
Light travels through tissue and interacts with moving blood cells.
Readout
Speckle and flow dynamics
Fluctuation patterns become a non-invasive signal linked to cerebral blood flow.
The device does not read pressure directlyIt learns the mapping
SafeICP measures optical dynamics tied to blood flow, not ICP itself. To produce a pressure estimate, the system needs a learned time-series translation from signal behavior to pressure behavior.
Measured
Optical time series
Flow-related fluctuations, speckle behavior, and temporal signal structure.
Predicted
Intracranial pressure estimate
Clinically meaningful pressure output aligned to paired invasive reference data.
The prediction tracks the shape,not just the average
A representative InceptionTime example shows that the model follows the overall temporal behavior of invasive ICP well enough to make the learned relationship visually credible.
What to noticeThe solid blue line is the mean prediction averaged across five cross-validation folds. The shaded blue band is the 95% confidence interval across those folds. Together they follow the broad temporal structure of the green invasive ICP reference.
Where it remains hardSharp transitions and sudden ICP spikes are still more difficult to capture than slower baseline trends. The confidence interval widens noticeably around these rapid changes, reflecting higher model disagreement at the hardest moments.
Competitive errorCautious confidence
SafeICP reaches competitive MAE, but its uncertainty remains under-calibrated.
Best overall MAE
5.3 mmHg
mWDN on the right hemisphere
Low-error zone
52%
Predictions within 0–4 mmHg
PICP coverage
37%
Too low for trustworthy confidence
ModelMAE LeftMAE RightRead
InceptionTime
6.0
5.4
CompetitivemWDN
6.3
5.3
Strongest overallTCN
7.1
5.6
Useful baseline
Read of the chart
Dense low-error region52% within 0–4 mmHgCoverage still lags
Accuracy is now a matter ofclinical refinement, not feasibility
SafeICP already produces error levels that are close enough to clinical benchmarks to focus the next phase on generalization, cohort breadth, and stability on unseen data.
Where the work shifts now
The sensing stack is credible. The remaining gap is reducing the validation-to-test spread and proving robustness across broader patient conditions.
Reference standard
WHO invasive ICP thresholds
0–20 mmHg±2 mmHg
>20 mmHg±10%
Observed performance
Model error on held data
Validation2.6 mmHg
Test5.3 mmHg
Interpretation
The result is strongest on validation and still competitive on test, which makes the next milestone less about inventing a new approach and more about tightening generalization.
Four partners coveringthe whole translation chain
Each partner covers a part of the pipeline the others cannot.
Already operational
The collaboration is already operating across data, devices, and bedside validation.
Photonics & Device
SCOS hardware and optical sensing
Machine Learning
Time-series models and uncertainty analysis
Clinical Translation
Recruitment, validation, and bedside workflow
Industrial Validation
Safety engineering and product validation
The people behind SafeICP
Researchers, clinicians, and engineers from four institutions, each contributing to the domain
Institute of Photonic Sciences
Turgut
Durduran
Principal Investigator
Mirko
Fornasier
Doctoral Researcher
Carolina
Vega
Doctoral Researcher
Monica
Torrecilla
Doctoral Researcher
Pompeu Fabra University
Viacheslav
Danilov
Research Scientist
Gemma
Piella
Professor
Anton
Makoveev
Research Scientist
Vall d'Hebron Hospital
Maria
Poca
Head of Neurosurgery
Juan
Sahuquillo
Neurosurgeon
Murad
Al-Nusaif
Doctoral Researcher
ProCareLight
Youcef
Lebour
Research Engineer
Questions & Answers
Is it possible to directly measure ICP non-invasively?
Not reliably. A direct ICP sensor must communicate with the intracranial CSF/brain compartment, so non-invasive signals can only estimate or trend ICP.
How can ICP be estimated without measuring pressure directly?
The device captures optical dynamics tied to cerebral blood flow, then learns a mapping from synchronized optical windows to invasive ICP references.
Which model is strongest in the current comparison?
mWDN is the strongest overall in the reported results, reaching 5.3 mmHg MAE on the right hemisphere and the highest concentration of low-error predictions.
What makes the current result clinically credible already?
The models are trained and evaluated against invasive ICP, not a proxy label, and they reach competitive error in the 5-6 mmHg range while tracking temporal ICP behavior.
What is the main limitation today?
The hardest problem is elevated ICP and trustworthy uncertainty. Coverage stays well below the nominal target, so confidence estimates are still too optimistic.
Why compare several model families instead of picking one architecture first?
Because this is a time-series translation problem. InceptionTime, mWDN, and TCN test different temporal biases and reveal which one fits the paired signal-to-ICP mapping best.
