The intelligence is an ensemble with strict boundaries. Imaging models segment vessels and candidate findings, register examinations across time and quantify change. Language systems recover relevant facts from earlier reports but always link back to the source document. Temporal models combine observations without treating missing data as normal. A separate safety layer verifies that the requested use matches the model's authorised population, modality and clinical purpose. The system may prioritise or draw attention; it cannot declare that a person is safe, replace a full evaluation or autonomously choose treatment. Its most important output may be a well-calibrated statement that the evidence is insufficient.
PROJECT 015 · Research venture
Quiet Signal
“Can we see neurovascular danger before it becomes an emergency?”
MOVE TO EXPLORETHE IDEA IN ONE LINEvessel
01 / PREMISE
The beginning
Quiet Signal begins with a difficult asymmetry. The brain can change catastrophically in minutes, while the conditions that make a stroke or aneurysm dangerous may develop quietly over years. A scan is normally read as a moment. A blood-pressure value is a moment. A brief neurological examination is a moment. I imagine a system that can remember those moments together and help a specialist notice when the pattern has changed. The aim is not to predict every stroke, screen everyone for aneurysms or place a diagnostic alarm in a consumer app. It is to build a careful, longitudinal neurovascular intelligence layer for people whose history, symptoms or known findings already justify clinical attention.
Neurovascular care is fragmented across time and institutions. Previous CT, CTA, MRI and MRA examinations may sit in different archives. Reports summarise what one reader considered important, but subtle morphology, vessel changes or disagreement between measurements can be difficult to compare years later. Risk factors such as hypertension, atrial fibrillation, smoking, diabetes, medication and family history live in other systems. When acute symptoms appear, speed matters, yet clinicians must reconstruct the person while also interpreting new imaging. Existing AI tools can help triage specific image findings, but a notification is not a diagnosis and many tools are intentionally limited to narrow indications. The missing object is a trustworthy memory of the individual neurovascular system: what is known, what changed, what remains uncertain and what needs a human decision now.
02 / THE PRODUCT
What it could become
Quiet Signal would be a clinical workspace with three clearly separated modes. Prevention mode helps authorised care teams assemble modifiable risk factors and identify evidence-based follow-up questions; it does not recommend population-wide aneurysm screening. Surveillance mode aligns serial vascular images, measures known aneurysms and other relevant findings, visualises uncertainty and highlights changes that deserve specialist review. Acute mode analyses new imaging in parallel with standard interpretation, prioritises suspected time-sensitive findings and retrieves the most relevant prior context for the neurovascular team. For selected patients already under care, short supervised speech, facial symmetry, gaze, fine-motor or balance tasks could establish a personal baseline, but any detected deviation would prompt clinical assessment rather than generate a diagnosis. Every output would show its source, confidence, comparison and intended use.
For whom
- People with a known unruptured intracranial aneurysm under specialist surveillance
- People with previous stroke, transient ischaemic attack or elevated neurovascular risk
- Vascular neurologists, neuroradiologists and neurointerventional teams
- Emergency departments and stroke networks coordinating time-critical care
- Researchers building better longitudinal evidence for brain vascular disease
Core capabilities
- Longitudinal registration and comparison of CT, CTA, MRI and MRA
- A personal neurovascular timeline linking images, reports, risk factors and interventions
- Aneurysm morphology and growth measurement with explicit uncertainty
- Acute image triage and specialist notification in parallel with standard care
- Optional supervised speech, vision and motor baselines for selected clinical pathways
- Audit, provenance, bias monitoring and structured clinician override
The project becomes meaningful only when a new technical possibility is translated into a clear human advantage, an experience people can understand, and a system capable of earning trust over time.
03 / FOUNDATIONS
Intelligence and mathematics
The core is a time-varying geometric graph of the cerebral vasculature. Vessel segments are edges; bifurcations, stenoses, aneurysms and anatomical landmarks become typed structures whose geometry can be compared across registered scans. Bayesian hierarchical models separate true biological change from scanner, protocol and measurement variation. Survival and competing-risk models can support research into personalised follow-up, while conformal prediction or calibrated uncertainty sets communicate where a model's estimate is unreliable. Change-point detection searches for meaningful deviation from an individual baseline. Decision-curve analysis asks the practical question: at a given threshold, does using the model produce more clinical benefit than harm? No single risk number should collapse anatomy, time, evidence quality and patient preference.
04 / VENTURE LOGIC
How it might live
I would develop Quiet Signal as regulated clinical software with hospitals and neurovascular specialists, not as a direct-to-consumer diagnosis product. A first narrow product could focus on longitudinal surveillance of known unruptured aneurysms: automatically retrieving prior studies, aligning images, measuring change and preparing an inspectable comparison for the clinician. A second pathway could support stroke-network workflow and prior-context retrieval. Hospitals could subscribe per site or care network, with research collaborations used to build diverse longitudinal datasets and prospective evidence. The durable advantage would not be an opaque model alone, but trusted integration, longitudinal data quality, evaluation across devices and populations, and a workflow that saves specialist time without hiding uncertainty.
For me, a venture is more than an interesting technology. It needs a narrow first user, a repeated problem, a distribution path, a credible advantage and a reason to improve as more people use it. I would test those conditions before deciding whether this idea should become a company, a product, an open technology or an ongoing research programme.
05 / DESIGN PRINCIPLES
Rules for making it real
- 01
Time is clinical data: never interpret a new image without searching for the relevant past.
- 02
Triage is not diagnosis, and a risk estimate is not a treatment decision.
- 03
Do not encourage population screening where evidence does not support it.
- 04
A model must reveal uncertainty, source evidence and its authorised purpose.
- 05
Optimise for earlier appropriate human attention, not for more alerts.
- 06
Build with patients, neurologists, neuroradiologists, emergency teams and regulators from the beginning.
06 / ROADMAP
From question to company
- 01Frame
Choose one narrow clinical question: longitudinal comparison of known unruptured aneurysms.
- 02Prototype
Create a multi-centre retrospective dataset with scan protocols, repeated measurements and adjudicated outcomes.
- 03Prove
Build registration, segmentation and uncertainty-aware change measurement, then compare it with specialist variability.
- 04Build
Run a silent prospective study to measure errors, workflow effect, subgroup performance and alert burden.
- 05Advance
Seek regulatory clearance for the narrow intended use before clinical deployment.
- 06Advance
Only then investigate adjacent stroke triage, prior-context retrieval and supervised neurological baselines.
07 / HONEST QUESTIONS
What could go wrong
Serious imagination includes the possibility that an idea should change radically—or should not exist. These are the tensions the project would need to resolve:
- A false negative delaying urgent assessment or creating dangerous reassurance.
- A false positive increasing anxiety, unnecessary imaging or invasive intervention.
- Scanner, protocol and population differences producing unequal performance.
- Incidental findings expanding surveillance without clear clinical benefit.
- Sensitive imaging and neurological data being reused beyond meaningful consent.
- A polished visualisation making uncertain morphology appear more precise than it is.
- Combining prevention, surveillance and emergency triage into one misleading score.
EVIDENCE / BOUNDARIES
Built from evidence, not medical theatre
This is an early research idea, not a medical product or medical advice. Its intended uses, datasets, claims and evaluation would need to be defined with clinicians, patients and regulators before any real-world deployment.