Research
Five hard, unsolved problems.
Foundational gaps that have resisted decades of effort, selected because they are tractable using formal and computational methods.
-
HP-01
Formal verification of AI clinical decision support
AI clinical decision tools are deployed across health systems with no mathematical proof that they cannot recommend harmful treatment sequences. The MHRA 2026 AI device framework requires safety cases. No accepted formal methodology exists.
-
HP-02
Non-invasive intracranial pressure — the signal fusion gap
Neurosurgeons still drill skulls to measure brain pressure. Every non-invasive approach — transcranial Doppler, optic nerve sheath, tympanometry — fails calibration. The 95% limits of agreement remain ±7–15 mmHg, clinically unacceptable.
-
HP-03
Pre-clinical sepsis trajectory modelling
Sepsis kills 11 million people annually. Existing AI tools detect sepsis after clinical recognition — when the patient is already deteriorating. The pre-clinical trajectory has never been formally modelled. Point-in-time scores cannot reason about trajectory dynamics.
-
HP-04
Information-theoretic limits of surface physiological sensing
The wearables industry has invested billions without asking a foundational question: what physiological information is theoretically recoverable from surface measurements, given the physics of signal propagation through biological tissue?
-
HP-05
Causal digital twin for personalised clinical decision
Hospitals are buying correlation engines. Clinical AI predicts but cannot answer counterfactual questions: what would happen if we intervened now versus in four hours? The gap between predictive and causal clinical AI has not been bridged by any deployed system.