Treatment optimisation, not prediction
Diagnosis, triage and alerting tools are useful — but they stop before the treatment decision, and that market is now commoditised. A Minerva app recommends the next treatment action, continuously, as physiology changes. Reinforcement learning is literally the mathematics of agents: systems that choose actions over time to reach an outcome, rather than assigning a probability to a moment.
We pioneered off-line reinforcement learning and off-policy Policy Estimation methods for medical treatment enabling us to learn from operational real-world data. We developed grading based methods for privacy-preserving transfer learning enabling us to use a multitude of multi-modal data, differing from ward to ward, hospital to hospital, region by region, country by country. Our methods realise transform each patient's longitudinal data stream into digital twin, that we can interrogate and perform counterfactual and what-if questions to optimise decision making across a patient's entire trajectory.
Building one such agent is a research problem that we pioneered. Building a family of them, each safe enough to deploy in a hospital and defensible enough to clear a regulator, is a platform problem. That is what the Minerva AI Clinician Platform is for.
- Ingest clinical data
Vitals, laboratory results, prior fluids and vasopressors — 48 time-varying variables from routine care. - Update the model & infer the patient state
A trajectory-aware model of the patient with calibrated uncertainty, decomposed into what is noisy and what is unknown. - Recommend the next action
For Sepsis fluid management we need to prop up the patients circulatory system for them to survive: the AI, delivers an intravenous fluid rate and a noradrenaline-equivalent vasopressor dose recommendation — an action, not a probability. - Clinician reviews
The bedside clinician accepts, modifies or rejects. Advisory decision support, but designed for safety and intepreability at closed-loop standard. - Audit and learnFeedback, safety monitoring and governance. Every AI recommendation and action takeb by clinicians is captured, the system learns on the job and provide accountability.
Risk-aware reinforcement learning under uncertainty
The policy inside each app is a risk-aware reinforcement-learning agent. For Minerva AI Clinician Sepsis it was trained on 96,000 patient records and externally validated on independent data from 128 US intensive care units. It is the first system in its class to combine a risk-aware policy with structural action-masking under uncertainty: the model decomposes aleatoric uncertainty (irreducible noise in the data) from epistemic uncertainty (what the model has not seen), and is structurally prevented from recommending when it should not speak.
Predefined safety constraints refine the policy’s behaviour without degrading its performance, and off-policy evaluation quantifies how the recommended strategy would have performed against observed care — the methodological backbone regulators expect for treatment-recommendation AI. Because these methods live in the core rather than in any one app, each new disease inherits them.
A medical device from the first commit
Our software is designed to IEC 62304 Class B with full requirements traceability, a maintained Design History File, calibration verification, a cybersecurity baseline, and summative usability evidence. The regulatory position — UK Class IIb, EU Class IIa, and a US pathway via De Novo. The quality system, safety case architecture and technical file structure are built once at the platform level. Each subsequent app extends an existing regulatory file rather than opening a new one from scratch — which is what makes a catalogue of regulated clinical apps tractable at all.
