Approach

The clinical problem

The project began as a structured response to a recurring clinical problem: a single ICU patient generates more streams of monitored data than clinicians can integrate continuously and accurately, particularly under the cognitive load of overnight cover, multi-patient handover, or simultaneous deteriorations. The working memory available for integrating multiple variables in real time is bounded — originally estimated at around seven items by Miller (1956), and revised downwards by later work (Cowan, 2001). Modern ICU bedside data routinely and substantially exceeds those bounds.

Ecological interface design

The ecological-interface-design tradition was initiated by Rasmussen and Vicente in the late 1980s and early 1990s for high-stakes process-control environments such as nuclear power plant operation. The central EID claim is that interfaces should make the constraints of a given problem perceptually available. This means not just the data, but the structural relationships between data, are displayed in a form that the user can engage with directly.

Applied to critical care, a patient's clinical state can be represented as a position within a constraint space, where the axes are the physiological dimensions along which the patient is being supported. Polygon visualisation makes these constraints simultaneously visible. Worst-metric highlighting and zone bands map published physiological thresholds directly onto the polygon geometry. A 24-hour temporal loop view extends the static snapshot to a trajectory, allowing the clinician to perceive direction and rate of change as well as the current state.

Evidence-anchored physiological modelling

Each constraint axis is grounded in published evidence: zone bands are calibrated against landmark trials and established physiological thresholds, for example:

• Driving pressure cut-offs follow Amato et al. (2015)
• Mechanical power values follow Serpa Neto et al. (2018) and Gattinoni's 2023 review
• Tidal volume per kilogram aligns with the ARDS Network trial (2000)
• Vd/Vt is computed by the Enghoff modification of the Bohr equation

Suggestion rules generated by the model encode published differential-reasoning patterns rather than improvised heuristics. Reference material is provided alongside each suggestion and the basis for any boundary is modifiable as evidence evolves.

Knowledge-guided machine learning

The longer-term framing of the project draws on knowledge-guided machine learning (KGML), in which established domain constraints are embedded directly into a model's architecture rather than left for the model to infer from data alone (Karpatne et al., 2017).

This contrasts with the currently dominant paradigm in clinical AI: imitation learning from retrospective, outcome-labelled datasets, in which a model reproduces the decisions associated with better outcomes in its training cohort. Such models inherit the variation, biases and dated practice of that cohort, together with a fundamental opacity about how any given recommendation was reached.

The paradigm case for the alternative is AlphaFold2 (Jumper et al., 2021), which embedded evolutionary, physical and geometric constraints directly into the network architecture, so that the model reasoned within an explicit domain-knowledge framework rather than from data alone. The analogy has clear limits: protein folding is governed by constraints that are mathematically precise and physically invariant, whereas ICU physiology is heterogeneous, measurement-noisy and patient-specific, and its constraints are probabilistic thresholds drawn from trial evidence rather than physical laws. The architectural principle nevertheless transfers — encode what is already known, explicitly and inspectably, and let any learned component operate within that structure. The project terms this a constraint scaffold.

This is a direction of travel rather than current functionality. No machine-learning component exists in the present prototypes.

Bayesian reasoning

The natural formal language for a constraint scaffold is Bayesian. Clinical reasoning at the bedside is itself a process of revising belief as new data arrives, and Bayesian networks encode prior knowledge explicitly, propagate uncertainty transparently, and produce an auditable inferential chain at each step — precisely what is lost in imitation learning (Pearl, 1988; Lucas et al., 2004).

One distinction is worth stating plainly. The current models are structured by Bayesian reasoning; they do not perform Bayesian inference. The zone boundaries are published thresholds, not fitted priors, and no posterior is computed. Bayesian network integration is a planned step for a future machine-learning layer, not a description of what the prototypes do today.

Trajectory

The project's first priority is to validate the constraint models as a synthesis display tool for human clinicians. This will be described in methodology papers followed by detailed analysis of the model's accuracy and usability on a retrospective set of real-world ICU data using the MIMIC-IV and eICU-CRD databases. The "scaffold-for-AI" framing becomes relevant when that validation evidence is in hand. At that stage, Bayesian network integration is the planned ML step: operating natively within the constraint scaffold, encoding uncertainty explicitly and preserving the full interpretability chain from prior evidence to posterior clinical recommendations.

Principal investigator