Three traditions, one framework

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 any clinician 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 — classically estimated at seven plus or minus two by Miller (1956), with more recent estimates lowering this further. Modern ICU bedside data routinely and substantially exceeds those bounds.

Overview

The methodological grounding of the project is best understood as the intersection of three established research traditions: ecological interface design, evidence-anchored physiological modelling, and the scaffold-first paradigm increasingly recognised in clinical AI. The constraint-state-space formulation is what emerges when all three are applied simultaneously to the problem of bedside ICU decision support.

Ecological interface design

The ecological-interface-design tradition was initiated by Rasmussen and Vicente in the 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 clinician can engage with directly.

Applied to critical care, this means the patient's clinical state should be representable as a position within a constraint space, where the axes are the physiological dimensions along which the patient is being supported. The polygon visualisation makes these constraints simultaneously visible. Worst-driver highlighting and zone bands map abstract physiological thresholds into a new visual paradigm. 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 rather than derived from local convention. Zone bands are calibrated against landmark trials and established physiological thresholds, for example:

• Driving pressure cutoffs 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 ARDSnet (2000)
• Vd/Vt is computed by the Enghoff modification of the Bohr equation

Suggestion rules encode published differential-reasoning patterns rather than improvised heuristics. Provenance metadata is preserved 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 what is increasingly termed knowledge-guided machine learning (KGML) — an approach in which domain knowledge and established constraints are embedded directly into the ML architecture, rather than leaving the model to derive all structure from data alone. This contrasts with the currently dominant paradigm in clinical AI: imitation learning from retrospective outcome-labelled datasets, in which models reproduce the decisions found to be associated with better outcomes in the training cohort. This approach intrinsically inherits the variation, biases, and dated practices of that cohort together with a fundamental opacity about how those decisions were reached.

The knowledge-guided alternative encodes the reasoning structure of a problem first, then machine learning operates within that structure. A widely cited example is the Google DeepMind project AlphaFold2 (Jumper et al., 2021), which achieved a step-change in protein structure prediction by embedding evolutionary, physical, and geometric constraints — including knowledge-based scoring functions — directly into the neural network architecture, so that the model reasoned within an explicit domain-knowledge framework rather than from data alone. The Critical Care State Space Project applies the same principle to bedside physiology: encoding evidence-anchored physiological constraints as an explicit reasoning structure — what the project terms a constraint scaffold — within which future ML can operate. The aim is to improve accuracy and usability whilst preserving interpretability — the reasoning behind any output remains legible and open to clinical scrutiny.

The formal language underlying this approach is Bayesian. Zone thresholds in the constraint model function as evidence-based priors calibrated against landmark trials rather than local convention. Clinical reasoning at the bedside is itself a process of updating beliefs as new data arrives. Bayesian networks are therefore the natural candidate architecture for the ML layer: they encode prior knowledge explicitly, propagate uncertainty transparently and produce auditable outputs at each inferential step. This is precisely what is lost in imitation learning, and what the constraint scaffold is designed to preserve.

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 a methodology paper 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 recommendation.

The project welcomes collaboration from experts within the field to refine and improve the models to create the best possible scaffold for future ML models to work from.

Principal investigator