A structured response to a recurring clinical problem

Motivation

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.

The project's premise is that a well-constructed visual constraint representation can extend rather than replace clinical reasoning by making the multivariate physiological state of the patient legible at a glance, while leaving the underlying numerical evidence one click away for any clinician who wishes to interrogate the synthesis. This is in the spirit of ecological interface design as applied to other high-stakes process-control domains, adapted to bedside critical care.

Research objectives

The project pursues five linked objectives.

1. Develop instance prototypes for major ICU subsystems. Ventilation and continuous renal replacement therapy are the first two instances. Hemodynamics, sedation depth, and electrolyte management are candidate subsequent instances. Each instance applies the same constraint-state-space architecture to a different physiological subsystem, with the goal of establishing the paradigm as generalisable rather than ventilator-specific.
2. Anchor every constraint to published evidence. Zone thresholds in the visualisation are not improvised. Every band, cutoff, and suggestion rule traces to a citable source — landmark trials, established physiological constants, current ICU practice guidance. Provenance metadata is preserved alongside each threshold in the implementation so that the basis for any boundary is always available for inspection and so that thresholds can be updated as the underlying evidence evolves.
3. Validate the synthesis against unaided clinical reasoning. A pilot scenario-based study is planned in which clinicians will be presented with patient data under two conditions — raw data alone, and raw data alongside the constraint model — and the difference in diagnostic speed, accuracy, and management completeness will be quantified. A parallel dual-database retrospective validation across MIMIC-IV and eICU-CRD will test whether the suggestion rules generalise beyond a single training cohort.
4. Publish the methodology before claiming the implementation. The first paper describes the framework — the constraint-state-space formulation, the dead-space synthesis, the design rationale — independent of the working software. The second paper reports the validation results. This ordering protects the methodological contribution from being conflated with code-level decisions and gives the framework standing in its own right.
5. Articulate the paradigm as a scaffold for interpretable clinical AI. The visual constraint model is structured reasoning made explicit. As clinical AI moves toward neuro-symbolic and scaffold-first architectures, the project positions the state-space framework as a concrete instance of the structured scaffold within which machine learning can later operate without sacrificing interpretability. This is held as a longer-term framing rather than a current implementation claim.

Principal investigator