Engine context
What the VYZR Care RPP Engine (Recognition · Prediction · Prevention) is intended to do, how it works, what it does not do, and what clinicians should know before interpreting its output. The full method is in docs/clinical-methods.md.
Intended use
Prioritise adult general-ward patients for clinician review by recognising evolving trajectories — small changes, repetition, velocity and several weak signals moving together — rather than single threshold breaches. The output supports, and does not replace, clinical assessment and local escalation pathways.
Engine architecture
Temporal features
Per-signal deltas, patient baseline, direction, short/long slopes, acceleration, persistence and freshness from irregular timestamps.
Pattern library
Sodium, respiratory, renal, hemodynamic, bleeding, infection, multisystem and recovery modules, each with its own timeframe and evidence.
Pattern strength
Explainable 0–100 index (weights: magnitude 18, velocity 16, concordance 16, persistence 12; chronic-stable -25, improving -30). Not a probability.
Projection
Short-horizon linear trend with a residual-based uncertainty band; refused below 3 points, under 3 h of span, or when the latest value is stale.
Known limitations
- All patients are synthetic; scenarios were authored to exercise the engine.
- Weights, thresholds and bands are heuristic and uncalibrated. Pattern strength is an explainable index, not a probability of harm.
- Projections are linear and short-horizon; they are refused when data are too few, too short or too stale, and will be wrong when physiology is non-linear.
- The engine sees only recorded observations — not medications, procedures, fluids or the reason for admission unless reflected in the numbers.
- No external validation, regulatory clearance or local calibration is claimed.