Clinical Decision Support (Comprehensive)

The rule library

~500 rules at go-live, ~50 literature-cited

Sepsis bundles, drug–drug and drug–allergy interactions, renal dose adjustment, VTE prophylaxis, early-warning escalation, and the bundle state machines for the operating room, emergency and ward. Citations surface at the moment of suggestion, so the clinician reads the evidence, not just the flag.

Hospital-authored growth

Over the first year a typical library grows well beyond the foundation — rules added, revised and retired continually by the hospital, every change versioned.

Automated clinical scores

Twelve foundational scores recalculate continuously as data arrives, each value carrying its literature citation. The same engine activates additional scores at go-live and supports hospital-authored scores throughout.

The foundational twelve

SOFA, qSOFA, APACHE II, NEWS2, CAM-ICU, KDIGO AKI staging, ARDS severity (Berlin), shock-type classification, SAT/SBT readiness, VTE adherence, glycemic tier, and pressure-injury risk — recalculated the moment the chart holds new data, never on a schedule.

Cited, not claimed

Every score value shows the calculation basis and the citation behind it. The audit reviewer sees the evidence alongside the number.

Care bundles as state machines

Sepsis bundle, timed end to end

The 1-hour, 3-hour and 6-hour bundle lights up the moment the chief complaint is named, tracks every element — lactate, cultures, antibiotics, fluids — and times each step against the protocol.

Pathway machines across units

Stroke, trauma, chest pain and peri-operative bundles run as the same state machines in the ED, OR and ward — adherence visible live, every deviation logged for the quality review.

Smart alarm management

Alarm fatigue is the largest single source of clinician burnout in the ICU. zMed cuts alarm load 60–85% — at three levels.

At the source

Waveform-derived lethal alerts carry self-cooldown; cycle anomalies (a cuff reading the floor of its range three times) prompt a re-cuff instead of three sequential alarms.

At the corroboration layer

Single-observation alerts are downgraded; alerts where two or more modalities agree are escalated. A corroboration score gates severity.

At the override-pattern layer

Every override is captured with a structured reason. A rule that fires often without acceptance reveals itself on the effectiveness dashboard — a noisy rule is a tunable rule, and a tunable rule is a hospital-owned rule.

Numeric analytics — the quiet trends

Trend computation

First and second derivatives of vitals over rolling windows. The slow drifts that never fire an isolated alarm are exactly what feeds the early-warning rules — and the predictive models in Clinical AI.

Cross-modal corroboration

An SpO₂ drop is upgraded when the plethysmograph confirms it and downgraded on obvious motion artefact; a blood-pressure spike is verified against the arterial trace when one is present.

False-alarm scoring

Every alert candidate carries a computed false-alarm likelihood. High-likelihood artefacts are gated — and the suppression itself is logged.

Audit & effectiveness

Every interaction logged

Fire, surface, acknowledge, resolve, override, escalate — the alert lifecycle is consistent and auditable end to end, with clinician identity at every step.

The override is the feedback

Override reasons are the system's primary improvement signal: rules dominated by "not clinically relevant" surface monthly to clinical informatics as retuning candidates. The closed loop runs from chart event to the next quarter's rule revisions.