Early warning score automation

A usable alert names the vital or lab, the clock time of that value, and the protocol threshold it crossed. Anything less is a colored tile, not a reason to leave the desk.

You are looking at a board of inpatients, not a diagnosis engine. The job of early warning score automation is to surface a calibrated deterioration signal for sepsis, decompensation, or clinical decline, then stop talking. Empty stays empty when the stream is ordinary. A rapid-response nurse still walks to the room, still examines the patient, still decides whether the protocol applies.

If intake already tagged the patient as high acuity, that context belongs in the chart, not as a second unexplained red badge. Pair this work with acuity risk stratification at intake so the overnight board is not restating a triage score you already knew.

The same discipline applies to labs. A lactate or creatinine spike is only useful when the alert shows the result, the draw time, and the cutoff your sepsis or AKI bundle actually uses. That is the same citation pattern as lab result anomaly detection, applied to a continuous inpatient stream instead of a one-off result.

Vendors sit in one class here: Epic, Oracle Health, Microsoft, and Philips-class monitors. Do not rank them. Do not assume a named model, including the Epic Sepsis Model, has already done your assessment. The model is a trigger. You are the response.

Start by loading the live stream the unit already trusts: heart rate, respiratory rate, blood pressure, temperature, oxygen saturation, and the labs your deterioration protocol names. Then load the protocol itself, as written for this hospital, this unit, this shift. Do not type a number you remember from another facility.

Map each protocol rule to a field and a comparator. If the bundle fires when respiratory rate is at or above a stated threshold for a stated duration, the automation must read that same respiratory rate, at that timestamp, against that threshold. If the protocol is silent on a value, the score must not invent one.

Calibrate the alert text to the rule, not to a vendor default label. "Sepsis watch" without a cited source is theater. "Respiratory rate 28 at 02:14, protocol threshold 24" is a reason to stand up.

Where the stream comes from does not change the mapping. A Philips-class monitor at the bedside, or a flowsheet in Epic, Oracle Health, or Microsoft, can all feed the same fields. Treat them as pipes. The protocol is the spec.

If you cannot bind a protocol line to a timestamped reading, do not light the tile. A red tile with no reading is a failure mode, not a conservative default. It trains the unit to either ignore color or to chase ghosts. Both waste a rapid-response nurse.

Watch for silent substitution. Some implementations fill a missing respiratory rate from an older value, or from a neighboring vital, so the score never looks empty. That is inventing a threshold by another name. Missing is missing. The board should show a gap, not a borrowed number dressed as a fresh score.

Most patients on a typical med-surg night will not meet a deterioration rule. The correct output for those rows is blank. No amber. No "watch." No score that looks urgent because a dashboard designer wanted a number in every cell.

Silence is a quality feature. If the board is always busy, you will stop believing it, and the first real decompensation of the night will look like every other tile. Leave ordinary empty so a cited alert is rare enough to act on.

Do not compensate for a noisy model by adding a second, unofficial cutoff in your head. Inventing a threshold ("we only go if it is really high") means the written protocol and the practiced protocol have diverged. Either change the protocol through governance, or honor the one on the wall.

Remote programs use a similar empty-unless-escalated pattern. Inpatient early warning is not remote monitoring, but the escalation logic is cousin to remote patient monitoring with escalation: ordinary traces stay quiet; only a cited breach pages a human.

Adverse events that are not vital-sign deterioration belong on a different channel. Do not overload this board with pharmacy or device signals. Route those through adverse event signal detection so a sepsis-style tile still means physiology plus protocol.

It is 02:14 on a med-surg unit. The board has been empty for an hour. One row lights: respiratory rate 28 at 02:14, protocol threshold 24; temperature 38.1 at 01:58, protocol threshold 38.0. Heart rate and blood pressure are listed and have not crossed. No sepsis rate is attached. No probability of death. The text names two readings, two times, two cutoffs.

You do not treat that tile as a diagnosis of sepsis. You treat it as a reason to go to the room. At the bedside you find a patient who is working to breathe, with a recent antibiotic change and a new oxygen requirement that was not yet in the flowsheet. You follow the unit protocol: assess, call the covering clinician, start the bundle steps that the protocol assigns to nursing. The model did not decide the patient had sepsis. You and the clinician did the assessment.

There is no claimed catch rate and no minutes-saved figure. The quality bar is the cite on the tile and the human response that followed.

If the same night the board had shown a red tile with no respiratory rate, no timestamp, and a banner that only said "high risk," you would have had to open the chart to reconstruct the story before you even stood up. That delay is the product failing, not the nurse moving too slowly.

A red tile with no reading is the first failure. Color without a vital, a lab, a time, and a threshold is not an early warning score. It is a notification. Reject it in design review. If it ships, teach the unit to treat it as incomplete until the cite appears, and log it as a defect, not as a clinical miss.

Treating the alert as a diagnosis is the second failure. Sepsis, decompensation, and clinical decline are judgments made after you look at the patient, the trend, and the differential. A model, including the Epic Sepsis Model or any counterpart on Oracle Health, Microsoft, or Philips-class monitors, can be wrong in both directions. A cited alert is a page, not a label you copy into the problem list.

Inventing a threshold is the third failure. It shows up as a shadow rule on the unit ("we ignore the first fire"), as a silently imputed vital, or as a vendor default that nobody compared to the hospital protocol. If the written threshold and the firing threshold differ, you do not have automation. You have two protocols. Fix the mapping or change the document. Do not split the difference at 02:14.

Secondary burns follow from those three. Alert text that names a disease instead of a reading. Scores that never go blank, so ordinary nights look like crisis. Duplicate pages because intake acuity, lab anomaly, and early warning all fire on the same creatinine without saying they share a source. Keep each channel cited and scoped.

The clinician still responds. Nursing assessment, covering-provider judgment, and the protocol's next steps are not optional because a tile lit. Document what you found, what you did, and whether the cited values were still true when you arrived. If the respiratory rate had already fallen, say so. The alert was true at 02:14. The patient at 02:22 is a new observation.

Feed the outcome back only through the governance path your hospital uses: true deterioration, explained abnormality, or artifact. Do not retune thresholds on the unit. Do not disable a row because last night was noisy. Calibration belongs to the people who own the protocol, with the same citation rule: change the written cutoff, then change the map.

When the stream is ordinary again, the row goes empty. That is the system working. Your attention returns to the rooms that still have a cited reason to be there.