Readmission risk scoring

At the point of discharge, a model can estimate 30-day readmission probability from the chart that exists right then. The quality output is a score that cites the factors that moved it and the vintage of the history those factors rest on. If the history is too thin, the field stays empty. A care-transitions nurse still decides whether to run a high-risk care transition protocol.

That split matters. The model answers how this stay compares to other stays it has seen, given this snapshot. It does not answer whether to start the next-day call, the home visit, or monitoring. Treating the score as an automatic protocol is a failure mode. You will over-enroll patients whose charts look risky because they are incomplete, and you will skip patients whose risk is real but sits just under a threshold someone set in a vendor console.

Epic, Oracle Health, Microsoft, and Biofourmis sit in the class of tools that can attach a prediction to discharge or to a post-acute workflow. Do not rank them here. None of them can manufacture a missing history, and none of them should fire a protocol without a named clinician decision. If a product only shows a color or a rank with no factor list and no vintage, do not treat that display as a quality score.

Do not invent a readmission percent to make the huddle slide look complete. If the model did not emit a probability, or if you blanked the field because history was thin, say that. A made-up 30-day rate is not a clinical communication. It is a documentation problem.

Score the stay against a locked discharge snapshot, not against a chart that is still moving. Discharge is a busy hour: med rec is finishing, the after-visit summary is being edited, a last lab may still be pending, and a social-work note may land after the patient has left the unit. If the model reads whatever is current at query time, two nurses can get two scores for the same stay, and neither can explain which version of the chart produced the number.

Lock means freeze the data the model is allowed to see at a defined event, typically the discharge order or the discharge timestamp, and store that freeze with the score. Include the problem list, utilization in the lookback the model was trained on, medications at discharge, labs with their draw times, and any coded social determinants you actually have. Exclude notes that arrive after the lock. If a late addendum changes the clinical story, rescore from a new snapshot and keep both versions. Do not silently overwrite.

A snapshot without a timestamp is the same problem as a score without vintage. Put the lock time next to the score.

The snapshot is also where you notice gaps other discharge work has to handle. Missing living situation and caregiver status belong in social support gap detection, not in a guessed risk percent. A locked empty field is more honest than a filled one.

A usable score names the factors that contributed and how old the evidence is for each. Naming heart failure, prior admissions, and polypharmacy is a start only if you can also say which encounter those admissions came from and when the med list was last reconciled. Vintage is how a nurse decides whether the model is reading this stay or a stale registry.

Write factor cites in language a covering clinician can check in two minutes. Prefer a diagnosis present on the locked problem list; a count of inpatient stays in the model's lookback, with the date of the most recent one; a discharge med class with the reconciliation timestamp; a lab value with the draw time. Avoid opaque feature names, hashed variable IDs, and a "top contributors" widget that disappears when you print the after-visit summary.

If a factor has no dated source, drop it from the cite list or mark it unknown. A cite without vintage is how teams start believing a score they cannot audit. The same rule applies when a vendor dashboard shows risk over time with no as-of dates. Do not copy that display into the transition note.

Chronic disease course can inform the score when the history is dated and continuous enough to use. When you need the trajectory itself, not just a discharge probability, that work lives with chronic disease trajectory modeling. Do not collapse a multi-year course into a single unlabeled risk chip on the discharge banner.

Empty stays empty. Thin history is not a reason to emit a midpoint probability, a moderate label, or a hospital-wide average. Those substitutes are invented rates. They teach the team that every row in the discharge list has a number, which is convenient for a dashboard and unsafe for a protocol.

History is thin when the lookback the model needs is missing, when the patient is new to the system, when prior stays are at unaffiliated hospitals you cannot see, or when key factors exist only as unscored free text. You do not need a perfect chart. You need enough dated, checkable evidence that a clinician can defend the score. If you cannot, leave the score blank and say why: new to the enterprise, no utilization in lookback, last echocardiogram outside vintage, social history not recorded.

Here is the pattern in practice. A Thursday discharge: a patient with heart failure, two prior inpatient stays visible in the lookback, a med list reconciled this morning, and an echocardiogram from fourteen months ago. The model may cite the diagnosis, the prior stays with their dates, and the current med list. It should not treat the old echo as current systolic function. If the model requires a recent ejection fraction and you do not have one, that factor stays blank. The nurse does not plug in a typical heart-failure readmission rate to complete the note. The nurse records the partial cites, the missing vintage, and then decides whether the remaining evidence is enough to start a high-risk pathway or whether the next step is to obtain the missing history and rescore.

Blank tells the team this stay needs review, not that the patient is low risk.

The protocol is a clinical action: extra teach-back, a scheduled call, a home health referral, pharmacy follow-up, or enrollment in monitoring. The score is an input to that action. Automatic enrollment from a threshold is the second failure mode. Thresholds do not know that this patient has a daughter in the house, that the prior admissions were planned chemotherapy, or that the chart is thin and the score should have been blank.

Keep a named decision on the record: run high-risk transitions, run standard discharge, or hold for missing history. The decision cites the score when the score exists, cites the blanks when it does not, and names the person who decided. If the EHR can auto-order a pathway, require that order to pass a clinician sign-off. Configure tools from Epic, Oracle Health, Microsoft, or Biofourmis the same way: prediction first, protocol second.

When you do run the protocol, wire it to work that can actually change the next thirty days. A high-risk flag with no follow-up plan is theater. If the concern is whether the patient will take the new diuretic, that is care plan adherence monitoring. If the concern is decompensation at home, that is remote patient monitoring with escalation, with a human still behind the alert. The score does not choose those tools. The clinician does, using the factor cites as the reason.

Put the same locked snapshot time, probability or blank, factor cites with vintage, and protocol decision in the discharge summary, the huddle list, and the handoff. If you cannot explain the score in one spoken sentence, it is not ready for a protocol. A working sentence: score based on heart failure and two dated prior stays; echo vintage too old to cite; living situation unknown; next-day call yes, remote monitoring held until social support is documented. That sentence has no invented 30-day percent.