- Design
- Retrospective cohort with machine-learning prediction models
- Population
- 493,035 hospital encounters in 21 Kaiser Permanente Northern California hospitals
- Primary outcome
- Prediction of discharge disposition within the next 24 hours
- Effect
- Day 3 skilled nursing facility discharge: AUC 0.89 with PF5 vs 0.75 with clinician estimate
Kaiser Permanente Northern California built the five-item PROMIS physical function tool (PF5) into its record across 21 hospitals. Patients reported their prior function, and clinicians recorded current function daily. This retrospective study of 493,035 encounters tested how well models with PF5 predicted discharge to home, home with services, or a skilled nursing facility in the next 24 hours.
Models using PF5 outperformed clinician-entered expected discharge disposition, particularly early in the stay, when those entries were often missing. For nursing facility discharge on day three, the area under the curve was 0.89 against 0.75. Clinician-assessed current function was the strongest predictor, and the number of patients needing evaluation to find one nursing facility discharge fell from 4.0 on day one to 1.4 on day seven.
The point is not the algorithm. A short, structured function score recorded every day carries most of the signal, and any ward can start recording it. Whether it shortens stays has not yet been tested.
- Record current physical function in a structured way every day, not just in therapy notes.
- Use the gap between prior and current function to start post-acute referrals early.
- Clinician-assessed current function was the strongest single predictor of discharge destination.
- Fill in expected discharge destination early; it was often missing in the first days.
- Whether this approach shortens length of stay is still untested.
Why it matters
Discharge planning usually starts late because nobody records function early, and this shows the information is both cheap and predictive.
Don't overread it
This shows better prediction, not better outcomes — no one has yet tested whether acting on it shortens stays.
The statistics, in plain English
An area under the curve of 0.89 means that, given one patient who went to a nursing facility and one who did not, the model ranked them correctly 89% of the time; 0.75 is fair, 0.89 is good. A number needed to evaluate of 1.4 means about three in four flagged patients really did go to a facility.
Read the rest in the app
You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.
Scan to keep reading on your phone. No account needed to start.
Tomorrow morning, before your first patient
One edition a day for internal medicine, written by the desk, every claim tied to its paper. Six minutes.
Get the app — free