Predictive models for hospital readmission are one of the most studied problems in healthcare analytics. The literature is deep. The deployments are common. The clinical impact has been disappointing relative to the investment. I want to talk about why, and about what a model that actually changes outcomes looks like.
The grading problem
Most readmission models are evaluated on AUC or on calibration against the historical readmission rate. These are reasonable statistical metrics. They are not clinical metrics. A model can have an AUC of 0.78 and produce no change in readmissions, because the people the model identifies are people the clinical team already knew were at risk.
The right grading question is different. Of the patients flagged by the model, how many were not already on the discharge planner's high-risk list. Of those, how many got an intervention that was attributable to the model's flag. Of those, how many were not readmitted within thirty days, controlled against a matched cohort.
That is a much harder evaluation. It also produces a much smaller number than the AUC suggests. The smaller number is the real number.
The intervention problem
A model is only as useful as the action it triggers. In most deployments I have seen, the action is a flag in the EHR. The flag goes to the discharge planner. The discharge planner is already overloaded. The flag is dismissed or actioned in a way that is statistically indistinguishable from baseline.
The deployments that move the readmission rate share a property. They have an explicit intervention bundle attached to the flag, and a person who owns the intervention. The flag is not advisory. It triggers a specific workflow. A pharmacist review of the discharge medications. A scheduled call from a care coordinator within forty-eight hours. A booked follow-up appointment within seven days. The model is the routing mechanism for a clinical intervention that has independent evidence behind it.
Without the bundle, the model is a flag in a sea of flags. With the bundle, the model is the most efficient way to allocate a scarce intervention.
The features that matter
Most published readmission models lean heavily on prior utilisation. The features that move the needle in deployment are usually simpler than the literature suggests, and a few of them are not in most published models.
Days since last discharge, with a strong nonlinearity around seven and thirty.
Number of medications at discharge, especially when above ten. Polypharmacy is one of the strongest predictors I have worked with.
Whether the patient has a confirmed primary care appointment within fourteen days of discharge, and whether they have a phone number that has been recently used.
Social determinant flags from the intake assessment. These features are noisy, often incomplete, and meaningfully predictive when they are present.
Notably absent from this list. Most lab values. Most demographic features beyond age. Many of the comorbidity codes that the literature emphasises. The reason is operational. By the time labs are stable enough to be predictive, the discharge decision has been made. By the time comorbidity codes are coded, the patient has been discharged for two weeks.
The deployment shape
A model that gets used has a few operational properties.
It runs at the right point in the workflow, which is usually at the point where the discharge plan is being written, not at admission and not at discharge. The discharge planner needs the flag in time to act on it.
It produces a small list, not a large one. A model that flags twenty per cent of admissions is not actionable. A model that flags the top three per cent on a given day, ranked, is actionable. The cardinality has to match the capacity of the intervention.
It is recalibrated quarterly against the local population. The model that worked in the academic medical centre will not be calibrated for the community hospital without local recalibration. The model that was calibrated on 2022 data will drift by 2025.
It is monitored on the clinical metric, not the statistical one. The dashboard reads readmission rate among flagged patients, intervention completion rate, and time to first contact post-discharge. Not AUC.
What to build first
If I were starting a readmission programme today, I would not start with the model. I would start with the intervention bundle. I would identify the highest-evidence interventions the institution can credibly deliver. I would build the operational machinery around those interventions, including the staff who will own them and the workflows that will trigger them. I would document a baseline.
Then I would build the simplest possible model, deploy it as the routing mechanism, and measure the lift against baseline. The model is the cheap part. The intervention is the hard part. The deployments that fail get this ordering backwards.