Before a patient leaves hospital, a care-management team has to decide who gets a follow-up call, a home visit or a medication review. The team can reach only a fraction of discharged patients. The useful question is therefore not "who is at risk?" but "who should we contact first, and why?"
The decision and the constraint
The score had to be produced before discharge, using only information available at that point, and the output had to fit the care team's weekly capacity. That ruled out several attractive but unusable predictors: discharge destination, final billing codes and anything recorded after the patient left.
What we worked with
We used encounter-level hospital records for patients with diabetes, covering:
- Patient context and admission source, including age, gender and prior utilisation.
- The hospital stay itself: length of stay, procedures, laboratory tests, medications and coded diagnoses.
- Diabetes control indicators, including HbA1c and glucose results, and changes to diabetes therapy.
- Discharge information, used to define the cohort and the outcome but not as a predictor.
The traps we designed around
Readmission data is unusually easy to model incorrectly. Three issues mattered more than the choice of algorithm:
- Patients, not encounters. A patient can appear several times in the data. A random row split would place the same patient in training and validation, producing a model that looks better than it is. We split by patient.
- Discharge-disposition leakage. Some disposition codes describe death or hospice. Those encounters cannot be readmitted in the ordinary sense and were excluded from the target. Keeping them would have taught the model to associate end-of-life discharge codes with "no readmission".
- Definition drift. "Readmitted" can mean within 30 days, within a year, or to any hospital. We fixed the window and the cohort before modelling, then kept them fixed through validation.
How the score was built
- Feature selection with clinical input. Prior inpatient and emergency visits, medication burden, number of diagnoses, HbA1c control and therapy changes were reviewed with the care team. Features that could not be acted on were treated as context, not as targets for intervention.
- Transparent baseline, tested challenger. A logistic model provided an interpretable starting point; a gradient-boosted model was tested on discrimination and calibration. The challenger had to improve both without becoming unexplainable.
- Calibration before ranking. A ranked list is useful, but care managers also need to know whether "high risk" means one in five or one in two. We reported calibration by risk band and checked it on held-out patients.
- Capacity-based tiers. The cut-off came from the number of patients the team could follow up, not from a default probability. Each tier mapped to a different action: routine call, medication review or urgent clinical review.
- Subgroup checks. We examined performance across age, gender and race groups where sample size allowed, and documented where the evidence was too thin to support a conclusion.
What this changes
The care team gained a consistent first pass over every discharge, with a reason for each flag. Three practical shifts mattered:
- Prior utilisation moved to the centre of the conversation. Recent emergency and inpatient visits carried more signal than any single laboratory value, which changed which patients the team looked for.
- Risk was discussed in bands, not adjectives. "High risk" became a calibrated range tied to a follow-up action, which made capacity planning easier.
- The model stayed decision support. Care managers could override it, and overrides were recorded. The score was never presented as a clinical verdict.
What we delivered
- A pre-discharge risk score with patient-level reason codes.
- A daily follow-up list split into action tiers.
- A calibration report and a validation note covering cohort definition, leakage checks and subgroup limitations.
- A monitoring plan for score drift and changes in discharge practice.
Where this applies
The method fits hospitals, health systems and insurers working with encounter-level claims or EHR data. The minimum inputs are a stable patient identifier, admission and discharge dates, prior utilisation, diagnosis and medication fields, and a readmission definition agreed with the clinical owner. The first deliverable is usually the cohort definition and the leakage review, not the model.
This is an illustrative case study using anonymized data with no patient-identifiable information; it does not report results from a specific client engagement. It reflects our standard method for clinical risk modelling.
