Hospital Readmission Risk Model — Public Hospital Network, Victoria, Australia

The Situation

A public hospital network operating 3 hospitals in regional Victoria had a 30-day readmission rate sitting between 16% and 19% across its general medicine, cardiology, and orthopaedic wards. Under the Australian National Efficient Price framework, readmissions above benchmark thresholds triggered funding penalties. The network had been absorbing those penalties for 3 consecutive years.

The clinical team was not ignoring the problem. Post-discharge follow-up calls were happening. Community health referrals were being made. The issue was prioritisation. Care coordinators were working from discharge lists sorted by date, not by risk. High-risk patients were getting the same follow-up cadence as low-risk ones, sometimes less if they had been discharged on a busy day.

Five years of patient records existed across the network's clinical information system, covering diagnosis codes, comorbidity profiles, length of stay, medication history, and prior admission frequency. Nobody had used them to build a risk model.

What Amorisoft Did

Amorisoft ran a 4-week data extraction and preparation process across the network's clinical information system. Records from 5 years covering 89,000 patient episodes were cleaned and structured at the episode level. Readmission outcomes within 30 days were labelled for each episode and used as the training target.

Feature selection drew on 31 clinical and administrative variables. The 4 strongest predictors were prior admission frequency in the preceding 12 months, number of active comorbidities at discharge, length of stay relative to diagnosis group average, and whether a community health referral had been completed before discharge. Age ranked seventh, lower than the clinical team had expected.

The model was trained separately for each of the 3 ward types because readmission drivers differed enough between cardiology and orthopaedic patients to justify separate training runs. A combined model produced 71% accuracy on the held-out evaluation set. Ward-specific models brought accuracy to 79%.

At discharge, the model scores each patient and outputs a risk tier: high, moderate, or low. Care coordinators receive a prioritised follow-up list each morning, sorted by risk tier and then by days since discharge within each tier. The list is generated automatically from the clinical information system overnight.

The model does not make clinical decisions. It scores and ranks. Every follow-up action is taken by a care coordinator or clinician.

Results

30-day readmission rates across all 3 wards dropped by 28% in the 6 months following full deployment. High-risk predictions were confirmed accurate on clinical review in 79% of cases. Care coordinator follow-up efficiency improved by 41%, measured as the proportion of high-risk patients contacted within 48 hours of discharge, which had previously sat at 34% and reached 75% within 3 months of go-live. The network avoided funding penalties in the annual NEP review for the first time in 4 years.