The problem
After discharge, a hospital’s transition-care nurses phone patients to check medicines, symptoms and follow-up appointments. There are only so many calls a shift can make, so most services pick patients with a LACE-style score or by ward. Published programmes that pair risk screening with nurse calls show lower readmissions in the patients they reach, while calls made without targeting show little effect, which makes the choice of who to call the part that matters.
I built a nightly list that ranks discharges by readmission risk and measured how many 30-day readmissions it reaches at the same number of calls as a LACE-style score.
- Team
- Discharge or transitional-care nurses, ward managers, patient flow
- Systems
- Cerner or Epic discharge records, the patient administration system, a call worklist in the nursing system
- Measure
- Share of 30-day readmissions whose patient was on the call list
What I built
A nightly risk list. Each discharge from the day is scored from details already in the discharge record: where the patient went, prior inpatient and emergency visits, main diagnosis, length of stay, medicines and age. The list is sorted highest risk first and cut at the number of calls the team can make.
Results
On 14,811 stays of 12,469 patients the model never saw, calling the highest-risk 20% of discharges reaches 34.7% of everyone readmitted within 30 days. The LACE-style score reaches 27.2% with the same number of calls, so the same nurse hours reach 28% more of the patients who come back.
| Share of discharges called | Risk model | LACE-style score | Prior stays | Readmitted among those called (model) |
|---|---|---|---|---|
| 10% | 22.1% | 15.8% | 16.3% | 19.5% |
| 20% | 34.7% | 27.2% | 29.4% | 15.3% |
| 30% | 46.4% | 38.5% | 39.5% | 13.6% |
This is a modest signal, and I report it as one. Readmission depends on much that a discharge record does not hold: who is at home, whether the scripts were filled, how the patient feels on day three. The gain comes from ranking a fixed list better, not from predicting individuals with confidence.
Calibration
What drives the ranking
| Main diagnosis group | Test stays | 30-day readmission |
|---|---|---|
| Circulatory | 4,339 | 9.6% |
| Other | 2,696 | 8.8% |
| Respiratory | 1,975 | 7.6% |
| Digestive | 1,375 | 7.9% |
| Diabetes | 1,124 | 9.5% |
| Injury | 1,083 | 10.2% |
| Genitourinary | 867 | 8.2% |
| Musculoskeletal | 821 | 6.5% |
| Neoplasms | 529 | 9.6% |
Data and method
- Data
- Diabetes 130-US Hospitals for years 1999–2008 (Strack et al., 2014; UCI Machine Learning Repository, CC BY 4.0): 101,766 inpatient stays of patients with diabetes. Stays ending in death or hospice discharge are excluded, leaving 99,343
- Outcome
- Readmitted within 30 days of discharge
- Split
- By patient. Patients first seen in the earliest 80% of stays (57,521 patients, 84,532 stays) train the model; patients first seen later (12,469, 14,811 stays) are the test set, so no patient is in both
- Model
- Gradient-boosted trees (scikit-learn HistGradientBoosting) on discharge destination, admission type and source, diagnosis group, specialty, length of stay, lab and medicine counts, prior visits, HbA1c and glucose results, insulin and age
- Baselines
- A LACE-style score built from the fields available (length of stay, emergency admission, a diagnosis-count proxy for comorbidity, emergency visits) and prior inpatient stays alone
- Uncertainty
- Bootstrap 95% range on AUC over test stays
- Brier score
- 0.0787 against 0.0814 for a constant rate
Limitations
- The data comes from United States hospitals between 1999 and 2008 and covers patients with diabetes only. An Australian health service would train and test on its own discharges before any list goes live.
- The LACE-style baseline lacks a true Charlson comorbidity index, so it understates a full LACE score somewhat.
- Readmissions to another hospital are not visible in this data, as in most single-service records.
- A call list sorted by risk changes who gets called, so outcomes of called patients are tracked separately from the model’s accuracy.
How I would take it live
- Pull two years of discharges and readmissions from the patient administration system
- Retrain, recalibrate and agree the daily call capacity with the transitional-care team
- Run the list in shadow beside the current rule for a month and compare who each would have called
- Switch over, then report weekly on readmissions among called and uncalled patients
Technical depth
Patients are split by their first stay so that repeat stays of one patient never straddle training and test. Categorical inputs are passed to the trees as native categories. Importances are permutation drops in AUC on held-out stays. No patient-level details are shown on this page.