Which discharged diabetic patients should a care-transitions team call first? A readmission-risk model on public data from 130 US hospitals, shown in terms a quality-improvement team can act on.
Aryan Meena · M.S. Applied Data Analytics, Boston University · Code & methods · LinkedIn
Outreach capacity planner
Most teams can only call a fraction of discharged patients. Choose that fraction and compare who you reach when you call the highest-risk patients first versus calling at random.
10% of discharges · calls
readmissions reached, calling highest-risk first
readmissions reached, calling at random
Highest-risk first (model)Random calls
Who is most at risk
Patients ranked into ten equal groups by predicted risk. Bars show how many in each group were actually readmitted within 30 days; dots show what the model predicted. Bars and dots line up, so a predicted 20% risk means about 20%.
Actual 30-day readmission ratePredicted rateOverall rate
Show as table
Equity check
A risk tool should rank patients fairly across groups. Ranking accuracy (AUC, where 0.5 is a coin flip) by age, race and sex, with each group's actual readmission rate.
What this means for a quality-improvement team
Methods and limits
Data: UCI "Diabetes 130-US Hospitals" (1999–2008), public and de-identified. Encounters ending in death or hospice were excluded, since those patients cannot be readmitted.
Model: LightGBM with patient-history features, tuned inside nested cross-validation and grouped by patient, so no patient appears in both training and testing. Every number here comes from predictions on held-out patients, then isotonic calibration.
Benchmark: nested cross-validated AUC of 0.689, above a published 0.664 on the same dataset; full comparison in the repository.
Privacy: this page shows only aggregated results; no patient-level rows are published.
Limits: the data are 15+ years old and from one dataset; a tool like this needs local validation before any clinical use. It supports prioritising outreach, not individual clinical decisions.