Total Patients
2,000
Discharge cohort · 2024
30-Day Readmission Rate
22.9%
458 readmissions
↑ 2.9pp above target
Critical Risk Patients
271
Immediate intervention needed
13.6% of cohort
High Risk Patients
357
Priority follow-up required
17.9% of cohort
No Medication Counselling
37.2%
743 patients — discharge gap
Protocol gap
Model ROC-AUC
0.772
5-fold CV: 0.751 ± 0.023
Good discriminability
Risk Tier Distribution
Patient stratification across 4 readmission risk tiers
Critical: 271
High: 357
Moderate: 509
Low: 863
Readmission Rate by Diagnosis (ICD-10)
Percentage readmitted within 30 days by primary diagnosis group
Medication Counselling Impact
Readmission rates split by discharge counselling provision
Key Finding: Patients discharged without medication counselling had a 28.8% readmission rate vs 19.5% with counselling — a 9.3 percentage point gap. At scale, closing this gap could prevent ~69 readmissions annually in this cohort.
Top Risk Predictors (Odds Ratios)
Logistic regression — features ranked by adjusted odds ratio
Disparity Gap (Race)
+9.4pp
Black vs Asian patients
Significant disparity
Uninsured Rate
29.2%
Highest readmission group
vs 18.9% Medicare
Medicaid Readmission Rate
26.9%
622 Medicaid patients
+8pp vs Medicare
Reference Rate (White + Private)
21.5%
Equity benchmark group
Reference
Readmission Rate by Race / Ethnicity
Equity Analysis
Health Equity Alert: Black/African American patients show a 28.6% readmission rate — the highest of any group. This is 1.49× the rate of Asian patients (19.2%) and exceeds the population average by 5.7 percentage points. Intersecting factors include higher SDOH burden and Medicaid/uninsured concentration in this group.
Readmission Rate by Insurance Type
Disparity Matrix — Race × Insurance
Stratified Readmission Rates (%) — Reference: White + Private = 21.5%
Cells with disparity gap > 5pp are flagged. Minimum n=15 per cell.
| Race / Ethnicity | Insurance Type | N | Readmission Rate | Rate Visualised | vs Reference |
|---|
Avg SDOH Burden Score
1.3
Out of 4 factors
Max burden: 53.3% readmit
Food Insecurity Prevalence
35%
700 patients affected
Housing Instability
28%
560 patients affected
Transport Barrier
30%
600 patients affected
Low Health Literacy
40%
800 patients affected
Readmission Rate by SDOH Burden Score
Each point = cumulative SDOH factors present at discharge
SDOH Gradient: Patients with all 4 SDOH factors have a 53.3% readmission rate — 3.2× the rate of patients with no SDOH burden (16.9%). This gradient confirms SDOH as a primary driver of avoidable readmissions.
SDOH Factor Prevalence
Proportion of discharge cohort affected by each factor
🍎 Food Insecurity
35%
🏠 Housing Instability
28%
🚌 Transport Barrier
30%
📖 Low Health Literacy
40%
Compound Risk: 30 patients carry all 4 SDOH burdens simultaneously — this group has a 53.3% readmission rate. Targeted social prescribing and community health worker deployment could disproportionately reduce overall population readmissions.
SDOH Burden × Insurance Type
Logistic Regression — Model Performance Python · scikit-learn
ROC-AUC
0.772
CV AUC (5-fold)
0.751
Sensitivity
71.7%
Specificity
68.8%
Brier Score
0.196
CV Std Dev
±0.023
ROC Curve (Approximated)
True positive rate vs false positive rate at varying thresholds
Feature Importance (Odds Ratios)
Adjusted odds ratios from logistic regression — values > 1.0 increase risk
Methods
Technical approach and reproducibility
Dataset: 2,000 synthetic EHR-style discharge records generated with realistic distributions using NumPy/Pandas. Readmission outcomes are derived from a logistic data-generating process incorporating clinical, demographic, and SDOH covariates — not random noise.
Model: L2-regularised logistic regression (balanced class weights) with 80/20 stratified train/test split. Features standardised via StandardScaler. 5-fold cross-validation used for generalisation assessment.
Equity Analysis: Stratified readmission rates computed by race/ethnicity × insurance type, with disparity gaps expressed relative to White + Private Insurance reference cohort. SDOH burden scored 0–4 per patient.
Python 3.11
scikit-learn
pandas
numpy
ICD-10
Logistic Regression
SDOH Stratification
5-fold CV
Brier Score