Patient Readmission Risk

Problem

Predict which patients are at elevated risk of being readmitted to hospital within 30 days of discharge, enabling targeted interventions (community support, pharmacy follow-up, GP communication) that prevent avoidable readmissions. In the NHS, 30-day readmission rates are ~15% and are used as a quality indicator. In the US, CMS penalises hospitals for excess readmissions under the HRRP (Hospital Readmissions Reduction Program).

Users / Stakeholders

Role	Decision
Discharge coordinator	Identify high-risk patients needing enhanced discharge planning
Community care team	Prioritise post-discharge calls and home visits
Hospital administrator	Track readmission KPIs; CMS penalty avoidance
GP / primary care	Follow-up appointment scheduling
Patient / carer	Understand discharge instructions; support needs

Domain Context

Timing sensitivity: The intervention window is at discharge. Score must be available before the patient leaves the ward. Real-time score at discharge event required.
Social determinants: Readmission is strongly predicted by social factors (social isolation, housing instability, health literacy) that are often not in the EHR. Missing features limit model accuracy.
Condition specificity: Readmission predictors differ by diagnosis. A single global model underperforms condition-specific models for high-prevalence conditions (heart failure, COPD, pneumonia).
Intervention effectiveness: If no intervention is available or the patient declines support, the score is not actionable. Model should only flag patients where intervention is feasible.
GDPR / DSPT: NHS Data Security and Protection Toolkit. Patient data must not leave NHS systems without IG approval.

Inputs and Outputs

Features:

Clinical: primary_diagnosis (ICD code), Charlson comorbidity index, length_of_stay,
          procedure_codes, lab_values_at_discharge, medication_changes
Prior utilisation: n_admissions_12m, n_ED_visits_6m, n_GP_contacts_3m
Demographics: age, sex, IMD_deprivation_decile
Social: lives_alone, carer_present, care_package
Discharge: discharge_destination (home/care home/other), time_of_day_discharged

Output:

readmission_score:    P(readmission within 30 days) ∈ [0, 1]
risk_tier:            LOW / MEDIUM / HIGH
intervention_flag:    Recommended intervention type
top_risk_factors:     Clinician-interpretable explanation (SHAP top 3)

Decision or Workflow Role

Discharge planning meeting (24–48h before discharge)
  ↓
EHR triggers readmission score computation
  ↓
HIGH risk → Enhanced discharge: community referral, pharmacy review, 
             GP alert, 48h follow-up call scheduled
MEDIUM → Standard: GP letter, patient discharge summary
LOW → Routine discharge
  ↓
30-day outcome tracked → confirmed readmission or not
  ↓
Monthly model performance review → recalibrate if AUC drifts

Modeling / System Options

Approach	Strength	Weakness
LACE+ Index (rule-based)	Validated; regulatory accepted; simple	Limited accuracy; no personalisation
Logistic regression	Interpretable; fast; calibrated	Misses complex comorbidity interactions
LightGBM	Higher AUC; handles ICD code sparsity	Less interpretable; more validation effort
Condition-specific ensemble	Best performance; handles heterogeneity	Complex to maintain (one model per condition)

Recommended: LACE+ as baseline. LightGBM for high-volume conditions. SHAP explanations mandatory.

Deployment Constraints

Real-time scoring: Score must be available at discharge. EHR trigger-based computation.
NHS IG: Data must stay within NHS environment. On-premises deployment only.
Equity: Score must not disadvantage protected groups. Deprivation is a predictor — monitoring required to ensure interventions are equitably distributed.

Risks and Failure Modes

Risk	Description	Mitigation
Demographic bias	Model higher FPR for ethnic minorities	Subgroup AUC monitoring; representative training data
Label definition	Readmission vs planned readmission — planned should be excluded	Careful label curation; exclude planned admissions
Social factor blindspot	High-risk social factors not in EHR → missed predictions	Structured social data collection at admission
Intervention capacity	High-risk list longer than team can handle	Tiered intervention; capacity-constrained prioritisation

Success Metrics

Metric	Target	Notes
AUROC	> 0.75	vs LACE+ benchmark ~0.65
30-day readmission rate reduction	> 10% relative	Intervention group vs control
High-risk recall	> 80%	Fraction of readmitted patients flagged
Alert actionability	> 60% result in documented intervention	Not just flagged; acted upon
Equity: AUC gap	< 0.03 across demographic groups	Fairness metric

References

van Walraven, C. et al. (2010). Derivation and Validation of an Index to Predict Early Death or Unplanned Readmission.
Rajkomar, A. et al. (2018). Scalable and accurate deep learning with electronic health records. npj Digital Medicine.

Notes

Explorer

patient_readmission