Hospital Readmission Prediction

What is Readmission Prediction?

Readmission prediction estimates the probability that a patient will be re-hospitalized within 30 days of discharge, enabling targeted interventions to improve outcomes and reduce costs. Hospital readmissions affect 15-20% of Medicare patients, costing the US healthcare system 17 billion considered potentially preventable. The CMS Hospital Readmissions Reduction Program (HRRP) penalizes hospitals up to 3% of base DRG payments for excess readmissions, creating both clinical and financial incentives for accurate prediction.

The clinical value of readmission prediction extends beyond financial penalties. Early identification of high-risk patients enables targeted interventions: transitional care programs reduce readmissions by 20-25%, medication reconciliation reduces drug-related readmissions by 30%, and post-discharge follow-up calls reduce 30-day readmissions by 15%. These interventions require knowing which patients to target, as applying them universally would be cost-prohibitive. AI models achieve AUC of 0.72-0.78 for 30-day readmission prediction, compared to 0.62 for the LACE index (current clinical standard), representing a substantial improvement in risk stratification accuracy.

Traditional risk scores (LACE, HOSPITAL score) use hand-crafted features and logistic regression, achieving limited discrimination. The LACE index combines Length of stay, Acuity of admission, Comorbidities, and ED visits, achieving AUROC of 0.62—only slightly better than random (0.50). Machine learning models can capture non-linear interactions between hundreds of features, achieving 10-15% higher AUROC while maintaining interpretability through feature importance analysis.

Temporal modeling adds significant value by capturing patient trajectory across multiple visits. A Transformer-based model processing the last 10 visits achieves AUROC of 0.76, compared to 0.72 for models using only the index admission. The improvement comes from capturing patterns like: repeated admissions for heart failure decompensation, increasing medication burden over time, and declining functional status. These temporal patterns are invisible to models that only consider the current admission.

Key Risk Factors

• Prior utilization: Number of admissions in past 12 months (strongest predictor)
• Comorbidity burden: Elixhauser/Charlson comorbidity indices (31 categories)
• Discharge disposition: Home vs skilled nursing facility (OR = 1.8 for SNF)
• Medication complexity: Polypharmacy and discharge medication count (OR = 1.3 per 5 medications)

Evaluation Metrics

AUROC (Area Under ROC)

Where each parameter means:

• — True Positive Rate (sensitivity) at threshold : proportion of actual positives correctly identified
• — False Positive Rate at threshold : proportion of actual negatives incorrectly identified as positive
• — inverse function mapping FPR to corresponding threshold
• Intuition: AUROC measures the probability that a randomly chosen positive example is ranked higher than a randomly chosen negative example. Range [0.5, 1.0], where 0.5 = random, 0.7-0.8 = acceptable, 0.8-0.9 = excellent, >0.9 = outstanding. For readmission prediction, AUROC 0.72-0.78 is typical, reflecting the inherent difficulty of predicting future events from historical data.

AUPRC (Area Under Precision-Recall)

Where each parameter means:

• — proportion of positive predictions that are correct
• — proportion of actual positives correctly identified
• Intuition: AUPRC is more informative than AUROC for imbalanced datasets (15-20% readmission rate). A model that predicts "no readmission" for everyone achieves AUPRC = 0.15 (prevalence), while AUROC = 0.50. Thus, AUPRC of 0.40 represents meaningful improvement, while AUROC of 0.60 may not. For readmission prediction, AUPRC ranges 0.35-0.50, with 0.45 considered good.

Expected Calibration Error (ECE)

Where each parameter means:

• — number of calibration bins (typically 10-20)
• — set of predictions in bin (predicted probability between and )
• — number of predictions in bin
• — total number of predictions
• — actual accuracy (fraction of correct predictions) in bin
• — average predicted confidence in bin
• Intuition: Calibration measures whether predicted probabilities match observed frequencies. For clinical deployment, a model predicting 20% readmission risk should have ~20% of such patients actually readmitted. ECE < 0.05 is well-calibrated, 0.05-0.10 is acceptable, >0.10 requires recalibration. Poor calibration leads to incorrect risk stratification and inappropriate interventions.

Implementation

import torch
import torch.nn as nn
import numpy as np

class ReadmissionPredictor(nn.Module):
    def __init__(self, n_features=50, d_model=64, nhead=4):
        super().__init__()
        self.feature_proj = nn.Linear(n_features, d_model)
        self.visit_attn = nn.MultiheadAttention(d_model, nhead, batch_first=True)
        encoder_layer = nn.TransformerEncoderLayer(d_model, nhead, batch_first=True)
        self.transformer = nn.TransformerEncoder(encoder_layer, num_layers=2)
        self.readmission_head = nn.Sequential(
            nn.Linear(d_model, 32),
            nn.ReLU(),
            nn.Dropout(0.3),
            nn.Linear(32, 1),
            nn.Sigmoid()
        )
    
    def forward(self, visits):
        x = self.feature_proj(visits)
        h, _ = self.visit_attn(x, x, x)
        h = self.transformer(h)
        risk = self.readmission_head(h[:, -1])
        return risk

class LACEScore:
    """LACE Index: clinical baseline for readmission prediction."""
    
    def __init__(self):
        self.feature_names = ['Length of stay', 'Acuity', 'Comorbidities', 'ED visits']
    
    def compute(self, los, acuity, comorbidities, ed_visits):
        """Compute LACE score (0-19 scale)."""
        # Length of stay (days)
        if los <= 1: lace_los = 0
        elif los <= 3: lace_los = 1
        elif los <= 6: lace_los = 2
        elif los <= 13: lace_los = 3
        else: lace_los = 4
        
        # Acuity of admission
        lace_acuity = 0 if acuity == 'elective' else 3
        
        # Comorbidities (Elixhauser count)
        if comorbidities == 0: lace_com = 0
        elif comorbidities <= 3: lace_com = 1
        else: lace_com = 2
        
        # ED visits in past 6 months
        if ed_visits == 0: lace_ed = 0
        elif ed_visits == 1: lace_ed = 1
        elif ed_visits == 2: lace_ed = 2
        else: lace_ed = 3
        
        return lace_los + lace_acuity + lace_com + lace_ed

model = ReadmissionPredictor()
visits = torch.randn(16, 10, 50)  # 16 patients, 10 visits, 50 features
risk = model(visits)
print(f'Risk scores: {risk.shape}, mean={risk.mean():.3f}')
# Risk scores: torch.Size([16, 1]), mean=0.487

# LACE baseline comparison
lace = LACEScore()
score = lace.compute(los=5, acuity='emergency', comorbidities=4, ed_visits=2)
print(f'LACE score: {score}/19')  # LACE score: 9/19

Feature Engineering

Real-World Case Study

The Hospital Readmission Reduction Program (HRRP) analyzed readmission patterns across 3,000+ US hospitals, finding that AI-based risk stratification could reduce excess readmissions by 15-20%. A gradient boosting model trained on Medicare claims data (n = 3.5 million discharges) achieved AUROC of 0.72 for 30-day all-cause readmission, with the strongest predictors being: prior admissions in 6 months (OR = 2.8), discharge to SNF (OR = 2.1), and heart failure diagnosis (OR = 1.9). The model's top 20% risk group contained 65% of readmissions, enabling targeted intervention deployment.

At Partners Healthcare, deployment of an XGBoost readmission model (AUROC 0.74) across 6 hospitals reduced 30-day readmissions from 14.2% to 11.8% (17% relative reduction) over 18 months. The model identified 8,000+ high-risk patients annually, who received: (1) pharmacist medication reconciliation within 48 hours, (2) nurse follow-up call within 72 hours, (3) PCP appointment within 7 days, and (4) home health referral for patients with >5 comorbidities. The intervention cost 8.4 million in avoided readmission penalties and reduced length of stay.

For heart failure specifically, the HOSPITAL score achieved AUROC of 0.76 for 30-day readmission across 100,000+ discharges, compared to 0.68 for the LACE index. The score's components were: hemoglobin at discharge (OR = 0.88 per g/dL increase), sodium (OR = 0.93 per mEq/L increase), discharge to SNF (OR = 1.7), index admission length of stay (OR = 1.12 per day), and number of procedures (OR = 1.05 per procedure). The model's net benefit analysis showed that targeting the top 30% risk group for intervention achieved 80% of the potential readmission reduction while avoiding unnecessary interventions for 70% of patients.

Common Challenges

• Class imbalance: 15-20% readmission rate means 4:1 negative-to-positive ratio. Solution: Apply SMOTE oversampling, use focal loss with , or implement class weighting in the loss function (weight = 5× for positive class).

• Temporal leakage: Future data accidentally included in features (e.g., readmission diagnosis used to predict readmission). Solution: Use strict time-based train/validation/test splits with gap periods, and audit features for temporal dependencies.

• Heterogeneity: Different readmission causes (cardiac, respiratory, surgical) may require different models. Solution: Train disease-specific models for top readmission diagnoses, or implement multi-task learning that jointly predicts cause-specific readmissions.

• Intervention paradox: High-risk patients may receive interventions that prevent readmission, making them appear low-risk in retrospective data. Solution: Use causal inference methods (inverse probability weighting) to estimate what readmission rate would have been without intervention.

• External validation: Models degrade 5-15% AUROC across hospitals due to different patient populations and practices. Solution: Use domain adaptation (adversarial training), implement local recalibration with 100+ patients, and validate on diverse multi-center data.