Model Performance Dashboard

Technical deep dive into model accuracy, features, and validation

UCI Diabetes Model

Logistic Regression | 71,518 patients

AUC-ROC
0.564
Moderate discrimination
Readmission Rate
8.8%
Diabetes cohort

Performance Metrics

Sensitivity (Recall)
72%
Specificity
48%
Precision
11%
Clinical Interpretation

Model captures 72% of actual readmissions (high sensitivity) but has moderate specificity. Optimized for recall to minimize missed high-risk patients in care management workflows.

MIMIC-IV ICU Model

Gradient Boosting | 211,354 patients

AUC-ROC
0.630
+12% vs UCI
Readmission Rate
20.5%
ICU population

Performance Metrics

Sensitivity (Recall)
68%
Specificity
58%
Precision
28%
Clinical Interpretation

Enhanced model with 60+ clinical features (labs, vitals, procedures) improves discrimination by 12%. Better balance of sensitivity and specificity makes it suitable for broader ICU populations.

Model Selection Rationale

UCI Model (Production)

Chosen for general diabetes populations. Uses administrative features (visits, medications, diagnoses) available in most EHRs. Simple, interpretable, and fast to implement.

MIMIC Model (Research)

Enhanced for ICU/clinical settings. Incorporates vital signs, lab results, and procedures. 12% better discrimination but requires richer data infrastructure.

Trade-Offs

Both models optimized for recall. In care management, it's better to over-identify risk (some false positives) than miss truly high-risk patients who need intervention.

Validation & Model Governance

Validation Strategy

  • Train/Test Split: 80/20 holdout set never seen during training
  • Cross-Validation: 5-fold CV to assess stability (AUC variance: ±0.02)
  • Temporal Validation: Models tested on chronologically later admissions
  • Calibration: Isotonic regression applied to ensure predicted probabilities match observed rates

Fairness & Bias Analysis

  • Age Groups: Model performs consistently across age bands (AUC: 65+ = 0.59, <65 = 0.57)
  • No Protected Attributes: Race, ethnicity, gender excluded from model features
  • Proxy Monitoring: ZIP code, insurance type monitored for disparate impact
  • Ongoing Audits: Quarterly performance reviews stratified by demographics

Model Limitations & Disclaimers

  • Historical Data: UCI model trained on 1999-2008 data may not reflect current clinical practice patterns, medication formulations, or care delivery models.
  • Missing SDOH: Models don't account for social determinants of health (housing stability, food insecurity, transportation barriers) which significantly impact readmission risk.
  • External Validation Needed: Before clinical deployment, models must be validated on your specific patient population, geography, and care delivery system.
  • Not for Clinical Use: This platform is a demonstration tool. Risk scores should not be used for clinical decision-making without rigorous validation and regulatory compliance review.