IsaacT1992/hospital-readmission-xgboost

🏥 Hospital Readmission Prediction (XGBoost)

Author: Isaac Tosin Adisa · Florida State University

📌 Overview

This model predicts 30-day hospital readmission risk using structured clinical features derived from the MIMIC-IV dataset. It is part of an integrated multi-model comparative framework alongside Logistic Regression and LightGBM, designed to address three major barriers to clinical AI deployment: lack of explainability, inadequate fairness evaluation, and absence of production reliability infrastructure.

The model outputs calibrated probabilities suitable for downstream clinical risk stratification workflows. This model is released alongside a fully reproducible pipeline and open-source implementation to facilitate independent validation and reuse.

📊 Dataset

Property	Value
Source	MIMIC-IV (Beth Israel Deaconess Medical Center)
Total admissions	415,231 adult hospital admissions
30-day readmission prevalence	~18%
Feature count	26 clinically derived features
Split	Train / Validation / Test (temporal split)

Features include demographics, prior utilization, primary diagnosis category, comorbidity burden, medication count, lab value summaries, and length of stay.

⚠️ Raw data is not included and requires credentialed access via PhysioNet.

⚙️ Training

Setting	Value
Framework	XGBoost
Objective	Binary logistic
Class imbalance	scale_pos_weight tuned to prevalence
Hyperparameter tuning	Optuna (Bayesian search)
Calibration	Platt scaling (post-hoc)

📈 Performance

Metric	Value	Notes
AUC-ROC	0.696 (95% CI: 0.691–0.701)	Strong discriminative performance
Brier Score	~0.217	Calibration reference
Benchmark	Comparable to LACE Index (0.60–0.68)	Validated clinical tool

📊 XGBoost delivers the strongest discrimination among tree-based models in this framework. LightGBM achieves better calibration (Brier Score: 0.146), making the two complementary depending on the clinical use case.

🔍 Explainability

Per-patient explanations are generated using SHAP TreeExplainer, which is exact and computationally efficient for tree-based models.

• Global feature importance via SHAP summary plots
• Local patient-level force plots for individual predictions
• Compatible with standard clinical decision support workflows

Top predictors identified by SHAP:

Rank	Feature
1	Prior hospital admissions (12 months)
2	Medication count
3	Diagnosis count
4	Length of stay
5	Charlson Comorbidity Index

⚖️ Fairness Evaluation

The model was evaluated across 16 demographic and clinical subgroups, including stratifications by race/ethnicity, age group, sex, and insurance type.

All subgroups satisfy the following thresholds:

Metric	Threshold	Result
ΔAUC (vs. overall)	≤ 0.05	✅ Met
ΔFNR (vs. overall)	≤ 0.10	✅ Met

No subgroup exhibited clinically meaningful performance degradation under these criteria. No post-processing bias correction was required.

🚀 Usage

import joblib
import numpy as np

# Load model
model = joblib.load("xgboost.pkl")

# Replace with your 26 clinical features
X = np.array([[...]])

# Returns 30-day readmission probability
pred = model.predict_proba(X)[0][1]

print(f"Readmission risk: {pred:.3f}")

⚠️ Input features must match the 26 clinical variables used during training. See the repository for the full feature schema and preprocessing pipeline.

🎯 Intended Use

• Research and reproducibility
• Clinical ML benchmarking
• Demonstration of explainable and fair AI systems
• Benchmarking against validated clinical tools (e.g. LACE Index)

🧾 Ethical & Regulatory Considerations

This model is not a medical device and is not approved for clinical use. Deployment in any clinical setting requires:

• Prospective validation on local patient populations
• Institutional review and governance approval
• Applicable regulatory compliance

This work is aligned with:

• ONC HTI-1 — AI transparency requirements for health IT
• HHS Section 1557 — non-discrimination standards in healthcare AI

🔁 Reproducibility

All results are fully reproducible using the open-source pipeline at github.com/Tomisin92/readmission-prediction, which includes data preprocessing, feature engineering, model training, SHAP explainability, and fairness auditing.

⚠️ Limitations

• Retrospective validation only — model was trained and evaluated on historical MIMIC-IV data; prospective validation has not been performed.
• Single institution — MIMIC-IV reflects one academic medical center (BIDMC); generalizability to other institutions requires local validation.
• No causal claims — feature associations do not imply clinical causation.
• No real-time EHR integration — this model operates on static feature vectors; live deployment would require additional infrastructure.
• Requires local validation before any deployment in a clinical decision support context.
• Credentialed dataset — MIMIC-IV requires PhysioNet credentialing; this model card does not distribute the underlying data.

🔗 Links

• 📄 Paper: arXiv:2604.22535
• 💻 Code: github.com/Tomisin92/readmission-prediction

📜 Citation

@misc{adisa2025readmission,
  title={Hospital Readmission Prediction with Explainability and Fairness},
  author={Adisa, Isaac Tosin},
  year={2026},
  eprint={2604.22535},
  archivePrefix={arXiv},
  primaryClass={cs.LG}
}

License

This model is released under the MIT License. The underlying MIMIC-IV dataset is subject to its own PhysioNet credentialed access agreement.