Condition-Specific Readmission Risk Stratification in a Predominantly Black Statewide Cohort Using Machine Learning: Development of Subtype-Specific Models for Heart Failure, Acute Myocardial Infarction, Atrial Fibrillation/Flutter, and Hypertensive Heart Disease

This study developed and validated condition-specific machine learning models using administrative data from a large, predominantly Black statewide cohort to achieve moderate-to-high discrimination in predicting 30-day cardiovascular readmissions, identifying key clinical indices and insurance status as dominant predictors to support equity-informed interventions in safety-net healthcare systems.

The Gist

Imagine a hospital as a busy airport. Every day, patients (passengers) arrive, get treated, and are sent home. The goal is for them to stay home and stay healthy. However, sometimes, within 30 days, they have to come back for a "re-flight" (readmission). This is expensive for the system and stressful for the patient.

This paper is about building a super-smart weather forecast to predict which passengers are most likely to need that re-flight, specifically focusing on a group of travelers who are often overlooked: Black patients.

Here is the breakdown of the study using simple analogies:

1. The Problem: The "One-Size-Fits-All" Mistake

In the past, doctors and computers tried to predict readmissions by grouping all heart problems together into one big bucket. It's like trying to predict the weather for a whole continent by looking at the sky in just one city.

• The Issue: Heart failure, heart attacks, and high blood pressure are different "storms." They need different forecasts.
• The Gap: Most previous forecasts were built using data from mostly White populations. It's like building a weather model based only on Florida data and trying to use it to predict snow in Alaska. It doesn't work well for everyone.

2. The Solution: A Specialized "Flight Control" Team

The researchers used a massive database of 157,000 hospital visits from Virginia, where 96.6% of the patients were Black. This is a huge, focused dataset that finally gives a clear picture of this specific community.

Instead of one big model, they built four specialized "Flight Control" models, one for each specific heart condition:

1. Heart Failure (HF): The engine is struggling to pump.
2. Heart Attack (AMI): A sudden blockage.
3. Atrial Fibrillation (AF/AFL): The heart is beating irregularly.
4. Hypertensive Heart Disease (HHD): High blood pressure damaging the heart over time.

3. The Tools: The "Super-Brain" Algorithms

The team didn't just use one type of computer brain. They tested four different "super-intelligences" (Machine Learning algorithms) to see which one was the best detective:

• XGBoost, LightGBM, Random Forest, and Elastic Net: Think of these as four different detectives. One is great at finding patterns in crowds, another is great at connecting dots, and another is great at spotting outliers.
• The Winner: XGBoost turned out to be the star detective for three out of the four conditions. It was the most accurate at spotting who would come back.
• The Team Effort: They also tried a "Super Learner," which is like a committee where all four detectives vote on the final answer. It worked well, but the single best detective (XGBoost) was usually just as good.

4. The Clues: What Predicts a Return Trip?

The computer didn't just guess; it looked at specific clues. The study found that the most important clues weren't just medical numbers, but also social factors:

• The "LACE" Score: This is a pre-made checklist doctors use (Length of stay, Acuity, Comorbidities, ED visits). It was the single strongest clue.
• Insurance Status: Who pays for the care? This turned out to be a huge predictor. It's not just about money; it's about access. If you don't have good insurance, you might not get the right care at home, leading to a return to the hospital.
• Kidney Health: Patients with kidney issues were much more likely to return.
• The "Obesity Paradox": Interestingly, the model found that obesity sometimes acted as a "shield" in this specific group, likely because it's often linked with having more medical support systems in place, though this is complex.

5. The Results: How Good Was the Forecast?

The models were moderately to highly accurate.

• Heart Failure: About 71% accurate at distinguishing who would return.
• Heart Attack: About 71% accurate.
• Irregular Heartbeat: About 73% accurate.
• High Blood Pressure Heart Disease: This was the best performer, hitting 76% accuracy.

The Catch: While the models were good at ranking patients (saying "Patient A is higher risk than Patient B"), they sometimes overestimated the exact probability of return. It's like a weather app saying "80% chance of rain" when it's actually only 50%. The ranking is useful for prioritizing care, but the exact number needs a little "tuning" before it can be used to tell a patient, "You have a 30% chance of coming back."

6. Why This Matters: Equity in Healthcare

This study is a game-changer because it proves that you can build high-tech, accurate tools specifically for Black communities using data that already exists in hospital records.

• The Takeaway: We don't need to wait for new data. We just need to stop using "one-size-fits-all" models.
• The Future: Hospitals can use these specific models to identify high-risk Black patients before they leave the hospital. This allows them to send a "concierge team" to help with meds, transportation, and follow-up care, preventing the expensive and scary return trip.

In a nutshell: This paper built a specialized, high-tech radar system for Black heart patients. It found that by treating different heart conditions separately and paying attention to social factors like insurance, we can predict who needs extra help much better than before. It's a step toward making healthcare fairer and more effective for everyone.

Technical Summary

1. Problem Statement

Cardiovascular disease (CVD) readmissions impose a significant clinical and economic burden, particularly on safety-net hospitals serving predominantly Black and low-income populations. Existing machine learning (ML) models for predicting 30-day readmissions suffer from three critical limitations:

• Aggregation Bias: Most models aggregate diverse CVD subtypes (e.g., Heart Failure, AMI) into a single outcome, ignoring distinct pathophysiological drivers.
• Demographic Gaps: There is a severe underrepresentation of Black populations in training data, limiting the generalizability of models to minority-serving health systems.
• Feature Limitations: Many studies lack the integration of validated clinical indices and administrative proxies for Social Determinants of Health (SDOH), such as insurance status.

2. Methodology

Study Design & Data Source

• Design: Retrospective cohort study (TRIPOD Type 1a: development).
• Data: Virginia Health Information (VHI) statewide all-payer database (2010–2020).
• Cohort: 157,791 discharge records from 123,272 unique patients.
  • Demographics: 96.6% Black/African American, 3.4% Hispanic/Latino.
  • Conditions: Heart Failure (HF), Acute Myocardial Infarction (AMI), Atrial Fibrillation/Flutter (AF/AFL), and Hypertensive Heart Disease (HHD).
• Outcome: 30-day all-cause readmission to any Virginia acute-care hospital.

Data Preprocessing & Feature Engineering

• Unit of Analysis: Discharge records, grouped at the patient level to prevent data leakage (70/30 train-test split).
• Feature Set: 60 candidate features reduced to 45 consensus-selected features using four independent methods (Boruta, SHAP, Recursive Feature Elimination, mRMR).
  • Clinical Indices: LACE Index, Charlson Comorbidity Index, Elixhauser Comorbidity Count, van Walraven Elixhauser Index.
  • SDOH Proxies: Insurance type, hospital ownership, teaching status.
  • Administrative: Length of stay, admission order, total diagnoses/procedures.
• Handling Imbalance: Models were trained under two configurations: unbalanced (using class weights) and balanced (using SMOTE with K=5).

Model Development

• Algorithms: Four base learners (XGBoost, LightGBM, Random Forest, Elastic Net) and a Super Learner stacked ensemble.
• Total Models: 40 models trained (4 algorithms × 4 conditions × 2 configurations + 8 ensembles).
• Validation:
  • Patient-grouped 5-fold cross-validation for hyperparameter tuning.
  • Held-out test set evaluation.
  • Robustness Checks: Generalized Estimating Equations (GEE) to account for within-patient correlation; clustered bootstrap resampling for confidence intervals.
• Interpretability: SHAP (SHapley Additive exPlanations) values and algorithm-native importance measures were used to identify key predictors.

3. Key Results

Model Performance (Discrimination)
All models achieved moderate-to-high discrimination (AUC 0.70–0.76), outperforming many prior administrative-data benchmarks:

• Hypertensive Heart Disease (HHD): Best performance. XGBoost (balanced) achieved AUC 0.758 (95% CI: 0.735–0.777).
• Atrial Fibrillation/Flutter (AF/AFL): XGBoost (balanced) achieved AUC 0.732 (95% CI: 0.715–0.750).
• Heart Failure (HF): XGBoost (balanced) achieved AUC 0.708 (95% CI: 0.701–0.716).
• Acute Myocardial Infarction (AMI): XGBoost (unbalanced) achieved AUC 0.706 (95% CI: 0.691–0.721).

Algorithm Comparison

• XGBoost was the top-performing base algorithm for all four conditions.
• The Super Learner ensemble was competitive, particularly for HHD, but did not significantly outperform the best base learner in most cases.
• SMOTE Balancing: Generally improved sensitivity for lower-prevalence conditions (HHD, AMI) but slightly reduced specificity.

Calibration

• AMI: Excellent relative calibration (slope 0.993) but significant overestimation of absolute risk (E/O ratio 3.47), requiring recalibration.
• HF: Excellent calibration (slopes ~1.0).
• HHD: Suboptimal calibration (slopes 1.479–1.646), indicating overconfidence at risk extremes; requires Platt scaling or isotonic regression before clinical deployment.

Feature Importance
Three methods (algorithm-native, aggregated, SHAP) converged on the same dominant predictors:

1. LACE Index: The strongest predictor across all conditions (SMD=0.604 between readmitted/non-readmitted).
2. Comorbidity Burden: Charlson and Elixhauser indices.
3. Insurance Type: A critical SDOH proxy.
4. Renal Failure: Consistently high importance across all subtypes.

• Condition-Specific Nuances: HHD models uniquely highlighted obesity and depression; HF models were heavily driven by admission order and renal failure.

4. Key Contributions

1. Largest Predominantly Black Cohort: This is the largest ML-based readmission study focused on a Black population (96.6%), addressing a critical gap in health equity research.
2. Condition-Specific Modeling: Demonstrated that distinct CVD subtypes require tailored models rather than a "one-size-fits-all" approach, with HHD showing the highest predictive utility when modeled separately.
3. Validated Indices over Ad-hoc Composites: Proved that using established clinical indices (LACE, Charlson) combined with administrative SDOH proxies yields robust performance without needing granular clinical data (e.g., labs, vitals).
4. Scalable Framework: Developed a pipeline using only administrative data, making it immediately deployable in safety-net and minority-serving health systems lacking rich electronic health record (EHR) integration.

5. Significance and Limitations

Significance

• Equity-Informed Intervention: By identifying insurance status and comorbidity burden as dominant predictors, the study highlights modifiable structural targets for intervention.
• Clinical Utility: The models provide a scalable decision-support tool for risk stratification, particularly for HHD, which is often subsumed under broader categories in existing literature.
• Policy Implications: Supports the argument that readmission penalties (HRRP) should account for patient-level structural factors (like insurance) rather than solely hospital-level performance.

Limitations

• Generalizability: The 96.6% Black cohort limits applicability to diverse populations; external validation in other demographics is required.
• Data Constraints: Lack of clinical variables (ejection fraction, medications, labs) and specific administrative fields (APR-DRG severity) may cap performance ceilings.
• Bias in Uninsured: Low readmission rates among self-pay/uninsured patients likely reflect access barriers rather than better health, potentially causing the model to underestimate risk for this group.
• Internal Validation Only: While robust internal validation (GEE, clustered bootstrap) was performed, independent external validation in geographically distinct systems is needed before clinical deployment.
• Coding Transition: The study spans the ICD-9 to ICD-10 transition, which may introduce noise in comorbidity ascertainment despite temporal adjustments.

Conclusion
The study successfully developed condition-specific ML models that achieve moderate-to-high discrimination for CVD readmissions in a predominantly Black population. The findings validate the use of administrative data enriched with clinical indices and SDOH proxies to create equity-informed risk stratification tools, with XGBoost emerging as the superior algorithm and the LACE Index as the most critical predictor.