Heterogeneous readmission prediction with hierarchical effect decomposition and regularization

1. Problem Statement

The paper addresses the challenge of predicting hospital readmission risks using Electronic Health Records (EHR). While EHR data offers rich, high-dimensional information, building accurate risk models is difficult due to two main factors:

• High Heterogeneity: Patient populations vary significantly based on their primary diagnoses. A risk factor (e.g., blood pressure) may have different predictive weights for patients admitted for heart failure versus those admitted for burns.
• Data Sparsity in Subgroups: To capture this heterogeneity, models should be tailored to specific Diagnosis-Related Groups (DRGs). However, many DRGs have very small sample sizes, making it impossible to fit separate, high-dimensional models for each without overfitting.
• Limitations of Existing Methods: Standard "one-size-fits-all" models ignore heterogeneity, while "black box" machine learning models (like Random Forests) often lack interpretability and fail to leverage the known hierarchical structure of medical diagnoses.

2. Methodology: The hierNest Framework

The authors propose a novel hierarchical modeling framework called hierNest (Hierarchical Nested Reparameterization). This approach decomposes covariate effects into three levels to borrow strength across related subgroups while maintaining interpretability.

A. Hierarchical Structure

The model utilizes the natural hierarchy of medical diagnoses:

1. Major Diagnostic Categories (MDCs): 25-26 broad categories representing major organ systems or diagnosis types.
2. Diagnosis-Related Groups (DRGs): Specific clinical groupings nested within MDCs.

B. Parameter Decomposition

Instead of estimating a single coefficient vector $\beta_d$ for each DRG $d$, the authors re-parameterize the coefficients as:
$$\beta_d = \mu + \eta_{M(d)} + \delta_d$$
Where:

• $\mu$: A common effect shared across all DRGs.
• $\eta_{M(d)}$: An MDC-specific effect shared by all DRGs within the same MDC.
• $\delta_d$: A DRG-specific effect unique to the specific DRG.

C. Regularization Strategies

To estimate these parameters efficiently and enforce the hierarchical structure, two penalization methods are proposed:

1. hierNest-Lasso:

   • Applies a standard Lasso penalty to the re-parameterized coefficients.
   • Uses sample-size-adjusted penalty weights (fourth root of the ratio of total sample size to subgroup size) to ensure fair shrinkage across groups of different sizes.
   • Allows shrinking $\delta_d$ or $\eta_{M(d)}$ to zero if the data does not support subgroup-specific effects.

2. hierNest-Overlapping Group Lasso (OGLasso):

   • Introduces a structured penalty to enforce hierarchical sparsity.
   • Constraint Logic: An MDC-specific effect ($\eta$) can only be non-zero if the common effect ($\mu$) is non-zero. Similarly, a DRG-specific effect ($\delta$) can only be non-zero if the corresponding MDC effect ($\eta$) is non-zero.
   • This prevents "false positives" where a specific subgroup has an effect while the broader group does not, ensuring the model adheres to the clinical hierarchy.
   • Computation: Since no off-the-shelf software solves this specific overlapping group problem, the authors developed a Group-wise Majorization-Minimization (MM) algorithm combined with Sequential Strong Rules to accelerate convergence.

D. Implementation

• The hierNest-Lasso can be solved by constructing a modified design matrix (using Khatri-Rao products) and feeding it into standard Lasso software (e.g., glmnet).
• The hierNest-OGLasso requires the custom MM algorithm provided by the authors.

3. Key Contributions

• Novel Reparameterization: A method to decompose high-dimensional covariate effects into hierarchical components (Global $\to$ MDC $\to$ DRG) to handle heterogeneity without overfitting small subgroups.
• Structured Regularization: The introduction of an overlapping group lasso penalty that enforces logical constraints on the hierarchy (i.e., you cannot have a specific effect without a general effect), improving model interpretability and stability.
• Computational Efficiency: Development of a custom Majorization-Minimization algorithm and strong rules to solve the complex overlapping group lasso problem efficiently, even with over 1 million parameters.
• Sample Size Adjustment: A specific penalty weighting scheme that accounts for the varying sample sizes of different DRGs, ensuring that small subgroups are not penalized disproportionately.

4. Results

A. Simulation Studies

• Setup: Simulated data with 40 DRGs (nested in 10 MDCs) and 60 predictors. Varied subgroup sample sizes ($n=30$ to $200$) and sparsity levels.
• Performance:
  • hierNest-OGLasso consistently outperformed standard Lasso, Group Lasso, Sparse Group Lasso, and Random Forests, particularly when subgroup sample sizes were small ($n=30, 50$).
  • It achieved higher AUROC (Area Under the Receiver Operating Characteristic curve) and better worst-case AUROC (performance on the smallest subgroups).
  • The method effectively "borrowed strength" from larger groups to improve estimates for smaller groups.
  • Standard Lasso and hierNest-Lasso performed suboptimally because they lacked the strict hierarchical constraints of the OGLasso.

B. Real-World Application (Medicare Readmission)

• Dataset: 67,804 Medicare patients from a Midwestern academic health system (2016–2020) with ~3,300 predictors and ~736 DRGs.
• Validation: Tested on external hospital data and temporal data (year 2020).
• Findings:
  • hierNest-OGLasso achieved the highest mean and median AUROC and AUPRC (Area Under Precision-Recall Curve) across MDCs compared to all competitors.
  • Model Sparsity: Despite having a design matrix with 1.8 million parameters, the final model selected only 13,116 non-zero coefficients.
  • Interpretability:
    • Only 168 of the 3,291 original predictors were selected.
    • 53 predictors had only a global effect (applicable to all patients).
    • Only 10 out of 24 MDCs had unique MDC/DRG-specific effects; the rest shared the same risk model, demonstrating the model's ability to collapse unnecessary complexity.
  • Random Forests performed the worst, highlighting the difficulty of black-box methods in capturing structured heterogeneity with limited data.

5. Significance and Implications

• Clinical Utility: The model provides interpretable risk scores that can be tailored to a patient's specific diagnosis (DRG) while borrowing data from related diagnoses. This helps clinicians identify high-risk patients more accurately than generic models.
• Resource Allocation: By accurately predicting readmission risks for diverse patient subgroups, healthcare systems can better target interventions and allocate resources to those most in need.
• Generalizability: While demonstrated on readmissions, the hierNest framework is applicable to any domain with hierarchical group structures (e.g., genetics with tissue-specific effects, multi-center clinical trials).
• Scalability: The computational framework allows for the analysis of massive, high-dimensional EHR datasets that were previously intractable for hierarchical modeling.

In conclusion, the paper presents a robust statistical solution to the "heterogeneity vs. sparsity" trade-off in medical risk prediction, offering a method that is both statistically powerful and clinically interpretable.