AutoScore-Ordinal

AutoScore-Ordinal automates construction of interpretable point-based clinical scoring models for ordinal outcomes to enable risk prediction and stratification.

Key Features:

• Extension to ordinal outcomes: Extends the original AutoScore framework for binary outcomes to handle ordinal outcomes.
• Variable ranking: Implements variable ranking to prioritize candidate predictors from input data.
• Variable transformation: Supports variable transformation to convert predictors into score-friendly formats.
• Score derivation: Derives point-based scores using proportional odds models.
• Model selection: Includes model selection procedures to choose predictor subsets.
• Score fine-tuning: Provides score fine-tuning to adjust point allocations for interpretability.
• Model evaluation: Evaluates model performance using metrics such as mean area under the receiver operating characteristic curve (AUC) and generalized c-index.
• High-dimensional data support: Identifies potential predictors from high-dimensional datasets.
• Flexible variable selection: Employs a flexible variable selection procedure to derive compact models (example models used eight predictors).
• Dataset splitting: Supports training/validation/testing splits (example split: 70% train / 10% validation / 20% test).

Scientific Applications:

• Clinical risk prediction: Development of interpretable risk prediction models for ordinal clinical outcomes to support risk stratification.
• EHR validation: Applied to electronic health records from Singapore General Hospital emergency department (2008–2017) on 445,989 inpatient cases.
• Outcome modeling: Models ordinal outcomes such as alive without 30-day readmission (80.7%), alive with 30-day readmission (12.5%), and death during hospitalization or within 30 days post-discharge (6.8%).
• Model comparison: Enables comparison to alternative models, with example mean AUCs of 0.758 and 0.793 and generalized c-index scores of 0.737 and 0.760 for two point-based models.
• Clinical decision support: Supports applications in clinical decision-making and resource allocation through interpretable score outputs.

Methodology:

Computational steps comprise variable ranking, variable transformation, score derivation using proportional odds models, model selection, score fine-tuning, and model evaluation, with example training/validation/test split of 70%/10%/20% and reporting of mean AUC and generalized c-index.