AutoScore

AutoScore generates interpretable point-based clinical risk scores using machine learning to support risk stratification and predictive modeling.

Key Features:

• Modular framework: Six modules are implemented: variable ranking, variable transformation, score derivation, model selection, score fine-tuning, and model evaluation.
• Interpretable point-based scores: Produces point-based scoring models intended to be more interpretable than complex black-box models.
• Machine learning integration: Automates score generation using machine learning techniques to select and weight predictors.
• Predictive accuracy and calibration: Demonstrated competitive predictive accuracy and model calibration compared to logistic regression, stepwise regression, least absolute shrinkage and selection operator (LASSO), and random forest on a dataset of 44,918 intensive care admission episodes.
• Reduced predictor complexity: Generates models using fewer predictors while maintaining performance, reporting an area under the curve (AUC) of 0.780 versus 0.778 for a 24-variable logistic regression model.
• Software implementation: Implemented as an R software package for application to clinical datasets.

Scientific Applications:

• Mortality prediction: Derives point-based scores for predicting mortality in clinical cohorts.
• Risk stratification: Supports stratifying patients by predicted risk to inform clinical analysis and comparisons.
• Resource allocation: Provides quantitative risk estimates that can be used to inform resource allocation decisions in healthcare settings.

Methodology:

Computational methodology comprises six modules—variable ranking; variable transformation; score derivation; model selection; score fine-tuning; and model evaluation—and evaluation using discrimination (AUC) and calibration with comparisons to logistic regression, stepwise regression, least absolute shrinkage and selection operator (LASSO), and random forest on 44,918 intensive care admission episodes; implemented in R.