Observational Health Data Sciences and Informatics

Observational Health Data Sciences and Informatics provides a standardized analytics pipeline for development and rigorous validation of prediction models using observational health data.

Key Features:

• Standardized analytics pipeline: A structured approach for rapid development and validation of prediction models using observational health data.
• Bias mitigation strategies: Validates phenotypes, specifies target populations, conducts large-scale external validations, and publicly shares analytical source code to reduce sources of bias.
• Machine learning methods: Implements AdaBoost, random forest, gradient boosting machine, decision tree, L1-regularized logistic regression, and MLP neural network for model development.
• End-to-end process: Supports steps from problem design through model development and evaluation.
• Data scale and sources: Applicable to observational datasets such as claims databases (e.g., development on a USA claims database with >20,000 hospitalizations) and supports external validation across multi-country datasets (>45,000 hospitalizations).

Scientific Applications:

• COVID-19 mortality risk prediction (0 to 30 days): Development and validation of models predicting 0 to 30 day post-hospitalization mortality for COVID-19 patients.
• Cross-population external validation: External validation of predictive models across datasets from the USA, South Korea, and Spain.
• Comparative algorithm evaluation: Comparative assessment of AdaBoost, random forest, gradient boosting machine, decision tree, L1-regularized logistic regression, and MLP neural network performance.
• Model performance assessment: Evaluation of discrimination and calibration during internal and external validations.

Methodology:

Model development used AdaBoost, random forest, gradient boosting machine, decision tree, L1-regularized logistic regression, and MLP neural network; phenotype validation and explicit target population specification were performed; internal and large-scale external validations were conducted using observational claims data (development: >20,000 US hospitalizations; external: >45,000 hospitalizations across South Korea, Spain, and the USA); discrimination and calibration were evaluated and analytical source code was shared.