Enhancing risk prediction base on health administrative data using high-dimensional prediction model.

Area of Science:

• Epidemiology
• Health Informatics
• Biostatistics

Background:

• Health administrative datasets often lack crucial clinical variables for accurate risk prediction.
• High-dimensional prediction models (hdPMs) can leverage abundant health-care variables to compensate for missing clinical predictors.

Purpose of the Study:

• To compare the predictive performance of hdPMs against conventional models relying solely on investigator-specified clinical predictors.
• To evaluate the utility of health administrative data in enhancing prediction models.

Main Methods:

• A plasmode simulation using tuberculosis (TB) patient data (n=2923) was employed to simulate time-to-event outcomes.
• Conventional and hdPMs were developed, with hdPMs utilizing health-care variables and incorporating LASSO regularization.
• Model performance was assessed using internally validated time-dependent c-statistics and calibration.

Main Results:

• LASSO-regularized hdPMs demonstrated superior predictive performance for TB mortality, achieving a c-statistic of 0.90 compared to 0.78 for the conventional model.
• While non-penalized hdPMs showed overfitting, LASSO-based hdPMs exhibited improved cross-validated discrimination and calibration.
• Sensitivity analyses confirmed consistent results across varying numbers of health-care variables and different outcome types.

Conclusions:

• Health administrative data, particularly when integrated into LASSO-regularized hdPMs, can substantially enhance predictive accuracy for medical outcomes.
• This approach offers a robust method for risk stratification and assessment in epidemiological research, compensating for limitations in clinical data.
• The findings support the use of hdPMs for identifying high-risk individuals and informing targeted interventions.