Variable importance and prediction methods for longitudinal problems with missing variables.

Area of Science:

• Biostatistics
• Medical Informatics
• Data Science

Background:

• Current medical practice for severe trauma prognosis relies on simplified scoring systems, often ignoring dynamic, high-dimensional patient data.
• Longitudinal data with missing continuous and binary exposures present significant analytical challenges.

Purpose of the Study:

• To introduce novel prediction and variable importance (VIM) methods for longitudinal data with missingness.
• To apply these methods for improved prognosis in severe trauma patients, informing clinical decisions.

Main Methods:

• Developed VIM parameters analogous to regression coefficients but model-independent and causally interpretable.
• Employed a flexible SuperLearner ensemble method for prediction model fitting.
• Utilized targeted maximum likelihood estimation (TMLE) for doubly robust and locally efficient inference.

Main Results:

• The proposed VIM methods identified significant effects missed by traditional parametric approaches like stepwise regression or LASSO.
• The prediction model demonstrated improved cross-validated fit, including enhanced area under the curve (AUC) for receiver-operator curves (ROC).

Conclusions:

• The new VIM methods offer a robust, interpretable, and flexible alternative for variable importance estimation in high-dimensional data.
• These methods can enhance clinical decision-making by leveraging comprehensive patient data for prognosis.