A DEEP NEURAL NETWORK TWO-PART MODEL AND FEATURE IMPORTANCE TEST FOR SEMI-CONTINUOUS DATA.

Area of Science:

• Biostatistics
• Machine Learning
• Health Data Science

Background:

• Semi-continuous data, characterized by a mix of zero and positive values, are prevalent in clinical settings, such as postoperative pain (POP) scores.
• Traditional two-part models struggle with non-linear and non-additive relationships common in health data.
• Existing methods may not adequately capture the complex interactions influencing health outcomes.

Purpose of the Study:

• To develop a robust deep neural network (DNN)-based two-part model for semi-continuous data.
• To enhance model stability and interpretability for complex health outcome analyses.
• To introduce a feature importance testing procedure for improved predictive performance and statistical inference.

Main Methods:

• A novel stable deep neural network (sDNN) model was developed, incorporating bootstrapping and filtering for enhanced stability.
• A feature importance testing procedure (fsDNN) was derived to identify key predictors for each data process.
• The models were applied to postoperative pain data and validated through extensive numerical simulations and comparisons with other machine learning methods.

Main Results:

• The sDNN and fsDNN models demonstrated superior performance compared to existing parametric and semi-parametric two-part models.
• fsDNN successfully identified important features and improved the predictive accuracy of the sDNN model.
• The proposed methods consistently outperformed conventional approaches across various data complexities.

Conclusions:

• The DNN-based sDNN and fsDNN models offer a powerful and flexible alternative for analyzing semi-continuous health data.
• fsDNN enhances interpretability and predictive power by identifying crucial influencing factors.
• These advanced methods provide significant advantages in understanding and predicting health outcomes like postoperative pain.