Feature Selection for Health Care Costs Prediction Using Weighted Evidential Regression.

Area of Science:

• Health Informatics
• Machine Learning
• Data Science

Background:

• Accurate prediction of healthcare costs is crucial for budget management.
• Existing predictive models often lack transparency, hindering the understanding of influencing factors and potentially leading to discrimination.
• There is a need for interpretable models that provide similar accuracy to black-box methods.

Purpose of the Study:

• To develop an interpretable regression method for predicting patient healthcare costs.
• To achieve prediction accuracy comparable to black-box models while maintaining result interpretability.
• To identify key factors influencing healthcare costs and ensure fairness in predictions.

Main Methods:

• Developed an interpretable regression method based on Dempster-Shafer theory using Evidential Regression (EVREG).
• Incorporated a discount function based on feature contribution and used gradient descent for optimal weight learning.
• Employed the k-nearest neighbor (k-NN) algorithm for accelerated calculations and feature selection.

Main Results:

• The proposed transparent EVREG model demonstrated performance comparable to Artificial Neural Networks and Gradient Boosting.
• Achieved an R-squared (R2) value of 0.44 in predicting healthcare costs.
• Successfully identified relevant features for cost prediction and provided interpretable reasoning.

Conclusions:

• The Evidential Regression (EVREG) model offers a transparent and accurate approach to predicting healthcare costs.
• This interpretable method facilitates fair healthcare budget management by revealing prediction drivers.
• The model's transparency addresses the limitations of black-box methods in healthcare cost prediction.